Loshin Enterprise Knowledge Management

PREFACE
While data quality problems are widespread, it is rare for an event to

take place that provides a high-profile example of how questionable
information quality can have a worldwide business effect. The 2000 US
Presidential election and the subsequent confusion around the Florida
recount highlights the business need for high quality data. The winner
of this election is in a position to influence economies around the world.
The uncertainty associated with the lack of a clear winner had an
immediate effect the day after the election when stock prices plum-
meted. Whether it is unintuitive data presentation, questions about the
way information is aggregated, or the method by which information
policy regulates the use of data, valuable information quality lessons
regarding at least six data quality issues that can be learned from the
election.
POOR DATA REPRESENTATION
A poor decision with respect to data presentation resulted in voter con-

fusion. The use of the butterfly ballot in Palm Beach County, FL is an
example how the presentation of information did not correlate to users'
expectations, leading to a large number of voting errors. In fact, many
voters later claimed that they were dismayed to learn they may have
voted for Pat Buchanan, a right-wing Conservative candidate, instead
of Al Gore, the Democratic party candidate.
XIII
XIV PREFACE
DATA VALIDATION
With no built-in mechanism to vaHdate the data before it enters the sys-
tem, the use of punch cards and the "butterfly ballot" leads to problems
with vote validation. When using a punch card ballot, (which, accord-
ing to the LA Times was used by more than 37 percent of registered
nationwide voters in 1996), the voter selects a candidate by poking out
the chad^the perforated section that should be ejected when the hole is
punctured. The cards are read by a tabulation machine, which counts a
vote when it reads the hole in the card.
The validation issue occurs when the chad is not completely
ejected. The automated tabulation of both "hanging chads" (chads that
are still partially attached) and "pregnant chads" (chads that are
bulging but not punched out) is questionable, and so it is not clear
whether all votes are counted. What constitutes a valid vote selection is
primarily based on whether the tabulation machine can read the card.
In the case of recounts, the cards are passed through the reader multiple
times. In that process some of the hanging chads are shaken free which
leads to different tallies after each recount.
In addition, if someone mistakenly punches out more than one
selection, the vote is automatically nullified. It is claimed that 19,000
ballots were disqualified because more than one vote for president had
been made on a single ballot. This is an example where a policy to pre-
qualify the ballot before it is sent to be counted could be instituted.
Since the rules for what constitutes a valid vote are well described, it
should be possible to have a machine evaluate the punch card to deter-
mine whether it is valid or not, and notify the voter that the ballot
would be invalidated before it is cast.
INVALID ANALYTICAL MODELS
For many people on election eve, it is customary to sit in front of their

televisions and watch as their favorite newscasters predict the alloca-
tion of electoral votes. These predictions are based on the results of exit
polls and election results provided by an organization called the Voter
News Service (VNS), which is jointly owned by a coalition of news
companies in order to cut the cost of data collection. Typically, the VNS
feeds both vote counts and winner predictions to all the news media
simultaneously, which is why all the different broadcasters seem to pre-
dict the winners all around the same time.
PREFACE XV
In the case of the 2000 election, the networks were led to predict
the winner of Florida incorrectly, not just once, but twice. The first
error occured because predicting elections is based on statistical models
generated from past voting behavior that 1) were designed to catch vote
swings an order of magnitude greater than the actual (almost final) tal-
lies and 2) did not take changes in demographics into account. This
meant that the prediction of Gore's winning Florida was retracted
about 2 hours after it was made.
CONFLICTING DATA SOURCES
By 2:00AM the VNS (and consequently, the reporting organizations)

switched their allocation of Florida's electoral votes from Gore to Bush,
and declared Bush to have enough electoral votes to win the election.
However, a second retraction occured when actual vote tallies were dis-
puted. While the VNS report indicated that Bush led in Florida by
29,000 votes, information posted on the Florida Board of Elections web
site indicated that Bush's lead was closer to 500 votes, with the gap nar-
rowing quickly. A computer glitch in Volusia County led to an overesti-
mation of Bush's total by more than 25,000 votes.
EXPECTATION OF ERROR
According to Title IX, Chapter 102 of Florida law, "if the returns for
any office reflect that a candidate was defeated or eliminated by one-
half of a percent or less of the votes cast for such office...the board
responsible for certifying the results of the vote...shall order a recount
of the votes..."
This section of the law contains data-accuracy implication that
there is an expected margin of error of one-half of one percent of the
votes. The automatic recount is a good example where the threshold for
potential error is recognized and where there is defined governance
associated with a data quality problem.
TIMELINESS
Timeliness is an important aspect of information quality. We have come

to expect that the results of a national election are decided by the time
XVI PREFACE
we wake up the next day. Even in a close election when the results are
inconclusive, there are timeliness constraints for the reporting and certi-
fication of votes.
DATA QUALITY ON THE PERSONAL LEVEL
My personal connection to data quality began at a very early age,

although I did not realize the connection until recently. My parents,
through family coercion, decided to give me the first name "Howard,"
although they had wanted to give me the name "David." So David
became my middle name, but even though they named me Howard, I
have always been called by my parents and, subsequently, everyone
else David.
Hence the source of the problem. This fixed-field database-oriented
world is ill prepared to deal with a person who is called by his middle
name. Everywhere you go, you are asked for your first name and middle
initial. So, officially, I am Howard, but I always go by the name David.
At school, at camp, at college, at work, filling out applications, opening
bank accounts, filing tax returns, and so on, I fill out my name in its
official form.
People try to bend the rules: "Put down David as your first name
and H as your middle initial." "Just use David." "Scratch out first name
and change it to middle name." Unfortunately, these ideas are too radi-
cal for the poor data entry people, so I either end up as David H. Loshin
or I am stuck with being called Howard.
This really happened: At a doctor's office recently, the receptionist
insisted that if my name were officially Howard D. Loshin on my in-
surance card, they were only going to call me Howard. I have three dif-
ferent credit cards, all with different names on them. Not only that
my last name, Loshin, sounds the same as "lotion," and I find my last
name is consistently misspelled: Lotion, Loshen, Loshion, Loshian.
When it comes to how my mail is addressed, I never know what to
expect except for one thing: I will get at least two of every direct
marketing sales letter.
Despite my inherited data quality connection, it was not until a few
years ago that I found a new motivation with data quality. I was asso-
ciated with a securities processing group at a large financial services
company, who were analyzing their accounts database. What was inter-
esting was that up to that point they had considered their accounts as
PREFACE XVII
just that: accounts. Over a period of time, however, some people there
became convinced of the benefits of looking at the people associated
with those accounts as customers, and a new project was born that
would turn the accounts database inside out. My role in that project was
to interpret the different information paradigms that appeared in the
accounts database name and address field. For it turned out that a single
customer might be associated with many different accounts, in many dif-
ferent roles: as an owner, a trustee, an investment advisor, and so forth.
I learned two very interesting things about this project. The first
was that the knowledge that can be learned from combining multiple
databases was much greater than from the sum total of analyzing the
databases individually. The second was the realization that the prob-
lems that I saw at this organization were not limited to this company
in fact, these problems are endemic and not only within the financial
industry but in any industry that uses information to run its businesses.
The insight that brought full circle the world of data quality was
this: Every business process that uses data has some inherent assump-
tions and expectations about the data. And these assumptions and
expectations can be expressed in a formal way, and this formality can
expose much more knowledge than simple database schema and Cobol
programs.
So I left that company and formed a new company. Knowledge
Integrity Incorporated, (www.knowledge-integrity.com) whose purpose
is to understand, expose, and correct data quality problems. Our goal is
to create a framework for evaluating the impacts that can be caused by
low data quality, to assess the state of data quality within an enterprise,
to collect the assumptions and expectations about the data that is used,
and recast those assumptions and expectations as a set of data quality
and business rules. In turn, these rules are incorporated as the central
core of a corporate knowledge management environment, to capture
corporate knowledge and manage it as content.
This book is the product of that goal. In it, we elaborate on our
philosophy and methods for evaluating data quality problems and how
we aim to solve them. I believe that the savvy manager understands the
importance of high-quality data as a means for increasing business
effectiveness and productivity, and this book puts these issues into the
proper context. I hope the reader finds this book helpful, and I am cer-
tainly interested in hearing about others' experiences. Please feel free to
contact me at [email protected] and let me know how
your data quality projects are moving along!
XVIII PREFACE
I have been collecting what I call "Data Quality Horror Stories"

and placing them on our corporate Web site (www.knowledge-
integrity.com/horror.htm). If you have any interesting personal experi-
ences, or if you see news stories that demonstrate how poor data quality
has serious (or comical) effects, please e-mail them to me at Loshin
knowledge-integrity.com.
I would like to thank the people who have helped make this book
possible. First and foremost, my wife, Jill, and my children, Kira and
Jonah, are always there when I need them. Ken Morton acted as acqui-
sitions editor and general enabler. Thomas Park, who took over the
project from Ken, was invaluable in helping me get this project com-
pleted. I thank Thomas Redman, with whom I worked for a short
period of time and consulted on some of the concepts in this book.
Thank you to Mary O'Brien, who read through early drafts of the pro-
posal and was a big supporter.
Thanks also go to Bob Shelly, whose experienced eye validated my
content, and to the rest of the Morgan Kaufmann staff involved in this
project. Sheri Dean and Julio Esperas also provided significant help in
the preparation of the book.
I also must thank Justin Kestelyn of Intelligent Enterprise maga-
zine, who has vetted some of my ideas by publishing abridged versions
of a few chapters. Thanks also go to Dennis Shasha at New York Uni-
versity, who gave me the opportunity to teach this material as a special
topic graduate class at the Courant Institute. I also thank my wife's par-
ents, Marty and Phyllis Fingerhut, who are two of my biggest support-
ers. Last, I want to thank and remember my mother, Betty Loshin, who
was a source of inspiration and who passed away earlier this year.
1
INTRODUCTION
Without even realizing it, everyone is affected by poor data quality.

Some are affected directly in annoying ways, such as receiving two or
three identical mailings from the same sales organization in the same
week. Some are affected in less direct ways, such as the 20-minute wait
on hold for a customer service department. Some are affected more
malevolently through deliberate fraud, such as identity theft. But when-
ever poor data quality, inconsistencies, and errors bloat both companies
and government agencies and hamper their ability to provide the best
possible service, everyone suffers.
Data quality seems to be a hazy concept, but the lack of data qual-
ity severely hampers the ability of organizations to effectively accumu-
late and manage enterprise-wide knowledge. The goal of this book is to
demonstrate that data quality is not an esoteric notion but something
that can be quantified, measured, and improved, all with a strict focus
on return on investment. Our approach is that knowledge management
is a pillar that must stand securely on a pedestal of data quality, and by
the end of this book, the reader should be able to build that pedestal.
This book covers these areas.
Data ownership paradigms
The definition of data quality
An economic framework for data quality, including steps in build-
ing a return on investment model to justify the costs of a data
quality program
The dimensions of data quality
Using statistical process control as a tool for measurement
ENTERPRISE KNOWLEDGE MANAGEMENT
Data domains and mappings between those domains

Data quality rules and business rules
Measurement and current state assessment
Data quality requirements analysis
Metadata and policy
Rules-based processing
Discovery of metadata and data quality and business rules
Data cleansing
Root cause analysis and supplier management
Data enhancement
Putting it all into practice
The end of the book summarizes the processes discussed and the steps
to building a data quality practice.
Before we dive into the technical components, however, it is worth-
while to spend some time looking at some real-world examples for
motivation. In the next section, you will see some examples of "data
quality horror stories" tales of adverse effects of poor data quality.
1.1 DATA QUALITY HORROR STORIES
1.1.1 Bank Deposit?
In November of 1998, it was reported by the Associated Press that a

New York man allegedly brought a dead deer into a bank in Stamford,
Connecticut, because he was upset with the bank's service. Police say the
70-year-old argued with a teller over a clerical mistake with his checking
account. Because he was apparently unhappy with the teller, he went
home, got the deer carcass and brought it back to the branch office.
1.1.2 CD Mail Fraud
Here is a news story taken from the Associated Press newswire. The
text is printed with permission.
Newark For four years a Middlesex County man fooled
the computer fraud programs at two music-by-mail clubs,
using 1,630 aliases to buy music CDs at rates offered only to
first-time buyers.
INTRODUCTION
David Russo, 33, of Sayerville, NJ, admitted yesterday that

he received 22,260 CDs by making each address even if it
listed the same post office box different enough to evade
fraud-detection computer programs.
Among his methods: adding fictitious apartment numbers,
unneeded direction abbreviations and extra punctuation
marks. (Emphasis mine)
The scam is believed to be the largest of its kind in the
nation, said Assistant U.S. Attorney Scott S. Christie, who
prosecuted the case.
The introductory offers typically provided nine free CDs
with the purchase of one CD at the regular price, plus ship-
ping and handling. Other CDs then had to be purchased later
to fulfill club requirements. Russo paid about $56,000 for
CDs, said Paul B. Brickfield, his lawyer, or an average of
$2.50 each. He then sold the CDs at flea markets for about
$10 each. Brickfield said. Russo pleaded guilty to a single
count of mail fraud. He faces about 12 to 18 months in
prison and a fine of up to $250,000.
1.1.3 Mars Orb/Cer
The Mars Climate Orbiter, a key part of NASA's program to explore the
planet Mars, vanished in September 1999 after rockets were fired to
bring it into orbit of the planet. It was later discovered by an investiga-
tive board that NASA engineers failed to convert English measures of
rocket thrusts to newtons, a metric system measuring rocket force, and
that was the root cause of the loss of the spacecraft. The orbiter
smashed into the planet instead of reaching a safe orbit.
This discrepancy between the two measures, which was relatively
small, caused the orbiter to approach Mars at too low an altitude. The
result was the loss of a $125 million spacecraft and a significant setback
in NASA's ability to explore Mars.
1.1.4 Credit Card Woes
After having been a loyal credit card customer for a number of years, I
had mistakenly missed a payment when the bill was lost during the
move to our new house. I called the customer service department and
explained the omission, and they were happy to remove the service
charge, provided that I sent in my payment right away, which I did.
A few months later, I received a letter indicating that "immediate
action" was required. Evidently, I had a balance due of $0.00, and because
of that, the company had decided to revoke my charging privileges! Not
only that, I was being reported to credit agencies as being delinquent.
Needless to say, this was ridiculous, and after some intense conver-
sations with a number of people in the customer service department,
they agreed to mark my account as being paid in full. They notified the
credit reporting agencies that I was not, and never had been, delinquent
on the account (see Figure 1.1).
1.1.5 Open or Closed Account?
Three months after canceling my cellular telephone service, I continue

to receive bills from my former service provider indicating that I was
being billed for $0.00 "Do not remit."
1.1.6 Business Credit Card
A friend of mine is the president of a small home-based business. He

received an offer from a major charge card company for a corporate
charge card with no annual fee. He accepted, and a short time later, he
received his card in the mail. Not long after that, he began to receive the
same offer from the same company, but those offers were addressed dif-
ferently. Evidently, his name had been misspelled on one of his maga-
zine subscriptions, and that version had been submitted to the credit
card company as a different individual. Not only that, his wife started
to receive offers too.
Six months later, this man still gets four or five mail offers per week
in the mail from the same company, which evidently not only cannot fig-
ure out who he is but also can't recognize that he is already a customer!
1.1.7 Direct Marketing
One would imagine that if any business might have the issue of data
quality on top of its list, it would be the direct marketing industry. Yet, I
INTROOUCnON
li':;"-,fW.^W^ :>' *'

fUMwdirieArtiofillelilmt
III '
Ytmr '^ QMI i gBWiiilllM
> Yon iiii# 4ift icitfl CQIIM^ lil 904

IfmtisldelliiU
If mtlslde llii U i r / | | i r hmm'air oomrMlnti t.1KI 11.111. to li^ mj^t M m i ^ flir<M|fi
fridny, wMl 1:00 .tn. t 4i\ SdliirA^ (EMIQIII Stiiiârtf Time).
Si(icerly
<%ifi. "-.-.v sr-jtNmcx*
ll:iiiiii|iillifiiK^^
iiiiiiiiiiliiiiiiiiiiiiiii^
W^^.:M0m'X llllllllljj^^^^
mmm m^tn lamin it~"
Figure 1.1 Mysterious bill

recently received two identical pieces of mail the same day from the
local chapter of an association for the direct marketing industry. One
was addressed this way.
David Loshin
123 Main Street
Anytown, NY 11787
Dear David,. . .
The other was addressed like this
Loshin David
123 Main Street
Anytown, NY 11787
Dear Loshin,. . .
1.1.8 Tracking Backward
I recently ordered some computer equipment, and I was given a track-

ing number to follow the package's progress from the source to my
house. If you look at the example in Table 1.1 (which has been slightly
modified from the original), you will see that the package was scanned
at the exit hub location in a specific state on June 26, was (evidently)
scanned in Nassau county, NY, at 12:30 A.M. the following day but was
scanned as a departure from the airport in the same state as the exit hub
at 1:43 P.M., which is practically 11 hours later. The rest of the tracking
makes sense from the XX airport to an airport local to my area, then
onto my locality, and finally to the delivery point.
Obviously, the June 27, 12:30 A.M. scan in Nassau has either the
incorrect location or the incorrect time. It is most likely the incorrect
time, since packages are scanned on entry to a location and on exit, and
this scan appears between the location scan at EXIT HUB and the
departure scan at ANYTOWNINTL, same state.
1.1.9 Conclusions?
These are just a few stories culled from personal experience, interac-
tions with colleagues, or reading the newspaper. Yet, who has not been
subject to some kind of annoyance that can be traced to a data quality
problem?
INTRODUCTION
TABLE 1
Tracking history for the equipment I ordered.
PACKAGE PROGRESS
Date Time Location Activity
June 28, 2000 5:25 P.M. NASSAU-HICKSVILLE,

NY US DELIVERED
3:42 A.M. NASSAU, NY DESTINATION
3:31 A.M. NASSAU, NY US LOCATION SCAN
June 27, 2000 11:21 P.M. NEWARK INTL, NJ DEPARTURE SCAN
4:45 P.M. NEWARK INTL, NJ ARRIVAL
1:43 P.M. ANYTOWN INTL, XX DEPARTURE SCAN
12:30 A.M. NASSAU, NY US ARRIVAL
June 26, 2000 11:29 A.M. EXIT HUB, XX US LOCATION SCAN
June 23, 2000 9:11 P.M. ADDISON, IL US LOCATION SCAN
1:38 P.M. SHIPMENT DATA
RECEIVED
1.2 KNOWLEDGE MANAGEMENT AND DATA QUALITY
Over the past 30 years, advances in data collection and database tech-
nology have led to massive legacy databases controlled by legacy soft-
ware. The implicit programming paradigm encompasses both business
policies and data validation policies as application code. Yet, most
legacy applications are maintained by second- and third-generation
engineers, and it is rare to find any staff members with firsthand experi-
ence in either the design or implementation of the original system. As a
result, organizations maintain significant ongoing investments in daily
operations and maintenance of the information processing plant, while
mostly ignoring the tremendous potential of the intellectual capital that
is captured within the data assets.
1.2.1 What Is Knowledge Management?
An organization's data collection is a valuable business resource that

until now has been largely underutilized. In the past, when data were
mostly locked up in databases, ferociously guarded by organizational
overlords, the ability to share and benefit from enterprise knowledge

was limited. Today, as technology evolves to unlock and distribute these
databases, a procedural methodology has evolved in tandem to help
integrate the technical, organizational, and behavioral issues associated
with enterprise knowledge. This methodology is referred to as "knowl-
edge management."
According to Karl Erik Sveiby, knowledge management is "The art
of creating value by leveraging the intangible assets." The Gartner Group
states it in a more down-to-earth way: "Knowledge management is a dis-
cipline that promotes an integrated approach to identifying, managing,
and sharing all of an enterprise's information assets. These information
assets may include databases, documents, policies and procedures as well
as previously unarticulated expertise and experience resident in individ-
ual workers."^
In other words, knowledge management is a strategic process
meant to capture the ways that an organization integrates its informa-
tion assets with the processes and policies that govern the manipulation
of those intellectual assets. The desired goal of knowledge management
is the determination and the harnessing of the value of the information
resources within the enterprise.
While knowledge management encompasses many disciplines, such
as document management or e-mail technology, our goal is to focus on
the embedded knowledge in data sets that can be expressed as a set of
business rules. Having expressed these rules, we can then validate our
expectations of the information that we use by testing it against the busi-
ness rules.
1.2.2 What Are Business Rules?
A business rule is an assertion about the state of a business process. All

business processes and data sets have rules. Unfortunately, these rules
are frequently expressed as "lore" instead of being properly docu-
mented, passed from one "generation" of managers or technicians to
another by word of mouth. Even worse, sometimes these rules are for-
gotten over time, having been implemented in software and then left to
happily chug away until a crisis strikes.
1. http.7/cestec 1 .mty.itesm.mx-laava/sdsites/cursos/pqg_base/clefinicion 1 .htm

INTRODUCTION
Business rules are likely to be expressed in a natural language. An

example of a business rule for an employee database might be "No
employee makes more than five times the salary of the lowest-paid
employee." An example for an electronic data interchange system might
be "If the header of an incoming message includes the symbol #, then
the message is routed to the accounts payable office." A business rule
might govern the way operations proceed. An example for a sales office
might be "No customer is pitched for a new sales transaction within 10
days after the customer's last sales transaction."
When business rules are undocumented, the chances are high that
the meanings or implications of these rules will be lost within a short
period of time. Knowledge is lost when employees leave the company or
change positions internally, when managers have too much control over
information, and when there is entropy in communication protocols.
When business rules are lost, the opportunity to take advantage of the
information resources is squandered as well.
1.2.3 Why Data Quality Is the Pivot Point for Knowledge Management
The opportunity to take advantage of the data and information re-

source can only be enabled if there is an understanding of the structure
and knowledge about the collections of information. The critical point
is that a formal method is needed for collecting, documenting, and vali-
dating business rules. This methodology revolves around ensuring that
the information that is present in the system meets or beats the expecta-
tions of what is in the system. Ensuring data quality is a process of stat-
ing information requirements followed by a process of validating that
those requirements are being met.
A significant effort is made today to bring data out of the transac-
tional framework and into an operational data store that can be used
for analytical processing. These operational data stores are embodied in
data marts and data warehouses, which are useful knowledge manage-
ment tools when used effectively. A major component of the data ware-
house process is the extraction and transformation of data from source
and legacy data systems into the target data warehouse. This extraction
process is an inflection point at which business rules can both be discov-
ered and used for ensuring enterprise-wide data quality.
Based on the growing awareness of data quality as an enterprise
responsibility, there is a burgeoning need for inline data quality policy
10 ENTERPRISE KNOWLEDGE MANAGEMENT
management. According to the Gartner Group, "It is critical for enter-

prises to develop a data quality program and ensure that it is carried
out. Key to this effort is identifying data stewards^ in end-user areas
where data ownership is clearly defined. . . . Enterprises can minimize
these data inconsistencies by better understanding the parameters gov-
erning the meaning and movement of data."^
The use of data quality management as a tool for knowledge man-
agement along with these definitions are the initial motivators for this
book. But we won't stop at that! The next section gives nine reasons for
caring about data quality.
1.3 REASONS FOR CARING ABOUT DATA QUALITY
The data quality problem is pervasive in organizations across all indus-

tries. Bad data cost money and reduce productivity time spent diag-
nosing and fixing erroneous data is time not spent productively. Low
data quality eventually leads to reduced customer satisfaction. For
example, customers exposed to incorrect reports or statements are less
likely to trust the organization providing those reports. Finally, strategic
decisions based on untrustworthy information are likely to result in
poor decisions.
1.3.1 Low Data Quality Leads to Operational Inefficiency
The use of a manufacturing chain assumes that multiple stages are asso-
ciated with the final product, and at each stage there is an expectation
that the partially completed product meets some set of standards. Infor-
mation processing is also a manufacturing chain pieces of informa-
tion flow into and out of processing stages where some set of operations
are performed using the data.
To continue this analogy, when a product developed on a manufac-
turing chain does not fit the standards required at a specific stage, either
the product must be thrown away or fixed before it can continue down
the manufacturing line. Information is the same way: When a data
2. Explored later in this book.

3. J. Hill, and S. Laufer, Data Transformation: Key to Information Sharing, Gart-
ner Group Strategic Analysis Report, September 29,1998.
INTRODUCTION 11
record is found to be incorrect, the record needs to be deleted or fixed

before the processing can continue. Sometimes this "break" means the
delay of the entire processing stream, although it is more likely that the
records will be shunted aside and the stream continued on the next set
of records, with the erroneous records being dealt with at a later time.
As the level of data quality decreases, the more frequent the breaks in
operation. As more employees are allocated to fixing and reconciling
incorrect records, the rate at which information is processed decreases
in other words, operational inefficiency. This inefficiency is manifested as
error detection, error correction, and rework. We will look at these issues
more closely in Chapter 4.
1.3.2 Low-Quality Data Constrains Decision Making
Information production can be used for either operational processing

or analytical processing. The same information can be used for both
purposes. Yet, if the data are used for analytical processing or decision
support, the quality of the data can affect the analysis. If senior man-
agers rely on the results of the analysis, they may rely on conclusions
drawn from faulty assumptions. If these same managers are aware of
the low quality of the input data, they may choose to delay making
decisions until better information can be collected or until the same
information can be improved.
The same factors that cause delays in decision making can also
produce constraints. When one unit of an organization depends on the
results of analytical processing or decision support analysis from
another unit, the level of quality of the source information will affect
the first unit's ability to take action on the analysis and may cause an
erosion of trust between the two units.
For example, an airline may use an online reservation system as an
operational process as well as the basis for analytical processing. The
information may be targeted for use in determining what kinds of fre-
quent flyer promotions to create and to which sets of customers to
make these offers. If the data is of low quality, the conclusions pre-
sented by an analysis may not be trustworthy, and managers making
any decisions about the new promotions may be hesitant if they cannot
rely on those conclusions.
1.3.3 Good Data Enhances Data Warehouse Utility
According to an article presented by the Data Warehousing Institute,"^ a

survey conducted by survey.com ("Database Solutions III") estimates that
the worldwide data warehousing market is growing at the rate of 43% a
year, and will reach $143 billion by 2003, with sharp growth outside of
North America. Additionally, it is commonly recognized (and attributed
to Data Warehousing pioneer Bill Inmon) that 80% of the effort of a data
warehousing project is spent in extracting, cleansing, and loading data.
Considering this significant investment as well as the fact that data
warehouses are used for analytical processing, anything that improves
the customer's ability to analyze data increases the value of the data
warehouse. A major component of the data warehouse solution process
is the extraction and transformation of data from a legacy source before
the warehouse is populated. If the information in the warehouse is of
poor quality, a significant amount of time is spent in tracking and
removing errors. And since many warehouses are populated on a short
period basis (such as completely reloaded daily, with hourly sweeps), if
the error correction time frame exceeds the refresh period, the ware-
house would be nothing but a white elephant.
This stage of data population is the best opportunity to include data
quality validation and standardization, since good data in the warehouse
enables its use. We will focus on the data warehouse certification process
that uses a data quality rules engine in Chapters 7, 8, and 12.
1.3.4 Bad Data Leads to Incorrect Conclusions
Just as a house built on a weak foundation cannot stand, conclusions

based on incorrect input will not withstand scrutiny. As more and more
data warehouses are being built for critical business analytical applica-
tions, the criticality of ensuring data quality increases. Analytical results
based on bad data will be bad results period.
Not only that, operational decisions that rely on poor data can
cause inefficiencies in application systems. For example, load-balancing
accesses to a database based on a distribution of data in one attribute
can lead to skewed balancing if half of the records referenced have an
empty index field!
4. http://www.dw-institute.com/whatworks9/Resources/warehousing/warehous-
ing.html
INTRODUCTION 13
1.3.5 Bad Data Lead to Customer Attrition
Have you ever been billed for service that you have not received? Or
have you been threatened with being reported to a credit bureau for
being late on a payment on a balance due of $0.00? Many people have
some nightmare experience with which they can relate, always associ-
ated with some incorrect information that causes pain in the pocket-
book. These stories always seem to end with the customer ending his or
her relationship with the vendor or product provider over the matter.
These errors are typically due to some mistake on behalf of the ser-
vice or product provider, whether it is in customer records, customer
billing, product pricing, or during data processing. No matter what, the
problem is worsened by the fact that it is apparent that the organization
at fault has no evident means of proactive error detection in place. This
conclusion may be drawn because it is the customer who is doing the
error detection. We can claim that while a significant expense is made to
acquire new customers, it is worthwhile to invest the time and money
into improving the data collected on the current customers, since cus-
tomer attrition may be tied directly to poor data quality.
1.3.6 Good Data Enhances New Customer Acquisition
Just as poor data foster mistrust among current customers, it also can
cast doubt in the minds of potential customers. When a potential cus-
tomer is presented with an offer backed by high-quality data, the image
of the seller is enhanced, which can improve the opportunity to turn a
potential customer into a real customer.
As an example, consider this real-life pitch we recently received in
the mail. My wife and I recently took a trip with our 5-month-old baby.
We purchased a seat for our child at a discount rate because of her
young age. About a month after our trip, our baby received a letter
from the airline and a major long-distance carrier, offering her 10,000
frequent flyer miles if she switched her long-distance service. Clearly,
this cooperative sales pitch, cobbled together between the airline and
the long-distance carrier (LDC), may be effective some of the time, but
consider this: The airline knew that we had bought a discount seat for
our baby, and typically babies don't have authority to make purchasing
decisions in a household. In addition, both my wife and I have received
the same pitch from the same airline-LDC combination on the same
day! Because we saw that the long-distance carrier was unable to keep
their records straight about who we were, we weren't sure that we

could trust them to handle our telephone records correctly either.
1.3.7 Poor Data Quality Leads to Breakdown

in Organizational Confidence
On behalf of both customers and employees, when an organization dis-

plays an inability to manage simple information management issues, it
sows a seed of doubt that the organization can manage critical processes.
When customers lose faith in a provider's ability to provide, it leads to
customer attrition. When employees no longer believe in the company's
ability to do business, it leads to employee turnover and loss of strategic
business knowledge.
1.3.8 Bad Data Restricts System and Data Migration Projects
Having been involved in some legacy migration projects, I can say from
direct experience that the most frustrating component of a migration
project is the inability to accumulate the right information about the
data and systems that are being migrated. Usually, this is due to the ten-
dency of implementers to programfirst,document later, if at all. But as
systems age, they are modified, broken, fixed, or improved but without
any updates to documentation.
This situation forces the integrators to become information archae-
ologists to discover what is going on within the system. Naturally, undi-
rected discovery processes will increase costs and delay the actual
implementation, and the amount of time needed to determine what is
going on cannot be predicted ahead of time.
1.3.9 Good Data Increases ROI on IT Investment
When we have assessed the data quality requirements of our system and
put in place the right kinds of processes to validate information as it
passes through the system, we can limit the downtime and failed
processes based on low data quality. In turn, this means that without hav-
ing to diagnose and fix data quality problems, processing can proceed
INTRODUCTION 15
with a greater bandwidth, which subsequently allows for an increase in

processing volume without an increase in resources. In a company whose
business depends on increasing processing volume without increasing
overhead (such as securities processing or order fulfillment), this can
increase the return on information technology investment.
1.4 KNOWLEDGE MANAGEMENT AND BUSINESS RULES
Business operations are defined using a set of rules that are applied in
everyday execution. When the business depends on the correct flow of
information, there is an aspect of data quality that intersects the opera-
tional specification.
In essence, in an information business, business rules are data quality
rules. This implies that data quality is an integral part of any operational
specification, and organizations that recognize this from the start can
streamUne operations by applying data quality techniques to information
while it is being processed or communicated. This in turn will prevent bad
data from affecting the flow of business, and denying the entry of incor-
rect information into the system eliminates the need to detect and correct
bad data. Because of this, a "data quality aware" operation can execute
at lower cost and higher margin than the traditional company.
1.4.1 Competition in Knowledge-Intensive Industries
Businesses that traditionally rely on the leverage of timely information

(retail securities sales, market data providers, analytical market analy-
sis) suddenly have competition from players whose barrier to entry is
low. For example, in the retail securities investment business, the com-
bination of low-cost "electronic" securities trading combined with the
widespread availability of financial and investment recommendations
demonstrates that margins on securities transactions can be lowered,
but businesses can still profit.
As the availability of online analysis increases, there will be a de-
crease in the need for a large sales staff as a distribution channel for stan-
dard financial products. In the future, the perceived value of the
information provided by an investment analyst will not necessarily be
based on "brand name," but instead, the business rules will focus on
presentation, accuracy and timeliness, which are critical data quality

dimensions.
1.4.2 Micromarketing
Businesses that traditionally rely on marketing analysis will find that

"micromarketing" analyses will provide opportunities for micromarket
penetration. Businesses that have not traditionally relied on marketing
analysis are beginning to use information gleaned from online transac-
tions. The rules for accumulating and merging these data sets are man-
aged as business rules. The capability provided by desktop access to
data sources (such as consumer databases) allows more information to
enhance the analysis. This strategic analysis can only be trusted if the
data on which it is based are of high quality.
1.4.3 Cooperative Marketing Arrangements
With increasing frequency, organizations from different industries join

forces in cooperative or cross-linked marketing programs. A common
example is an offer of frequent flyer miles coupled with a change in
long-distance service. Because of the potential disparity between differ-
ent organizations' data models and the currency of their stored data,
these arrangements provide strong opportunities for data quality
improvements. The rules for value-added information merging, which
incorporates both known data and inferred information, can be encap-
sulated as business/data quality rules.
1.4.4 Deregulation in the Communications, Energy,

and Financial Services Industries
The recent consolidation of the insurance and brokerage industries in

anticipation of deregulation and the (expected) waves of deregulation in
the conmiunications and the energy industries pave the way for consoli-
dation of multiple business entities. When two (or more) independent
organizations merge their operations, their combined data resources need
to be combined as well. A streamlined data merging, migration, and qual-
ity assurance methodology will uncover synergistic opportunities when
INTRODUCTION 17
combining data sets. Leverage in affecting the organizational bottom line

can be obtained through the qualified merging of data resources by
decreasing operational costs (for example, error detection and correction)
and by increasing customer response and customer satisfaction.
1.4.5 The Internet as a Knowledge Transport Medium
The rapid ascension of the Internet as a business medium has created

opportunities for industries based on information and knowledge. When
the business relies on the use of information, the quality of that informa-
tion suddenly takes on a much higher visibility. This is an entirely differ-
ent dimension of applicability for the data quality business, since the
growing boom of "e-commerce" cuts across all major industries.
The World Wide Web can be can be seen as the largest uncoordi-
nated database service in the world. All Web sites can be characterized
in terms of some database service. Some sites act primarily as data pre-
sentation systems, others act as data collection systems, and some act as
database query systems. Success in the arena of electronic commerce
will be largely dependent on customer-focused micromarketing and
automated customer relationship management and sales force automa-
tion. We will provide guidelines for both the acquisition of valid data
over the Internet, as well as guidelines for accurate, customer-focused
data presentation.
1.5 STRUCTURE OF THIS BOOK
1.5.1 Chapter 1: Introduction
This chapter is an introduction to the ideas behind enterprise knowl-

edge management and the importance of data quality in this endeavor.
In this chapter, we enumerate the most important issues regarding
information quality, how business is aversely affected by poor data
quality, and how business can be improved when data quality is high.
1.5.2 Chapter 2: Who Owns Information?
Because the data quality problem ultimately belongs to the data con-
sumer, it a good idea to start out by establishing ownership and bound-
aries. Chapter 2 focuses on the issues of data ownership. The chapter
begins by discussing the data processing activity as a manufacture of
information. The final product of this factory is knowledge that is
owned by the data consumers in a business enterprise.
Who are the data producers and data consumers in an enterprise?
We look at internal data producers (internal processes like account
opening, billing, marketing) and external data producers ("lead lists,"
consumer research, corporate structure data). We also look at the enter-
prise data consumers, ranging from the operational (customer service,
billing, resource planning), the tactical (middle management, schedul-
ing), and strategic consumers (directional management, strategists).
There are complicating notions with respect to data ownership.
The means of dissemination, collecting data from the public domain, as
well as acquiring data from data providers confuse the issues. There-
fore, a set of ownership paradigms are introduced, including decision
makers, sellers, manipulators, guardians, and workers. These para-
digms bound the "social turf" surrounding data ownership. Finally,
Chapter 2 focuses on a finer granularity of ownership issues, including
metadata ownership, governance of storage and repositories, and
accountability for data policies.
1.5.3 Chapter 3: Data Quality in Practice
In Chapter 3, the notion of data quality is introduced, loosely defined

through the term "fitness for use." Background in traditional quality
systems is provided. Chapter 3 also provides some insight into informa-
tion theory as it can be applied to data quality. Finally, statistical
process control is discussed as critical to improving data quality.
1.5.4 Chapter 4: Economic Framework for Data Quality

and the Value Proposition
The rationale for designing and building data quality systems seems
logical, but an economic framework that can be used to measure the
INTRODUCTION 19
effects of poor data quality while highlighting its benefits is needed to

demonstrate the value of improving poor data quality. Chapter 4
reviews the knowledge hierarchy and the knowledge manufacturing
process, then defines a set of impact domains that can be used to model
the costs and missed opportunities that result from bad data. The chap-
ter focuses on strategic, tactical, and operational impacts, followed by a
description of a way to model these impacts.
1.5.5 Chapter 5: Dimensions of Data Quality
In Chapter 5, a number of dimensions of data quality are discussed.

These dimensions are grouped into quality of the data model, quality of
data values, quality of data presentation, and other data quality dimen-
sions.
1.5.6 Chapter 6: Statistical Process Control

and the Improvement Cycle
In Chapter 6, we look in detail at the notion of control and how that

idea can promote predictability of the quality of the product being pro-
duced. Statistical process control is a method for measuring the quality
of the product while its process is in operation, instead of relying on
measuring the product after it has been completed. SPC is a process of
gathering measurements during processing in order to identify special
causes of inconsistencies and variations that lead to a flawed product.
We will focus in particular on the notion of control charts and how
these charts are used to look for processes that are in control and
processes that are out of control. In addition, we will look at how con-
trol charts are used for setting up an improvement cycle with respect to
conformance with data quality specifications.
1.5.7 Chapter 7: Domains, Mappings, and Enterprise Reference Data
In this chapter, we begin to explore the ideas revolving around data types
and how data types are related to the notion of sets. We then describe
our definition of data domains, both descriptive and enumerated. Next,
we discuss the relations between domains, how those relations exist in
so ENTERPRISE KNOWLEDGE MANAGEMENT
databases, and the power of abstracting these mappings as reference

metadata. Finally, we propose a publish/subscribe model for the man-
agement and use of enterprise reference data.
1.5.8 Chapter 8: Data Quality Assertions and Business Rules
In this chapter, we really begin the investigation of the kinds of rules

that can be applied to the measurement of data quality. With respect to
any tabular data (such as database tables), we will look at these classes
of rules: (1) data value rules, (2) attribute rules (rules for data values
associated with a particular attribute), (3) domain rules, (4) tuple rules
(rules associating different attributes within a tuple), and (5) table rules.
We identify a number of rules that can be characterized by a measurable
method.
1.5.9 Chapter 9: Measurement and Current State Assessment
We look at measuring conformance to the rules described in Chapter 8

in two different ways. The first way is a static measurement, which
involves looking at the data after it has been delivered into its target
location. The second way is in-process (dynamic) measurement, which
is consistent with the notions of statistical process control as described
in Chapter 6. This way involves integrating the measurement of the
conformance to the data quality rules while the data is in transit.
The goal of the initial measurement is to gauge the degree of any
data quality problem as well as use the economic framework to mea-
sure the effects of any poor data quality. The current state assessment is
a report that focuses on the important dimensions of data quality, the
degree on conformance, and how the lack of conformance affects the
bottom line. The result of the CSA is a characterization of the state of
data quality and what needs to be done to improve it.
1.5.10 Chapter 10: Data Quality Requirements
How do we determine data quality requirements? We use a technique

borrowed from object-oriented design called use-case analysis, which is
used to specify the system in terms of actors, use-cases, and triggers.
INTRODUCTION 21
Actors represent the roles that users play, use-cases represent what the
actors do with the system, and triggers represent events that initiate use
cases. We then select from the list of data quality dimensions from Chap-
ter 5 those dimensions that are of greatest importance to the actors and
define data quality rules, as described in Chapter 8. We can choose
thresholds for conformance to the data quality rules as a baseline for
acceptance of the data. These baseline thresholds are defined so that
when met, the data set consumers can be confident of the levels of data
quality.
1.5.11 Chapter 11: Metadata, Guidelines, and Policy
When defining data quality rules and requirements, it is necessary to

build a central repository for the collection of all metadata, which is
information about the data. All data processing systems maintain some
form of metadata, whether it is explicit using a DBMS system's data dic-
tionary, or implicit, such as the methods and inheritance embedded in a
C++ class definition.
The traditional view of metadata includes mostly static informa-
tion about the data, such as data types, sizes, and perhaps some rudi-
mentary rules such as enforcing a primary key or indexing. Metadata,
though, could and should encompass much more, including data qual-
ity and business inteUigence rules that govern the operational manufac-
ture and use of data. This is not to say that this kind of metadata isn't
being maintained now because a lot of it is! This metadata, however,
is maintained in a format unsuitable for easy access and understanding
by the actual data consumer.
1.5.12 Chapter 12: Rules-Based Data Quality
Having defined an assertional framework in Chapter 8, how would data

integrity assertions be defined and validated.^ The answer is through the
use of rule-based systems that a user employs to both define and test
rules. In Chapter 12, we describe rule-based systems: definitions of
"rule" and rule-based system and how we incorporate our assertion sys-
tem from Chapter 8 into an information integrity language. Last, we
describe how a rules engine can use these validation specifications to
provide a test framework for both data integrity validation and the data
for Statistical process control, all within the context of the data con-
sumer's constraints.
1.5.13 Chapter 13: Metadata and Rule Discovery
Chapter 12 presents a rule system for defining and testing information

validity. In Chapter 13 we discuss what we learn from the rule definition
and rule-testing process. Included is domain discovery, association rules,
map discovery, functional dependencies, overlapping and intersecting
domains, and attribute splitting. These all form a link in a feedback
chain in the metadata analysis and data quality improvement process.
1.5.14 Chapter 14: Data Cleansing
Understanding data quality rules, assertions, and validity checks is one

thing; actually "fixing" the problems is another! Chapter 14 focuses on
techniques for cleansing data to the point where it meets or exceeds the
conformance thresholds and is therefore acceptable. We start with
metadata cleansing identifying and documenting what we know
about our data. We look at using our assertions to form information fil-
ters that prevent bad data from entering the system.
When identifiability constraints specify uniqueness, it is worthwhile
to eliminate duplicate entries. This takes on two faces: absolute duplicates
and near-match duplicates. To address the latter, we discuss approximate
searching and matching techniques. These techniques are also useful in
householding, a special kind of duplicate elimination process that looks
to accumulate information about individuals sharing a single residence.
When there is a standard form for an information set (as specified
by a structured domain or such as the United States Postal Service stan-
dard), we use a technique called standardization to make our data to
look the way we want it to. We will discuss issues surrounding the
opportunity for cleansing data during a data migration process.
1.5.15 Chapter 15: Root Cause Analysis and Supplier Management
One technique not yet discussed is using the results of the information
validity exercise to prevent the continuation of low data quality events.
INTRODUCTION 23
When using the rule-based system for validity checking, we can use the
information in the reports to look for root causes of the occurrences of
bad data quality. This technique is the last Unk in our improvement chain,
since fixing the sources of bad data will directly improve overall data
quality.
1.5.16 Chapter 16: Data Enrichment and Enhancement
Data enrichment is a process of enhancing the value of a pool of informa-

tion by combining data from multiple sources. The goal of enrichment is
to provide a platform deriving more knowledge from collections of data.
A simple enrichment example is the combination of customer sales data
with customer credit data to build a data mart that can be used to make
special offers to preferred customers (customers with high purchasing
profiles and good credit profiles). Other examples would be merging
health insurance claim information with professional billing information
to search for insurance fraud and the enrichment of financial electronic
data interchange messages to enable straight-through processing.
In any case, data quality is critical to successful enrichment. In
Chapter 16, we discuss how data standardization, clustering techniques,
and the use of data quality rules can be used to express enrichment direc-
tives. We also look at some enrichment examples and how they can be
affected by poor data quality.
1.5.17 Chapter 17: Data Quality and Business Rules in Practice
In Chapter 17, we review our data quality rules in a context that

demonstrates ways to actually make use of these rules. Specifically, we
talk about the way these rules partition the data into two sets, the set of
conforming records and the set of nonconforming records. There is
already a means for specifying set partitions in databases the query
language SQL, which is not just used for performing operations on
databases but to show how to turn the rules into executable queries.
The rest of this chapter is designed to show how the implementation of
a data quality rules process can add significant value to different operat-
ing environments.
1.5.18 Chapter 18: Building the Data Quality Practice
Ourfinalchapter is basically a "run book" for building a data quality

practice. We walk step by step through the issues in building the prac-
tice, starting with problem recognition, followed by the way to gain
senior-level support, adopting a methodology, and creating an educa-
tion program as well as actual implementation.
2
WHO OWNS INFORMATION?
Before we delve into the details of what data quality means and how it
relates to knowledge management, we should establish where the respon-
sibility for data quality falls within a company. Without a clear assign-
ment of accountability, it is almost impossible to measure the quality of
data, much less effect improvements.
This chapter examines the question of data ownership as the first
step in establishing a knowledge-oriented organization. We begin by dis-
cussing data processing activity as a manufacture of information, and
knowledge, which is owned by the data consumers in a business enter-
prise and is the final product of this factory.
Who are the data producers and data consumers in an enterprise?
We look at internal data producers (internal processes like account
opening, billing, marketing) and external data producers ("lead lists,"
consumer research, corporate structure data). We also look at the enter-
prise data consumers, ranging from the operational (customer service,
billing, resource planning), the tactical (middle management, schedul-
ing), and strategic consumers (directional management, strategists).
There are complicating notions with respect to data ownership.
The means of dissemination, collecting data from the public domain, as
well as acquiring data from data providers confuse the issues. There-
fore, a set of ownership paradigms is defined, including decision mak-
ers, sellers, manipulators, guardians, and workers. These paradigms
bound the "social turf" surrounding data ownership.
Finally, we try to resolve some of these issues by investigating data
policy paradigms. This includes metadata ownership, governance of
storage and repositories, and accountability for data policies.
25
2.1 THE INFORMATION FACTORY
A relatively simple analogy for processing data that we use throughout the
book is the information factory. Any information processing activity can
be viewed as a small factory that takes some data as raw input, processes
that input, and generates some information result, potentially generating
data by-products and side effects in the process. Inside the factory there
may be smaller subfactories, each with its own input/output production
activity. The raw input data are provided by data suppliers external to the
organization or by data manufacturers within the organization. The ulti-
mate data customers may be internal or external consumers.
2.1.1 Actors in the Information Factory
To be more precise, let's look at the different roles that exist in the con-
text of the information factory. These roles may represent real people or
automated proxies within the system.
1. Suppliers: Data suppliers provide information to the system.
2. Acquirers: Acquirers accept data from external suppliers for
provision into the factory.
3. Creators: Internal to the factory, data may be generated and
then forwarded to another processing stage.
4. Processors: A processor is any agent that accepts input and gen-
erates output, possibly generating some side effects.
5. Packagers: A packager collates, aggregates, and summarizes
information for reporting purposes.
6. Delivery Agents: A delivery agent delivers packaged informa-
tion to a known data consumer.
7. Consumer: The data consumer is the ultimate user of processed
information.
8. Middle Manager: The people responsible for making sure the
actors are correctly performing their jobs.
9. Senior Manager: The senior manager is responsible for the over-
all operation of the factory.
10. Deciders: These are senior-level managers associated with
strategic and tactical decision making.
Each of these actors plays a well-defined role in the data processing
operation, and each is responsible at some level for quality assurance
WHO OWNS INFORMATION? 27
within each activity domain. In a perfect world, these responsibiUties

will all propagate up to the enterprise level, providing some degree of
quality assurance overall, but in reality there are complicating factors
that may prevent this from happening. It is clear, though, that at any
stage of processing, it is difficult to specifically assign ownership to the
information being created or processed.
2.1.2 Processing Stages
In any data processing system, it is helpful to be able to break down the

entire information flow into a series of processing stages, most of which
relate directly to the activity associated with one of the previously enu-
merated actors. Ultimately, we would like to be able to precisely iden-
tify all input and output information streams, as well as all affected
data stores associated with each processing stage, and then associate
responsibility with a particular actor.
When truly decomposed, the information manufacturing process
contains many different stages, each of which in itself might represent
an entire instance of an information factory. We end up with a hierar-
chical description of the information processing chain.
2.1.3 Data Producers
Data producers are those organizations that create, compile, aggregate,

package, and provide information to be inserted into an information
processing system. This includes organizations that generate internal
data (audit data, workflow messages, intermediate processed data for
incorporation into other processing stages) or external information,
such as marketing data, sales reports, invoices, corporate structure
data, credit reports, and so forth.
2.1.4 Data Consumers
Data consumers can be categorized into three groups: operational data

consumers, tactical and strategic data consumers, and external cus-
tomers. Operational data consumers are manifested as the role of
processors described in Section 2.1,1. They are any internal processing
Stage that requires input to operate, including any transaction process-

ing activity, message passing or routing, or workflow activities.
The tactical and strategic data consumers are those who use pro-
cessed information to make tactical and strategic decisions. This includes
sales management, marketing, enterprise resource planning, mergers and
acquisitions, and so on.
The external customers are those who receive information processed
by the information factory. The kinds of information in this category
include invoices, customer billing, sales teams data, geographic data, and
data provided by government agencies.
2.2 COMPLICATING NOTIONS
What complicates the ownership question is that there are factors

orthogonal to data creation or consumption that create real or artificial
boundaries around information. The ownership issue is essentially a
control issue control of the flow of information, the cost of informa-
tion, and the value of information.
Here is an example of how control over the flow of information
can have a major impact. Some government interest rate financial data
was mistakenly released early, prompting trading associated with sell-
ing bonds on the market before the official release time. The same
traders, by buying back the same bonds at a discount after the official
release time, made a huge profit. In this case, one party's loss of control
over the flow of information allowed other parties increased control
over the value of that same information!
A few of the many issues that complicate the notion of data owner-
ship are value, privacy, turf, fear, and bureaucracy. This list is by no
means inclusive, and we make no attempt to resolve them only to
expose them.
2.2.1 Value
The value of information drives a particular wedge into the question of

ownership. In any environment where there is shared data ownership,
how does the degree of ownership relate to the responsibility of care of
that same information? Presumably, the degree of ownership may be
related to more mundane aspects of the system, such as who initially
created the database or what silo currently manages the system. But at
the core, the degree of ownership (and by corollary, the degree of
responsibility) is driven by the value that each interested party derives
from the use of that information.
2.2.2 Privacy
The issue of privacy as a complicating notion could take up an entire

volume, if not more. What information should remain private, and
under what circumstances? If a party willingly releases private informa-
tion under one set of conditions, does that allow the receiving party the
right to use that information in other situations? Consider credit infor-
mation. The fact that someone applies for any particular credit card
might be considered private information, although in order to receive
that credit, a credit bureau must be consulted. At that point, the fact
that the credit card application has been taken is now added to the
credit record.
And once information is released from one party to another, who
exercises control over that information? For example, if a pharmacy
fills an individual's medicine prescription, can the pharmacy report
back to the pharmaceutical company the information regarding which
doctors prescribed the medication, which patients filled the prescrip-
tions, and which came back for refills?
When private knowledge can be inferred from public data, does
this violate any privacy constraints? For example, the amount of money
one borrows on a mortgage to pay for a house might be considered a
private matter between the borrower and the bank, but in fact this
information is lien information that is frequently filed under various
state codes and is not private information at all.
2.2.3 Turf
On a different level, the question of data ownership within an enter-

prise is often complicated because of the notion of "turf." In many
organizations, the control of the flow of information is regarded, as are
many other forms of control, as a means of job security. "As long as I
am in charge of this report," thinks the middle manager, "and no one
else has access to the data, I can't be fired!" Being in charge of creating,
packaging, and distributing the report naturally leads one to the con-
ception of owning the data that makes up the report.
2.2.4 Fear
As organizational employees carve out their little fiefdoms of control,

the overall ability of the organization to react to inconsistencies flowing
out of these fiefdoms decreases. Any suggestion that there is a reason
for an "outsider" to assess or analyze the current state of the data is
taken as an attack on the fiefdom's turf and therefore can be construed
as a direct challenge to the data controller's job.
An even worse fear is that any closer viewing of what goes on within
one's organization will reveal that what appeared to be stellar work is
actually mediocre, or worse. People can be so desperate to conceal their
own mistakes that they will sabotage any attempt to uncover them. Of
course, this is a prime example of conflict within the enterprise what is
good for the individual is terrible for the organization, and vice versa.
2.2.5 Bureaucracy
Another major issue that complicates data ownership is institutional

bureaucracy. As organizations grow, the intermediate management
structure grows as well, thereby diluting the responsibility for informa-
tion as it passes from one subsystem to another. When issues regarding
data quality problems arise, typically there tends to be a lot more finger
pointing than problem solving.
This may be due to the fact that organizations become divided
along project or system lines. Because of this, attribution of problems
associated with information that passes through individually managed
processing stages is hard to pin down, since each manager will pass the
buck further upstream.
Another bureaucratic issue involves decisions associated with up-
grades, renovations, and changes in the existing data processing infra-
structure. With a highly hierarchical organization, the ability to make a
decision hinges on building consensus among all interested parties both
across vertical departmental lines as well as up and down the manage-
ment hierarchy. Obviously, the effort involved is significant and, com-
bined with the turf and fear factors, may account for the failure of
many enterprise infrastructure renovation projects.
2.3 RESPONSIBILITIES OF OWNERSHIP
What do we mean when we talk about ownership of data? The essence

Ues in the control of information as an enterprise asset. That control
includes not just the ability to access, create, modify, package, derive
benefit from, sell, or remove data but also the right to assign these
access privileges to others. In this section, we discuss in greater detail
some of the responsibilities associated with data ownership.
2.3.1 Definition of Data
In any data environment, the data owner is responsible for understand-

ing what information is to be brought into a system, assigning the
meanings to collections of data, and constructing the data model to
hold the collected information. In addition, any modifications to the
data model and any extensions to the system also fall under duties of
the data owner.
2.3.2 Authorization of Access and Validation of Security
A major concern for any data system is the coordination and authoriza-
tion of access. In a system that contains data that is in any way sensi-
tive, whether it is confidential information, human resource data, or
corporate intelligence, it is necessary to define a security and authoriza-
tion policy and to provide for its enforcement.
2.3.3 Support the User Community
When information is provided to a user community, there is a responsibil-

ity to provide support for those users. This includes providing accessibil-
ity to new users, granting them access rights, providing documentation,
training, and addressing technical needs and questions. This also includes
defining and maintaining a service level agreement, which may entail

measuring system performance, and scaling or rebalancing resources to
provide the agreed upon level of service.
2.3.4 Data Packaging and Delivery
In addition to standard user support, the owner also holds the responsi-
bility for providing the data to the data consumers. This may include
data preparation, packaging and formatting, as well as providing a
delivery mechanism (such as a data portal or a publish/subscribe mech-
anism).
2.3.5 Maintenance of Data
Aside from the maintenance of the system itself, there is also the mainte-
nance of the information. This includes managing the data input
process, instituting gauges and measurements associated with the data,
and creating data extraction and loading processes.
2.3.6 Data Quality
The data owner is also accountable for maintaining the quality of the
information. This may include determining and setting user data qual-
ity expectations, instituting gauges and measurements of the levels of
data quality, and providing reports on the conformance to data quality.
This also includes defining data quality policies for all data that flows
into the system and any data cleansing, standardization, or other prepa-
ration for user applications.
2.3.7 Management of Business Rules
All data processing operations have business rules. Whether these rules
are embedded in application code, abstracted into a rules format, or
just documented separately from their implementation, the data owner
is also responsible for managing business rules.
2.3.8 Management of Metadata
Managing metadata involves the data definitions, names, data types, data
domains, constraints, applications, database tables, reference reposito-
ries, and dependence rules associated with different tables and databases,
users, access rights, and so forth.
2.3.9 Standards Management
Whenever information is shared between two or more parties, there

must be some agreement as to a format for that data. When multiple
parties agree to a representation format, that format is defined as a data
standard. The owner is also responsible for making sure that all rele-
vant data sets conform to their standard form, as well as negotiating
standards on behalf of the users.
2.3.10 Supplier Management
When data sets are built as a composition of supplier-provided data, the

data owner is also responsible for supplier management. This involves
negotiating arrangements with each supplier, determining data delivery
agreements, defining sets of data quality criteria, and enforcing these
requirements and arrangements with each supplier.
2.4 OWNERSHIP PARADIGMS
We can enumerate owner responsibilities, but that does not solve the prob-
lem of assigning (or declaring) data ownership. Instead of trying to pro-
actively dictate an ownership model, it is more helpful to explore different
existing ownership paradigms. In each one of these paradigms, we will
look at the question of value and how it relates to the claim of ownership.
2.4.1 Creator as Owner
In this paradigm, the party that creates or generates the data owns the
data. It represents a speculative investment in creating information as a
prelude to recognizing value from that information in the future.
An example of this is a geographic data consortium that analyzes

geographic regions, collects latitude/longitude measures, and enters
that information into a geographic database. The measurements in iso-
lation are essentially useless; it is the collection of all the measurements
that forms a useful data set. The consortium creates the information,
and most likely claims ownership of that data as well.
2.4.2 Consumer as Owner
This ownership paradigm indicates that the party that consumes the
data owns that data. This is a relatively broad ownership spectrum,
covering all aspects of data acquisition. In this paradigm, any party that
uses data claims ownership of that data. When the consumer requires a
high level of confidence in the data input into a process, this ownership
paradigm is very logical, since the party that cares most about the value
of the data claims ownership (and thus, responsibility). In this case, the
consumer derives the value from the data.
An example of this is a sales organization that uses information
provided from different organizations within a company. Once the data
lands at the sales staff's door, though, the information becomes integral
to the proper operation of the sales team, and so the sales team will
claim ownership of the data that it consumes.
2.4.3 Compiler as Owner
The operation of selecting information sources and compiling informa-

tion from these different sources constitutes an ownership model. By
combining data sets, the compiler is adding value and may expect to
reap the benefits of ownership.
A good example of data compilation is a news item retrieval com-
pany that provides, as a service, a search for newspaper articles. By col-
lecting and providing search capability, the data compiler has created a
body of information that is more valuable than the individual pieces
making up that body.
2.4.4 Enterprise as Owner
In larger corporate information technology organizations, there is a

notion that all data that enter the enterprise or are created within the
enterprise are completely owned by the enterprise. In effect, the com-
pany makes use of all input and generated data as fuel for its ongoing
data processing needs, and therefore the value derived from the infor-
mation resides with the organization as a whole.
The investment banking industry demonstrates this ownership
model when it accumulates information from external market data ven-
dors as well as data generated from internal sales and securities process-
ing. All data are absorbed into a single operational data center that then
redistributes the data, potentially with added value, out to data con-
sumers within the organization as well as individuals (such as clients
and customers) external to the enterprise.
2.4.5 Funding Organization as Owner
In this paradigm, the user that commissions the data creation claims
ownership. Here there are two parties involved: the one that pays for
the creation of data and the one that actually creates the data. In this
case, the patron claims ownership, since the work is being done on his
or her behalf.
An example is a company that commissions a research organiza-
tion to prepare a competitive intelligence report covering a particular
industry. The company may stipulate that the company is the sole
owner of the provided data.
2.4.6 Decoder as Owner
In environments where information is "locked" inside particular encoded

formats, the party that can unlock the information becomes an owner of
that information. The cost of the decoding process and implementations
is an investment in the value to be derived from the information.
A good example of this is embodied in the results of decoding DNA
sequences to isolate specific genes. The value of decoding the DNA struc-
ture can be expressed in terms of any improvement in the discovery,
prevention, or treatment of certain hereditary diseases. Bio-informatics
companies that decode genetic material can then sell the data that they
decode to the medical and pharmaceutical industries.
2.4.7 Packager as Owner
As opposed to the compiler ownership paradigm, the packager para-

digm focuses on the party that formats information for a particular use.
There is value added through formatting the information for a particu-
lar market or set of consumers.
An example of this is authors who publish public domain infor-
mation packaged as a book. The compilation process is most likely
straightforward the value added is in formatting the material to
make the collected data useful.
2.4.8 Reader as Owner
This is an interesting paradigm in that it implies that the value of any

data that can be read is subsumed by the reader, and therefore the
reader gains value through adding that information to an information
repository. The investment is in the reader's selection and consumption
of data.
For example, consulting firms establish expertise practices in par-
ticular areas. In order to become a principal in one of these practices, an
individual must acquire knowledge in the practice area by absorbing as
much information about that area as possible. Going forward, remain-
ing an expert requires active information gathering.
2.4.9 The Subject as Owner
This paradigm revolves around the subject data ownership issues, such
as personal privacy or image copyrights. In this view, the subject of the
data claims ownership of that data, mostly in reaction to another party
claiming ownership of the same data.
As an example of the privacy issue, consider a pharmacyfillingpre-
scriptions. Drug companies are interested in knowing which doctors are
prescribing their medicines, and doctors like to know which of their
patients are refilling prescriptions as a tool to see how well their patients
follow instructions. Recently, it was revealed that a national pharmacy

chain was providing to both health care providers and drug companies
detailed information about who was filling which prescriptions at each
of their sites. When this practice became public, naturally their cus-
tomers were incensed. In effect, the individual patients were claiming
ownership of personal information and insisting that the pharmacy
chain had no right to sell it.
Another example is the issue of corporate image and branding.
Companies will make a significant effort in establishing a connection
between the quality of the products or services that it provides and a cor-
porate image or logo. In this case, representations of the image are
equated with the company, and any misrepresentation or other unautho-
rized use could affect the branding, so the company claims complete own-
ership of the image (or logo) as well as protects the use of that image.
2.4.10 Purchaser/Licenser as Owner
Similar to the funder paradigm, the individual or organization that buys

or licenses data may stake a claim to ownership. In this paradigm, the
purchaser assumes that the investment made in acquiring the data
yields ownership. This holds for licensing as well, even if the terms of
the license specify some restrictions on use.
A good example is the sale of mailing lists for direct marketing
campaigns. One organization may own the lists and license their use,
but once the Usts are sold, the purchaser or licenser considers the data
its own.
2.4.11 Everyone as Owner
The final paradigm is the model of global data ownership. Some feel
that monopolization is wrong and data should be available to all with
no restrictions. Clearly, in the business world, this is a radical view, and
it has its benefits as well as detriments.
This ownership model is often in operation, to some degree, in sci-
entific communities, where experimentation, following by the publish-
ing of results, is common practice. In this situation, a common goal is
the increase in the global knowledge of a particular subject, and results
are subject to other experts' scrutiny.
2.5 CENTRALIZATION. DECENTRALIZATION. AND

DATA OWNERSHIP POLICIES
This section explores the issues regarding the distribution of ownership

across an enterprise. The question of centraUzation versus decentraUza-
tion is orthogonal to the responsibilities of ownership, yet it is distinctly
intertwined with it as well. As in our ownership model, the critical
point revolves around value.
In a centralized ownership model, there is a single entity (person or
group) responsible for all data ownership for the entire enterprise. Cen-
tralization implies that all ownership activities are coordinated from a
single point of control, as well as coordination of metadata, informa-
tion sourcing, and so forth. Centralized ownership yields the benefit of
the value added and whether the costs associated with centralization
are offset by it. The costs include the increased management overhead,
bureaucracy, and system integration, among others. The benefits include
enterprise standardization for data and systems, the ability to make use
of merged data for additional knowledge discovery, and increased lever-
age when dealing with external data suppliers.
In a decentralized model, the ownership roles are allocated to sepa-
rate areas of interest. A decision to opt for decentralization implies that
the value added from centralized control is more than offset by its asso-
ciated costs. On the other hand, most organizations do not explicitly
opt for decentralized control; instead, organizations evolve into it.
Therefore, the real question is whether migrating from a decentralized
ownership model to a centralized ownership model will increase the
value of the enterprise knowledge base.
Finding the answer to this question is not simple. It involves a
process of identifying the interested parties associated with all data sets,
determining each party's interest, identifying the different roles associ-
ated with all data sets, and assigning roles and responsibilities to the
right parties. All of these activities are embodied in an organization's
data ownership policy, which incorporates all governance rules regard-
ing data ownership and usage within the enterprise.
2.5.1 Creating a Data Ownership Policy
A data ownership policy is a tool used by the enterprise to establish all

roles and responsibilities associated with data ownership and account-
ability. The goal of a data ownership policy is to finesse the kinds of
complications discussed in Section 2.2, as well as hash out the strict def-
initions of ownership as described in Section 2.4. The data ownership
policy specifically defines the positions covering the data ownership
responsibilities described in Section 2.3. At a minimum, a data owner-
ship policy should enumerate the following features.
1 The senior level managers supporting the enforcement of the
policies enumerated
2. All data sets covered under the policy
3. The ownership model (in other words, how is ownership allo-
cated or assigned within the enterprise) for each data set
4. The roles associated with data ownership (and the associated
reporting structure)
5. The responsibilities of each role
6. Dispute resolution processes
7. Signatures of those senior level managers listed in item 1
A template for describing the ownership policy for a specific data set is
shown in Figure 2.1.
Data Set Name

Primary Owner
Data Set Location
Responsible Reports
Owner Party to Notes
Data Definition
Access/Security
User Support
Data Packaging
Data Delivery
Maintenance
Data Quality
Business Rules
Metadata
Standards Management
Supplier Management
FIGURE 2 . 1 Template for data ownership policy

These are the steps for defining a data ownership policy.

1. Identify the interested parties or stakeholders associated with the
enterprise data. This includes identifying the senior level man-
agers that will support the enforcement of the policy.
2. Catalog the data sets that are covered under the policy.
3. Determine the ownership models in place and whether these are
to continue or will be replaced.
4. Determine the roles that are and are not in place. Assign the
responsibilities to each role, and assign the roles to interested
parties.
5. Maintain a registry that keeps track of policies, data ownership,
roles, responsibilities, and other relevant information.
2.5.2 Identifying the Stakeholders
All stakeholders in the information factory, including all the actors

delineated in Section 2.1.1, should be considered interested parties. A
stakeholder is anybody who expects to derive some benefit or value
from the data, whether it is through the use of the data, the sale or
license of the data, or beneficially through association with the data.
For example, a business customer who uses the reports gets value
through the data, receives monetary compensation through the sale or
license of the data, and benefits from the jobs that may be dependent on
continued data center operations and application development.
In a small enterprise, stakeholder identification can be relatively
simple, but as the enterprise grows, the process can become extremely
complex due to the degrees to which information is processed and dis-
seminated. A good heuristic is to begin from the outside of the enter-
prise and work in. In other words, figure out who the end users are,
look at the data they are using, and follow it backward through the
information chain. While some business users may be outspoken in
terms of staking their claim, others may be blind to the fact that there is
any organizational process that generates the paper reports that land on
their desks. Also, just because people receive the reports, they may
never look at the data provided on a periodic basis.
The process of identifying the stakeholders will likely reveal areas
of conflict with respect to data ownership. This is a particularly valu-
able part of the process, as it provides a guide to deciding how the own-
ership responsibilities are assigned.
2.5.3 Cataloging Data Sets
Once the stakeholders have been identified, the next step is to learn
what data sets should fall under the ownership policy. The stakeholders
should be interviewed to register the data sets with which they are asso-
ciated and the degree to which each believes his or her stake in the data
is. The goal of this step is to create a create a metadatabase of data sets
to use in the enforcement of the data ownership policies. This catalog
should contain the name of the data set, the location of the data set, and
the list of stakeholders associated with the data set. Eventually, the cat-
alog will also maintain information about data ownership and respon-
sibilities for the data set.
2.5.4 Identifying and Assigning Roles
The next step is to determine the roles that are associated with each set
of data in the enterprise and describe the responsibilities of each role.
Here are some examples, although this list is by no means meant to be
exhaustive.
Chief Information Officer The CIO is the chief holder of
accountability for enterprise information and is responsible for deci-
sions regarding the acquisition, storage, and use of data. He or she is
the ultimate arbiter with respect to dispute resolution between areas of
ownership and is the ultimate manager of the definition and enforce-
ment of policies.
Chief Knowledge Officer The chief knowledge officer is responsi-
ble for managing the enterprise knowledge resource, which dictates and
enforces the data sharing policies, as well as overseeing the general
pooling of knowledge across the organization.
Data Trustee The data trustee manages information resources
internal to the organization and manages relationships with data con-
sumers and data suppliers, both internal and external.
Policy Manager The policy manager maintains the data owner-
ship policy and negotiates any modifications or additions to the data
ownership policy.
Data Registrar The data registrar is responsible for cataloging
the data sets covered under the policy as well as the assignment of own-
ership, the definition of roles, and the determination of responsibilities
and assignments of each role. The data registrar also maintains the data
policy and notifies the policy manager if there are any required changes
to the data ownership policy.
Data Steward The data steward manages all aspects of a subset
of data with responsibility for integrity, accuracy, and privacy.
Data Custodian The data custodian manages access to data in
accordance with access, security, and usage policies. He or she makes
sure that no data consumer makes unauthorized use of accessed data.
Data Administrator The data administrator manages production
database systems, including both the underlying hardware and the
database software. The data administrator is responsible for all aspects
related to the infrastructure needed for production availability of data.
Security Administrator The security administrator is responsible
for the creation of and the enforcement of security and authentication
policies and procedures.
Director of Information Flow The director of information flow is
responsible for the management of data interfaces between processing
stages, as well as acting as an arbiter with respect to conflicts associated
with data flow interfaces.
Director of Production Processing The director of production
processing manages production processing operations, transference of
data from one production source to another, scheduling of processing,
and diagnosis and resolution of production runtime failures.
Director of Application Development The director of application
development manages requirements analysis, implementation, testing,
and deployment of new functionality for eventual turnover to the pro-
duction facility.
Data Consumer A data consumer is an authorized user that has
been granted access rights to some data within the enterprise.
Data Provider A data provider is an accepted supplier of infor-
mation into the system.
These roles will then be integrated into a reporting structure where

there are clear lines of responsibility corresponding to degrees of own-
ership. Note that some responsibilities are assigned to multiple roles,
causing "role overlap," whose governance must be integrated into the
reporting structure as well. At this point, the senior manager responsible
for information (typically a chief information officer) will then assign
ownership roles and responsibilities to the different organizational
stakeholders.
2.5.5 Maintaining the Ownership Registry
The ownership registry is created from the data catalog and the assign-
ment of roles. It is the enterprise log that can be queried to determine
who has the ultimate responsibility for each data set. The ownership
registry should be accessible by all interested parties, especially when
new data requirements arise or there is a conflict that needs resolution.
Management of the ownership registry requires keeping a pulse on
the organization, as it is not unusual for employee turnover to affect the
data management structure. In addition, as new data sets are added to
the governance by the data ownership policy, the decisions regarding
the new data must be added to the registry.
2.6 OWNERSHIP AND DATA QUALITY
This brings us back to the issue of data quality. Once we have estab-
lished a chain of command for the ownership of data, we can look at
how the responsibility for data quality falls with respect to the policy. A
major factor is the relationship between ownership and care, which is
explored in this section, along with the enforcement of data policies and
an introduction to data quality rules.
2.6.1 Ownership and Care
This may seem pretty obvious, but a person demonstrates a greater

amount of care for an object that he or she owns than for something
that belongs to someone else. To this end, it is important to consider the
ownership stake associated with the parties in the enterprise responsible
for the quality of an organization's data. It is less likely that the quality
of data will be high when it is entrusted to someone who has no stake in
the value of the data.
To be more precise, there are two ways to incorporate the owner-
ship ideal with data quality: (1) Assign some degree of ownership to
those entrusted with the data quality, or (2) assign the responsibility of
data quality to the party with the highest degree of ownership.
In the first case, a person has been assigned the responsibility of
maintaining the integrity of a data set but does not feel any personal
attachment to the information in his or her trust. By allocating some
ownership role to that person, such as creating some kind of bonus

compensation structure tied to maintaining or increasing the overall
value of the data, an organization can infuse the employee with a
renewed interest in data quality.
The second case reveals another common ownership issue where
the end users (in many cases, business users) rely on the quality of the
data but have no stake in ensuring that quality. This can be remedied by
forcing the business users to get involved in the definition and assurance
of data quality. This includes understanding the data that is being used,
defining data quality rules that are the basis for acceptability, and
enforcing data policies and data quality rules.
2.6.2 Enforcing Data Policies
It is not enough to have a set of data policies. Without a process for

enforcement, the policy has no "teeth" and will be useless. It is incum-
bent on the drafters of a data ownership policy to incorporate the meth-
ods for enforcement as well as the process for conflict resolution.
Enforcement takes the form of a means for validating that the policies
are being followed and actions to be taken if it is determined that the
policies are not being followed. It is the role of the policy manager to
ensure that the policies are being enforced.
Validating policies is a making sure that all parties are living up to
their responsibility agreements and are not overstepping the bounds of
their responsibility. This involves periodic reviews by the policy man-
ager, under the authority of the CIO, of whether each assigned role is
being performed adequately. If the parties are not doing their job, it
may be necessary to give them further training or decrease their respon-
sibility (and authority) and possibly remove them completely from the
position of responsibility.
No organization is immune to conflicts, and there should be a dis-
pute resolution process in place. Typically, this will involve bringing the
dispute to the next level in the management hierarchy with responsibil-
ity over the particular area of conflict (we will call this person the dis-
pute manager). If the parties dispute an ownership role or whether the
responsibilities of an ownership role have not been fulfilled properly,
the dispute manager should consult the ownership policy along with the
ownership registry and use that information to establish a ruling for the
dispute. If the issue is not covered in the data ownership poUcy, then the
policy needs to be modified to incorporate the issue.
2.6.3 Data Quality Rules
As we move toward an organization with negotiated ownership roles,

part of the responsibility of the management is to ensure that any data
that is shared across the enterprise lives up to a standard of use. Ensur-
ing this implies a few notions.
There is a notion of data quality that is well defined throughout
the enterprise.
There is a means to describe data quality requirements.
There is a means to measure conformance to the data quality
requirements.
There is an enterprisewide agreement as to the expected levels of
data quality.
There is a mechanism for improving data quality.
The concept of a definition of data quality is complex and impor-
tant and will be covered in Chapter 3, and details of data quality ideas
are discussed in Chapter 5. Ways to measure ongoing performance are
examined Chapter 4. A large part of this book is devoted to defining
and vaUdating data quality rules.
2.6.4 Education and Training
When using data ownership policies to regulate the movement and

management of information across an enterprise, it is critical that all
participants understand the policies and how they work. This means
that the managers defining the data ownership policy must also arrange
for the education and training of staff members who are expected to use
it. Data ownership training should cover at least these items.
1. An overview of data ownership, including ownership paradigms,
ownership value, and data ownership responsibilities
2. A survey of the enterprise knowledge system covering the system
architecture, the data architectures, the data set catalog, and all
management, delivery, and presentation applications
3. An overview of the data ownership hierarchy describing the dif-

ferent roles involved, the responsibiHties of each role, the report-
ing structure, and the conflict resolution process
4. A session covering enterprise metadata
5. A training session on the value of enterprise data quality, includ-
ing an overview of data quality, a discussion of continuous mea-
surement for improvement, and the definition and use of data
quality and business rules
2.7 SUMMARY
This chapter introduced the concept of data ownership as a manage-

ment issue. Because information is bought, used, created, modified,
propagated, and sold throughout an organization, the enterprise data
processing function can be contrasted to a factory where individuals or
automated processes play specific roles.
Complicating notions in the enterprise make data quality manage-
ment a difficult task, especially due to issues such as differing views of
the value of data, privacy issues, turf wars, or standard bureaucracy. A
list of the responsibilities of data ownership was provided for clarity.
There are different kinds of data ownership paradigms, and, de-
pending on the organizational point of view, different ownership rules
may apply in different situations. Ultimately, though, the management
of the organization must choose between tight, centralized control or
loose, decentralized control of data. These decisions are incorporated
into a data ownership policy where the stakeholders, data sets, respon-
sibilities, and dispute resolutions are all clearly defined. As stakeholders
agree to subscribe to the data ownership policy, a more ordered envi-
ronment enables better knowledge management overall and data qual-
ity management in particular.
3
DATA QUALITY IN PRACTICE
In this chapter, we explore what data quality means and how it can be
effected in the enterprise. Without a definition for data quality, how can
we ever hope to improve it? And since everybody has a different idea as
to what data quality means, how can we level-set the members of an
organization so that an improvement in data quality can be measured
and recognized?
With the simple definition of data quality as "fitness for use," we
can start to plan how to improve data quality across the organization.
We first outline our procedure for a data quality improvement program,
which incorporates the gaining of senior management consensus, train-
ing, analysis, and implementation in a way that allows for a continuous
improvement program to build on each individual success.
We will also spend some time discussing what data quality means
within certain implementation domains operations, databases, data
warehousing, data mining, electronic data interchange (EDI), and the
Internet. In each of these domains, we will explore the importance of
data quality and how a data quality program can be integrated into the
domain as an integral component.
3.1 DATA QUALITY DEFINED: FITNESS FOR USE
What does data quality mean? In practicality, almost everyone has a dif-
ferent view of data quality. To the mailing list manager, data quality
means cleansed delivery addresses and deduplification. To the account
manager, data quality means accurate aggregation of customer activity.
47
To the medical industry, data quality may mean refined ability for record
linkage. Clearly, each definition is geared toward the individual's view of
what is "good" and what is not. This leads to the conclusion that there is
no hard and fast definition of data quality. Rather, data quality is defined
in terms of how each data consumer desires to use the data.
In the most general sense, we will use a qualitative definition of
data quality and refine that definition on a case-by-case basis. In essence,
we define data quality in terms of fitness for use the level of data
quality determined by data consumers in terms of meeting or beating
expectations. In practice, this means identifying a set of data quality
objectives associated with any data set and then measuring that data
set's conformance to those objectives.
This is not to say that the tools used for static data cleansing of
names and addresses or products that link data records based on spe-
cific data fields are not useful. It is true, however, that the use of these
tools is not a solution to the data quality problem. Instead, the best way
to get a handle on an organization's data quality is to define a set of
expectations about the data, measure against those expectations, and
continuously improve until those expectations are satisfied.
What has proven to be difficult up until now is that because every-
one's data sets are different, there are no well-defined means for defin-
ing data quality expectations. In this book, we address this need by
developing all the tools needed to determine if there is a data quality
problem, to measure the cost effect of low data quality, to assess the
current state of the organization's data, and to develop data quality
rules that can be used for measurement. But the first step in any of these
processes is to understand the notions of data quality and getting
senior-level management support for the assessment and improvement
of enterprise data quality.
3.1.1 Freedom from Defects
If we want to refine the definition of fitness for use, the first area of
focus is limiting the "badness" of the data. We can refer to this aspect as
"freedom from defects," where a defect is any situation where data val-
ues are not accessible or do not correspond in accuracy to an estab-
lished frame of reference. While we will explore the areas where defects
can crop up in Chapter 5, here is a short list of the kinds of defects we
want to avoid.
DATA QUALITY IN PRACTICE 49
Inaccessibility
Inaccuracy
Out-of-date information
Unmanageably redundant information
Inconsistency with other sources
Incomplete data
Incomprehensible data
3.1.2 Possesses Desirable Characteristics
The flip side of freedom from defects is that the information has the
characteristics of a high-quality environment. Again, we will explore
these characteristics in Chapter 5, but here is the short list.
The information is timely.
The data model completely and accurately models its real-world
counterpart.
The information is presented in a way that is easy to understand.
The appropriate level of detail is maintained.
The information captured is meaningful in its proper context.
3.2 THE DATA QUALITY IMPROVEMENT PROGRAM
In any improvement process, there are a number of steps that must be

taken to get to the point of a demonstrable, measurable improvement.
In this section, we review some of the steps, although we treat this in
much greater detail in Chapter 18.
3.2.1 Senior-Level Endorsement
A data quality improvement project will not be successful if there is no

senior-level management endorsement. Because of the kinds of compli-
cating ideas about data ownership that we discussed in Chapter 2, the
successful data quality engineer must first have in place a data owner-
ship policy and a process for conflict resolution within the organization.
Additionally, the conflict resolution process must take the value of the
data quality improvement into account in a way that does not decrease
the effectiveness of the data quality team.
Gaining senior-level management endorsement may prove to be a

more difficult task than one might imagine. Unfortunately, we have
often seen that data quality is like the weather everybody talks about
it, but no one ever does anything about it. There are some interesting
reasons for this.
No Mea Culpa Syndrome No one wants to admit that there
might be mistakes in their work. Low data quality implies that there are
errors somewhere in the system. This is especially true in environments
where employees are rewarded for a job well done, particularly when it
reflects well on the management chain. Rarely do employees admit their
own mistakes or call attention to those of their subordinates.
"The Emperor's New Clothes'' People often assume that the dik-
tats of a person in a senior authority position are naturally correct and
appropriate and it would be folly to challenge the status quo. In this
case, a comment about actually trying to find and fix poor data quality
problems might be more of a job security risk.
Denial of Responsibility The natural response when a mess is
discovered is to point the finger at someone else. Few are willing to
accept that the existence of poor data quality might be due to their own
misjudgment, and so the DQ "ball" gets bounced from desk to desk,
but it never lands anywhere.
Lack of Understanding Too often data quality is confused with
"data cleansing," "deduplification," or name and address standardiza-
tion. While all these items are an integral part of a data quality program, a
periodic run of a data cleansing application is not a replacement for an
ongoing knowledge management program integrated with continuous
data quality improvement.
Dirty Work There is a feeling that looking for problems and
cleaning up data "messes" are dirty work and no-glory activities. There-
fore, involvement in data quality activities is viewed as a "career-limiting
move."
Lack of Recognition Building on our last point, there is always
some expectation (by data consumers, customers, etc.) that things
should be right, and this implies that making things right is expected.
The employees who toil at making sure that everything looks correct
are viewed as just doing their job and are not recognized for performing
an extremely important knowledge management task.
The conclusion is that there must be an enlightenment among the
senior-level managers that breaks the negative connotations associated
with data quality and shows how an integrated set of data quality solu-
tions can add value to the organization. This enlightenment can be
effected through a number of steps, including initial training in knowl-
edge management and data quality, followed by the creation and
endorsement of a data ownership policy, along with the analysis that
demonstrates the economic impact of low data quality and the eco-
nomic value of measurable high data quality.
3.2.2 Data Quality Training
Training on data quality begins with understanding the principles of

what quality means in the context of information. A training program
in data quality should incorporate overviews and discussions of at least
these concepts.
1. Creation and usage of information
2. Storage of information
3. Data ownership
4. Quality concepts and the quality improvement cycle
5. Understanding the economic impact of data quality issues
6. Dimensions of data quality
7. Aspects of reference data domains
8. Data quahty and business rules
9. Metrics for measuring and assessing data quality
10. Metadata
11. Data quality requirements analysis
12. Data cleansing and standardization
13. Error detection, correction, and root cause analysis using data
quality rules
14. Data enhancement
3.2.3 Data Ownership Policy
As we discussed in Chapter 2, a data ownership policy governs the

responsibility over the collections of information used in an enterprise.
Without a policy, there is no process for accountability, so before any
further steps can be taken, the endorsing senior managers must draft,
accept, and publicize a data ownership policy, as well as define and
incorporate the procedures for enforcing the policy.
3.2.4 Build an Economic Model of Low Data Quality
Once the policy and its enforcement procedures are in place, the next
step is to identify those areas in greatest need of improvement. (Chapter
4 presents a framework for finding those areas.) In our economic model
of low data quality, we provide a mechanism characterizing the actual
impact of data quality both within and external to the organization.
This is done by taking these steps.
1. Looking for the signs of data quality problems
2. Mapping the flow of information into, through, and out of the
enterprise
3. Characterizing the impacts of low data quality at particular
stages in the information chain
4. Measuring the cost impact of low data quality
5. Building the data quality scorecard
The data quality scorecard is a tool used to focus on the locations
in the information chain where there are data quality problems that
have the greatest impact on the organization. The scorecard can be used
as input to the next step, current state assessment.
3.2.5 Current State Assessment and Requirements Analysis
Finding areas of high impact is one thing, butfiguringout how to address

them is another. Before we can fix a problem, we have to know how bad
the problem is. That is where the current state assessment comes in.
At each location of impact, we select some of the dimensions of
data quality covered in Chapter 5 and measure the level of data quality.
These levels are correlated to the degrees of economic impact deter-
mined by the scorecard analysis. We then use the current state analysis
to discover which areas of data quality account for the largest economic
impact and for a baseline measurement of the overall data quality of the
enterprise.
The baseline numbers coupled with the degree of impact allow us
to both determine what needs to be fixed and the levels needed to
demonstrate improvement. This is all determined during the require-
ments analysis stage. Requirements analysis will assign target levels of
data quality that need to be met in order to relieve the problems uncov-
ered during the economic impact analysis and current state assessment.
3.2.6 Project Selection
With the data quaUty scorecard, the current state assessment, and the
requirements analysis, there is enough data to select a project for
improvement. With senior management support, a team is assembled
and assigned a specific goal: to bring up the level of measured data qual-
ity to the target level determined during the requirements analysis.
Selecting a single project for execution is important. Unfortunately,
many data quality improvement projects are subject to failure because the
scope encompassed is way too large. It may be impossible to demonstrate
overall improvement if there is no particular focus. Remember that the
overall success of the program is determined by small successes in small
steps. Selecting a high-profile but small project for improvement, and suc-
cessfully completing that project, accomplishes three things.
1. It provides a measurable (both in hard metrics and in economic
benefit) improvement in the quality of information in the
enterprise.
2. It gains positive association within the organization for accom-
plishment, which in turn builds more senior-level buy-in and
general consensus.
3. It opens the possibility for additional improvement projects.
3.2.7 Implement and Deploy the Improvement
The implementation of the improvement is the critical denouement of

the data quality drama. The improvement team must be empowered to
take these steps.
1. Architect a solution This means determining what actual steps
must be done to address the data quality problem.
2. Implement the solution This incorporates the acquisition of a
development platform, access to the environment in which the
solution is needed, and access to the resources needed to perform
the implementation.
3. Deploy the solution This means being able to integrate the
solution with the production systems and move the new solution
into production.
3.3 DATA QUALITY AND OPERATIONS
Business operations are defined by a set of rules that are applied in

everyday execution. When the business depends on the correct flow of
information, there is an aspect of data quaUty that intersects the opera-
tional specification.
In essence, in an information business, business rules are data qual-
ity rules. This implies that data quality is an integral part of any opera-
tional specification, and organizations that recognize this from the start
can streamline operations by applying data quality techniques in
process. This in turn prevents bad data from affecting the flow of busi-
ness and stops incorrect information from entering the system, eliminat-
ing the need to detect and correct bad data. Thus, a "data quality
aware" operation can execute at lower cost and higher margin than the
traditional company.
Data quality and business rules can be used to incorporate a rule-
based execution engine that can validate information as it passes
through an organization (see Figure 3.1). Rules are defined that specify
Data quality and business rules Rules can direct the packaging
indicate how information and and presentation of information
control flows in the process. to the consumer.
Information must Low levels of data

be validated before quality can trigger
it enters the system. supervisor notification.
FIGURE 3 . 1 The need for data validation in the operational environment

what information must look like before it can enter a specific process,
as well as validate the information as it passes from producer to con-
sumer. In addition, data validation and business rules can be used as
triggers for particular events within an operational environment. For
example, threshold limits can be set for operational efficiency based on
the amount of invalid information in the system. When these thresholds
are encountered, an event is triggered that notifies a supervisor of the
presence of a problem and the location of the problem.
3.4 DATA QUALITY AND DATABASES
Of course, the concept of data quality pervades the world of databases.

Our lives are touched by databases every day, ranging from direct mar-
keting mailings that bombard us, utility and credit card bills that dis-
tress us, exasperating health insurance provider services and down to
taxes, our driving history, and those videos we rent.
It is unusual, though, for anyone to be untouched by bad data sit-
ting in one of these databases. Though these can be due to machine
errors, more often they are the result of human error. Data quality is
integral to the proper use of databases.
3.4.1 Databases and Data Quality
There has been longstanding recognition of the importance of data

quality within the world of databases. Databases are in fact designed
with certain safeguards for data quality, although they are often over-
ridden in practice. Processes such as embedded null testing, functional
dependence, data normalization, and referential integrity are all testa-
ments to the need for DBMS providers to include some degree of data
vaUdation within the database system.
Databases are often meant to represent a model of the real world,
and their quality can be measured by their conformance to the real-
world entities that they are designed to model. As a database ages and
no longer has the input of "positive energy," the real world tends to
diverge from what is represented in the database.
There are a number of data quality areas associated with database
systems, such as static cleansing applications that will examine data and
attempt to put it into some canonical format. Standardization can be
applied to name and address data, and deduplification is a process to

remove duplicate entries from a database. See Figure 3.2.
Rule-based data quality can be used to direct the entry of informa-
tion into a database (see Figure 3.3). If there are dependencies between
values in certain attributes, that information can be used to direct the
process of collecting the information. An example of this might be in
automated screen generation based on data dependence rules. The data
entry process would then be guided by the staged accumulation of data
values targeted to specific attributes. As the values are seen, they may
have some impact on the types and sets of attributes that must be filled
next. For example, consider an online medical history form, which
should only ask about a patient's most recent pregnancy only if the
David Loshin David Loshin

Data Quality Managei; Knowledge Integrity Incorp. 1 / Address \ J Data Quality Manager
123 Maine Street I Standardization j \ Knowledge Integrity Inc.
New York, NY 123 Maine St.
New York, NY 10001-0001
David Loshin 123 Maine Street New York NY

David Loshing 123 Main Street New York NY David Loshin 123 Main Street New York NY
H. David Loshin 123 Main St. New York NY
David Loshin 123 Maine Street New York NY

H. David Loshin 123 Main St. New York NY
Loshin 123 Main Street New York NY

Jill Loshin 123 Main Street New York NY Dauid
Jill Fingerhut 123 Main St. New York NY I j[ Householding j-J JiU
Kira Loshin 123 Main Street New York NY Kira
David & Jill Loshin 123 Main Street New York NY
Name: Loshin Data

Street: 123 Main St. Validation
City: New York
RULE: Order may
State: NY INVALID PURCHASE
Membership: Standard not be greater thart
Account: 1263587K purchase limit.
Purchase Limit: $5,000.00
Order: $7,500.00
FIGURE 3.2 Data quality applications

DATA QUAUTY IN PRACTICE 57
patient is determined to be female. We explore this in greater detail in

Chapter 17.
3.4.2 Legacy Databases
Legacy databases pose interesting data quality questions, since it is

likely that the people originally involved in the creation and develop-
ment of the database have long since left the organization. Legacy data-
bases hold more than just data they hold secrets about embedded
business rules folded into the way the information is stored, viewed,
manipulated, and so on.
The dependence
relationship between data
attributes is broken out into
Set of Data
Quality Rules
FIGURE 3 . 3 Data directed input screen generation

An interesting data quality issue surrounding legacy data is in dis-

covering the data quality rules and business rules embedded both in
associated application code and in the data itself. As systems evolve,
there is a growing reluctance to handle or modify any piece that is
working in production, so as time passes, many changes are applied
"around the edges" in ways that create new application execution
dependencies that are artificial.
For example, in some mainframe execution systems, the job sched-
uler must know exactly what the job execution dependencies are in
order to arrange an execution schedule. Lore regarding the completion
time associated with certain jobs was passed down as an imperative to
defining those dependencies. This stricture was imposed because of an
assumption that a certain processing phase must be finished before
another could begin. This did affect the completion time of the cycle of
processing, which by industry regulation had to be completed within a
prespecified time.
In reality, the actual dependence between jobs was much less rigid.
The actual set of rules showed a less restrictive dependence between the
initiation of the jobs (one had to wait until the other started, not fin-
ished). That discovery, along with an increase in the available technol-
ogy associated with database locking, had obviated the original
dependence. By removing the dependence and logging the interaction
rules between the two jobs, a much more efficient schedule could be
produced.
The benefit of a more efficient job schedule was a decrease in the
critical path of execution, which also decreased the completion of the
production cycle. The processing time freed by optimizing the job
schedule enabled an increase in processing volume as well as a decrease
in the per-item cost of processing. The result: an increase in cost margin
for the company!
Another interesting feature about legacy data is that while technol-
ogy has advanced, legacy systems are bound by their implementation
platform. Systems developed before the introduction of relational data-
base systems are limited to the power provided by the original system.
This can have two effects: (1) Many things that can be implemented eas-
ily with today's technology, are impossible in older technology, which
can be an argument for a database renovation project, and (2) clever
tricks that had been used in older systems to get around implementation
deficiencies become albatrosses, which prevents improving the system
DATA QUAUTY IN PRACTICE 59
in its current implementation. The innovators of these tricks move on to

other positions, and these secrets are lost, but the constraints remain.
3.4.3 Data Migration
An issue complementary to legacy databases is the migration of legacy

data to new databases. There are similar data quality issues involving
implementation decisions and how^ they affect overall information
quality.
We could say that in the migration environment, we really want to
perform information archaeology. We want to understand the historical
use, manipulation, and storage of information in the transactional sys-
tem and then characterize that knowledge in a way that eases the move-
ment of data from the legacy system to the new system. Since a lot of
knowledge is likely to be embedded in the data and in supporting appli-
cation code, we will need to document what we learn about the data
(that is, metadata) and maintain the learned metadata in a metadata
repository.
What are the reasons for data migration? Here are a few.
1. The underlying database management system (DBMS) is ob-
solete.
2. There is no longer enough support personnel for the transaction
system.
3. The company making the application code has gone out of busi-
ness.
4. There are better DBMS systems available.
5. There are system requirements that can be better met using a new
system.
6. The current DBMS is not supported on high-performance hard-
ware.
7. There are better ways to manage data in a new DBMS.
In any case, the reasons for migration yield insight into what kind of
metadata must be collected. This metadata can be used for the data
migration process followed by any transformations that must be per-
formed before loading it into the destination database. Both data quality
rules and business rules will govern the extraction and transformation
process. This is shown in Figure 3.4.
Metadata
Legacy
Database
Business
Rules
FIGURE 3.4 Learning from the migration process
3.4.4 Data Extraction, Transformation, and Loading
The data extraction, transformation, and loading process (referred to as

ETL) is almost completely governed by data quality rules. The
processes of extracting data from a set of data sources, combining that
data, perhaps applying some changes, and then loading that data into a
target database all comes under the category of the data quality rules
we will discuss in Chapters 7 and 8.
Many tools are designed specifically for the ETL process. These
tools often use a visual interface to define a mapping between the origi-
nal source data sets and the target data set. A library of transformation
functions, combination or aggregation functions, or insert user-defined
operations can be used to apply to data. The tools order the set of oper-
ations (data extractions, transformations, etc.) and then automatically
generate the application code needed to extract the data from the
source databases, perform the transformations, and insert the resuh

into the target. The resuh is a packaged apphcation from end to end to
perform the ETL.
3.5 DATA QUALITY AND THE DATA WAREHOUSE
Data marts and data warehouses are used for an analytical environment.
It is often said that the bulk of the work performed in implementing a
data warehouse is in the data extraction, cleansing, and transformation
phases of moving information from the original source into the ware-
house. Nevertheless, many data warehouse projects fail because not
enough attention is spent on either understanding the data quality
requirements or on the validation and quality assurance of information
imported into the warehouse.
3.5.1 Flawed Data = Flawed Results
Analyses performed using data warehouses with flawed information

will produce flawed strategic decisions. Many organizations report that
as much as 70 percent of their effort goes into supporting the data
cleansing and transformation process. This effort is split between data
auditing and data cleansing.
Ensuring that high-quaUty data enters a data warehouse decreases
the potential for flawed strategic decision making. What distinguishes
the approach presented here in data warehouse design and implementa-
tion is that our philosophy revolves around the idea of building data
quality into the product at the time of design. Instead of relying on data
cleansing tools that statically clean a database, this methodology pre-
sumes the validation of data items upon insertion into the database.
It is necessary to work with data consumers to understand, iden-
tify, and abstract data quality requirements, to determine the data qual-
ity rules, and to integrate tools to test and validate data items at the
insertion point. The methods described here allow the user to integrate
in-process qualification into the information manufacture and process-
ing chain to both measure and validate data quality. At the same time
this will provide feedback for the identification of root causes of data
quality problems.
3.5.2 Example: Data Warehouse Certification
As an example, you can use data quality and business rules for what we
call "data warehouse certification." Certification is a means of scoring
the believability of the information stored in a data warehouse. A data
warehouse is considered fit for use when the data inside conforms to a
set of data quality expectations embodied in a set of rules. Given these
rules, we assign a score to the quality of the data imported into a data
warehouse for certifying warehouse data quality (see Figure 3.5).
The first step is to define a set of rules that will qualify the data.
Again, we can use the rules framework that is described in Chapters 7
and 8. The next step is to import those data quality rules into a rules
engine. Each rule will have an associated validity threshold (as a per-
centage) based on the users' expectations of quality.
Data Quality
Rules
Rule
Thresholds
Data Warehouse Input
FIGURE 3.5 Data Warehouse certification

As records are fed into the engine, any relevant rules (that is, any
rules that refer to values of attributes defined within the record) are
tested. If no rules fail, the record is said to be valid and is successfully
gated through to the warehouse. If any rules fail, the record is enhanced
with information about which rules were violated, and the record is
output to a reconciliation system. The violating record can also be
passed through to the warehouse, but now it is marked as having not
conformed to the users' expectations, and this information can be used
when performing analysis. The count of failures and successes is main-
tained for each rule.
After the data is imported, each rule's validity value is computed as
the ratio of valid records to the total of records, A data quality certifica-
tion report delineating all validity percentages is generated. If all valid-
ity percentages exceed the associated thresholds, the warehouse is
certified to conform to the users' data quality requirements. Otherwise,
the warehouse is not certified, and until the percentages can be brought
up to the conformance level, the warehouse cannot be said to meet the
data quality requirements.
To qualify the warehouse after a failed certification, the records
output to the reconciliation system must be analyzed for the root cause
of the failures. This analysis and correction is part of a business work-
flow that relies on the same set of data quality and business rules used
for validation. After reconciliation, the data is resubmitted through the
rules engine, and the validity report is generated again. The root cause
information is used to return to the source of problems in the legacy
data and to correct those bad data at the source, gradually leading to
certification.
Data warehouse and data mart certification is an ongoing process,
and the certification report needs to be aligned with the ways that data
records are inserted into the warehouse. For example, if a data mart is
completely repopulated on a regular basis, the certification process can
be inserted as a component to the reloading process. Alternatively, if a
data warehouse is incrementally populated, the results of the certifica-
tion engine must be persistent.
3.6 DATA MINING
Data mining, or knowledge discovery, is the process of looking for pat-

terns or anomalies in large data sets. Data mining is really a compendium
of algorithms meant to discover knowledge embedded in data. This

includes decision trees, neural networks, genetic algorithms, and associa-
tion rule discovery, along with other statistical and correlation techniques.
It is well known that the bulk of the work associated with data
mining is in preparing the data for analysis. This data preparation can
be broken down into these phases.
Data Characterization Data miners will typically know how the
data is put together the base data types, the size of each attribute,
and other elements, but data characterization takes this to a higher
level. Characterization involves describing the nature of the content of
the data as opposed to just the format or the structure of the data.
Consistency Analysis Although we expect to use the same name
all the time, inconsistencies can creep into data that indicate different
names for the same entity for example, "IBM," "International Busi-
ness Machines," or "Int'l Bus. Mach." There can also be inconsistencies
on a broader scale, such as different terms used as attribute names in
several tables that model the same domain.
Domain Analysis In order to make sure the data values are valid,
the data types are subtyped into more restrictive domain descriptions.
Data Enrichment This involves incorporating more than one
data source into the data set used for analysis.
Frequency and Distribution Analysis Sometimes the value of cer-
tain data values is related to the number of times the values occur. Fre-
quency analysis can capture some of this information.
Normalization Normalization involves transforming data values
within one domain into values within another domain. For many pur-
poses it is desirable to relate data values drawn from a continuous
domain into a narrowed range. For example, for many neural network
applications, all continuous values should lie between 0 and 1.0.
Missing and Empty Value Analysis Missing data values poten-
tially skew analytical models, so there must be some mechanism for
detecting, characterizing, and dealing with missing values.
Data mining and knowledge discovery are particularly valuable
when in tandem with the data quality and business process because
data mining can be both a consumer and a producer of data quality and
business rules. Data quality techniques must be used to prepare the data
for the knowledge discovery process. In turn, the data mining will
expose data quality and business rules that can be used for further
knowledge discovery.
3.7 DATA QUALITY AND ELECTRONIC DATA INTERCHANGE
Electronic Data Interchange (or EDI) is the term used for any standard-
ized format for representing business information for the purposes of
electronic communication. EDI is enabled through a process of cooper-
ative data standardization within a particular business environment. It
is used in many industries today, such as the health care and financial
industries, and is firmly entrenched in interactions with the federal gov-
ernment.
EDI is used to eliminate manual processing when executing routine
business operations such as purchase orders, product orders, invoicing,
securities trading transactions, shipping notices, and so forth. This
increases efficiency and volume, thereby lowering the overall cost per
transaction.
EDI is more than just forwarding purchase orders and invoices via
e-mail or through Intranet postings. It is designed to use standards for
formatting and transmitting information that are independent of the
hardware platform. EDI enables a process known as straight-through
processing, (STP), which is the ability to completely automate business
operations with no manual intervention.
Any STP or EDI activity is almost by definition one that is based on
data quality. Because EDI is defined as a standard form for transmitting
information, there are rules about validation of the both the format and
the content of EDI messages. Precisely because STP is meant to replace
well-defined operational processes, there are specific business rules that
guide the STP applications in how to process the transaction.
What is more relevant is that not only must an STP application be
able to execute operations based on a set of predefined business rules, it
must also be able to distinguish between valid and invalid EDI messages.
Validity is more than just structural conformance to the EDI format it
must also include validation of the content as well. STP systems often
will only perform the structural validity checks, and not the content
validity, which may account for slips in actual complete automation.
3.7.1 XML
A good example of this can be seen in XML (Extensible Markup Lan-

guage) for EDI. XML is both a structure definition framework (for the
definition of data standards) and a well-defined set of rules for validating
based on defined standards. XML syntax imposes a structure on the

form of a document through the use of tags for markup metadata
directives describing the data in the document. These are the six kinds of
markup that can occur in an XML document.
Elements, which are delimited by angle brackets and denote the
kind of information they surround. There are two kinds of tags
that are used for elements: start tags, (denoted with <item>) and
end tags (denoted with <item>).
Entity references, which are alternate ways of expressing special
characters or symbols as well as frequently used text within a doc-
ument.
CommentSy which are delineated text items that are not part of
the document.
Processing instructions, which include pieces of information that
are to be passed to a processing application but are not part of the
document.
Marked sections, which section off areas from being parsed by an
XML parser (that is, all characters are taken verbatim).
Document type declarations, which is a set of declarations and
specifications for creating well formed documents.
A DTD can be used to define a data format standard for EDI, and
XML is a nice framework because there are inherent notions of "well
formed" and "valid."
3.7.2 XML: Well Formed
In the context of XML, a document is well formed if it conforms to the

syntax of XML. This means that an XML parser can successfully parse
the document. Also, in order to be well formed, a document must meet
a number of other conditions. For example, no attribute (a name-value
pair that attributes a tag) may appear more than once on the same start
tag, all entities must be declared, and nonempty tags must be nested
properly. Another condition is that neither text nor parameters may be
directly or indirectly recursive. Other conditions are beyond the scope
of this discussion. Suffice it to say, an XML parser can determine
whether a document is well formed or not.
3.7.3 XML: Valid
An XML document is valid if it is well formed and contains a DTD and

the document obeys the rules specified in the DTD. This means that ele-
ment sequence and nesting conform to the DTD, all required attribute
values are present, and their types are correct.
3.7.4 Structure vs Content
While there are notions of both well-formedness and (structural) validity,

note that there is no notion of content validation. Thus, even when the
format of an EDI message conforms completely to the syntax rules, the
data inside the message can still cause problems down the line.
As an example, consider a financial securities trading application.
The EDI format may specify that any trade must include the name of
the security, the security type, and the quantity traded. But if a message
came through with an invalid security in the security name field, the
message may still be gated through to the trade processing application,
where it will most likely create a failure in processing. The structure of
the message was valid, but the content was not. We will explore ways to
specify content validation rules that add additional insurance against
incorrect messages causing failure in straight-through processing.
3.8 DATA QUALITY AND THE WORLD WIDE WEB
As an information forum, it is hard to fully understand the impact of

the World Wide Web. One feature that permeates the Internet is the
complete lack of organization of the available information. Another is
the temporal quality of the information presented.
The Internet can be seen as the largest uncoordinated database in
the world. Here technology enables the use of information available
over the World Wide Web as a large data repository, with data collec-
tion and aggregation scripts acting as a "view-generator" for any par-
ticular user application. All Web sites can be characterized in terms of
some kind of data service. Some sites act as data presentation systems,
some as data collection systems, and some as database query systems.
The plethora of information that can be accessed and viewed by a
web user provides great potential for simulating a data warehouse or
business intelligence repository viewed through a "virtual portal." This

portal can aggregate collected information sources (both internal and
external) into a single business intelligence console coupled with "intel-
ligent agents" that seek out the desired data and deliver it directly to the
desktop.
3.8.1 What Is Missing from the Data Warehouse Analogy
The data warehouse analogy partially breaks down when it comes to

the actual aggregation and presentation of information at the presenta-
tion site. A critical value that can be derived from the aggregation of
information is its coagulation, enrichment, and presentation in terms of
critical business intelligence that is meaningful to the user.
The obvious reason for this is that similar types of information
available across the Web are represented in different ways. Whether
using different storage formats, different names, or different visual rep-
resentations, the data presented at different Web sites and collected and
delivered to the client location can be made much more valuable when
that data is massaged into business intelligence guided through client-
defined requirements.
Another reason for the breakdown in the analogy has to do with
governance (or more clearly, the lack thereof). Anyone can publish any-
thing on the Internet, yet there is no governing body that oversees the
quality of any information published. Therefore, it is up to the user to
characterize his or her data quality requirements to which the informa-
tion should conform before the information is presented.
3.8.2 The Web as a Transaction Factory
Another aspect of the use of the Internet is the growth of both business-
to-consumer (B2C) and business-to-business (B2B) EDI frameworks for
executing transactions. Whether we talk about amateur day traders,
competitors in online auctions, or B2B order processing and fulfillment,
there is an increase in the use of the Internet as a distributed transaction
factory.
3.8.3 Data Presentation
Both commercial sites for presenting information about a company's

services or products and personal Web sites for presenting personal
information are data presentation systems. A single Web page peppered
with text statements is the simplest form of data presentation.
As Web sites expand, they are frequently fragmented into different
presentation paradigms, including frames, tables, and pages. At some
point, there is too much information for a single Web page, so the infor-
mation is broken up into multiple pages. The presenter might also want
to use frames to allow more controlled navigation throughout the Web
site. Within a single page (or frame), tables are used both for data aggre-
gation and data presentation.
Another presentation device is the use of hyperlinks. Hyperlinks
are embedded references to other Web site addresses that can be consid-
ered "canned" information queries. Each hyperlink represents a request
to provide more information about a particular subject.
3.8.4 Internet Data Queries
There are at least five different data query paradigms over the Internet.
1. Hyperlinking As we said, hyperlinks can be seen as canned
queries that lead to (relatively) static data sets. As Web pages
evolve, sets of hyperlinks are grouped under distinct headings to
further refine and classify the virtual query.
2. Local data searching engines A local search engine that has
indexed the information present in a Web site's set of Web pages
can produce a set of references to Web pages that matches a
user's specific query.
3. Service queries This allows the web page to act as a front end
to an actual database service that is connected to the Web site,
although it may act independently of the Web site. Many
e-commerce sites operate in this manner. An example is a query
through an online catalog site as to the number of requested
items currently in stock and ready to be shipped.
4. Web-wide search engines These are web search engines that
front to databases of indexed Web pages. This can be considered
an extension to the localized Web searching.
5. Meta-Web-wide search engines These are Web sites that front

to other Web search engines.
3.8.5 Data Collection
Another primary database activity includes data collection. Anytime

the browsing user is required to submit some set of information, either
to gain access to a site, to register as a site user, to gate to more secure
sites, or just for informational purposes, a Web site acts as an update
mechanism to a database. The database may consist of customer infor-
mation, statistical information, or even simple logging information.
3.8.6 Example: Poor Data Collection
Once when I was using the Internet to collect information for a report I
was preparing on some data management tools, I visited a number of
Web sites for companies that provided those tools. At one particular
corporate site, I found a menu of white papers that were clearly relevant
to my task, but in order to download the papers, I was required to fill
out a form with pertinent contact information name, mailing address,
phone number, e-mail address, and so forth.
After submitting the form, I was gated through to the Web site to
download the white papers, at which point I made my selection and
downloaded the paper. After reading the paper, I realized that there had
been another paper at that site that I also wanted to read. I repointed
my browser at the site, only to find that I was required to fill out the
form again. Only after resubmitting all the same information was I able
to download the other white paper.
About two days later, I was called by a company representative
who introduced himself and asked if I had successfully downloaded the
white paper (not papers!) and inquired about my interest in their prod-
uct. I replied that I had gotten the papers, but that I had to fill out the
form twice. I asked him if my filling out the form twice meant that I
would appear in their database twice. He said it would. I then asked if
that was a problem, considering that if I were called more than once
about the product, I might get annoyed. The salesman answered that
while I would appear in the database more than once, he was aware of
the problem and usually was able to mentally keep track of whom he
had already called. We briefly talked about data quality and duplicate
record elimination, but he seemed convinced that this was not a serious
dilemma.
About four hours later, I found a message on my voice mail. Who
was it? It was the same salesman, calling to introduce himself and to see
if I had successfully downloaded the white paper! What was ironic is
that the salesman not only completely exposed his inability to keep
track in general, he completely forgot our earlier conversation focusing
on this exact problem!
The simple practice of collecting sales lead information through
Web-based forms is a compelling idea as long as the operational aspects
don't backfire. We can see at least three ways how the absence of a data
quality program associated with this Web site can diminish the effec-
tiveness of the process.
Across one dimension, by filling out that form I expressed interest
in the area associated with the products being sold. As a potential cus-
tomer, filling out the form is a proactive action, exposing me as a quali-
fied sales lead. Yet, by registering a visitor in the database more than
once, the system dilutes the effectiveness of this sales lead qualification.
Across a second dimension, a measurable cost of customer acquisi-
tion is increased each time the salesman makes a telephone call. Associ-
ated with each call is some expectation of converting the lead into a
sale. By calling the same person twice, however, the salesman is not only
performing rework but preventing himself from getting real work done,
such as following up on other sales leads.
Across a third dimension, there is a strategic weakening of the tar-
geted sales process. When targeting a potential customer, it is in the
salesman's best interest to foster a good working relationship with that
customer. Interrupting a decision maker once during a day to gauge
interest is probably bearable, but someone might be irritated by having
his or her work routine interrupted twice in the same day by the same
person. I would probably make the assumption that the company is not
very organized, which would make me question their ability to provide
a good product.
3.9 SUMMARY
In this chapter, we examined the definition of data quality in terms of

fitness for use and how understanding the measure of data quality may
be very dependent on the context and domain in which the data is scru-
tinized. We looked at the development of a data quality improvement
program and what it takes to get the ball rolling. We determined that
there are seven phases to the program.
1. Gaining senior level endorsement
2. Training in data quality
3. Creating and enforcing a data ownership policy (as discussed in
Chapter 2)
4. Building the economic model (will be discuss in Chapter 4)
5. Performing a current state assessment (as will be described in
Chapter 9)
6. Selecting a project for improvement
7. Implementing and deploying the project
Of course, with each successful project implementation and
deployment, there are opportunities to select new improvement pro-
jects, which turns phases 6 and 7 into a cycle, although the entire
process may need to be repeated.
We also looked at different contexts for data quality namely,
data quality in an operational context, data quality in the database
world, data quality and the data warehouse, data mining, electronic
data interchange, and the Internet. While there is a distinct need for
data quality in each of these contexts, the actual work that is done with
respect to data quality is understated but will grow in importance.
4
ECONOMIC FRAMEWORK OF DATA
QUALITY AND THE VALUE PROPOSITION
The key to understanding the importance of quality data Ues with an

examination of the costs and benefits associated with data quality.
Without an economic framework to measure the detrimental effect of
poor data in an organization, it is difficult to make an argument for the
investment in a knowledge management program.
Most frequently, poor data quality is measured by anecdotes or
descriptions of hazy feelings of difficulty in implementation, as well as by
tales of customer dissatisfaction. While these somewhat vague measure-
ments seem to verify that something is wrong, they give no clear indica-
tion of the severity of this "wrongness," nor do they offer solutions.
If there is a more concrete means of measuring the impact of poor
data quality on an economic basis, we can then determine the extent to
which the bottom line is affected by bad information. This is a good time
to discuss a framework for categorizing the costs and benefits in eco-
nomic terms. There are five interrelated steps to evaluating the cost-effect
of low data quality^ or COLDQ. The first step involves charting the flow
of data through the organization. The second is categorizing the costs
associated with bad data. The third is identifying and estimating the
actual effect of the presence of poor data quality. The fourth step is deter-
mining the cost of fixing the problems, whether by a static operation or a
continuous improvement program. The fifth and final step is the calcula-
tion of a return on investment to help senior-level management make a
decision as to the level of commitment to a data quality program.
In this chapter, we abstract the flow of information from original
source to final information product, whether the ultimate consumer of
that product is an operational process, a decision process, or customer.
73
We then develop a taxonomy of data quality cost categories. We look at

how to identify "break points" along the data flow/information chain
where data quality problems might arise. This is followed by a descrip-
tion of the analysis used to measure the actual costs of poor data quality.
The economic framework outlined in this chapter combined with the
more detailed qualification of data quality problems as described in
Chapters 5 through 8 and the current state assessment process described
in Chapter 9 will compile a complete "data quality score card" that can
guide an organization's path to better knowledge management.
4.1 EVIDENCE OF ECONOMIC IMPACT
How do we know that there is an economic impact on a company due

to poor data quality? Most likely, there will be evidence of an impact.
Let's look at some common issues that might be related to poor data
quality.
4.1.1 Frequent System Failures and Service Interruptions
An information-based system that experiences frequent system failures

or systems interruptions may be highly sensitive to poor data quality.
Failures may be caused by irregular or unexpected data values cropping
up in processing. Time spent searching for errors and fixing them before
restarting the processing will account for service interruptions.
4.1.2 Drop in Productivity vs Volume
As the volume of processing (transactions, data input) increases, a sys-

tem should scale appropriately and be able to handle the increased vol-
ume and meet the need with only an adjustment of the base resources
(processor power, storage, bandwidth). Any significant increase in other
resources (support staff, programmers, diagnosticians) may indicate an
inability to scale properly, which in turn might be associated with
inability to tolerate higher levels of bad data. If productivity decreases
as volume increases, this may be an indication of low data quality.
ECONOMIC FRAMEWORK OF DATA QUALITY AND THE VALUE PROPOSITION 7 5
4.1.3 High Employee Turnover
While it is a stretch to claim that high employee turnover is related to

poor data quality, there is certainly frustration associated with enter-
prise-wide employee mistrust. When the senior management is stymied
(due to insufficient or incorrect input assumptions) in its ability to set
and execute strategic initiatives on behalf of the company, the staff loses
faith in the core management's ability to get things done. This loss of
faith translates into a more mobile staff eager to bolt for more promis-
ing opportunities.
4.1.4 High New Business/Continued Business Ratio
It is said that the best new customers are your current customers. In
other words, you are more likely to sell something to your current cus-
tomer base than to new customers. So when your organization is
always focusing on new business and not closing deals with current cus-
tomers, this may be evidence of customer dissatisfaction, which can be
related to poor data quality.
4.1.5 Increased Customer Service Requirements
A high demand for customer service indicates that problems are leaving
the enterprise and making it to the customers. When the customer ser-
vice budget increases especially when the customer service depart-
ment mostly attends to billing problems this is likely indicative of a
data quality problem.
4.1.6 Decreased Ability to Scale
Small data sets with problems result in small sets of errors, but as data
sets grow, so do the size of the error sets. In some cases the growth of
the error sets is linear, but it can increase exponentially compared to the
source data. If the number of errors multiplies rapidly, the organization
will not be able to scale its systems to address those problems.
4.1.7 Customer Attrition
When customers stop being your customers, it is even stronger evidence

that some area of dissatisfaction has not been successfully resolved.
This is a more serious indication of data quality problems and shows
how much impact low data quality can have on an organization.
4.2 DATA FLOWS AND INFORMATION CHAINS
Although it is clear that data are used in both operational processing and
decision-making processes, outside the implementation arena, these
processes are often considered "black boxes" that take input data as "raw
material" and generate value-added information as output. This output
then proceeds to another processing operation (or another black box), or
it is summarized as a report as input to a decision-making process.
The first step in understanding the effect of low data quality is
peering inside the black box to identify the steps through which the
input data is converted into usable information. For simplicity, let's
divide the world into one of two data flow models: the strategic data
flow, used for decision-making, and the operational data flow, used for
data processing. Either model represents a data processing system, and,
also for simplicity, let's reduce the number of processing stages to an
abstract minimum. Note that there are, of course, exceptions to these
generalizations, but we'll use them as the generic model for determining
the COLDQ.
These are the processing stages that we will work with.

1. Data supply Data suppliers forward information into the system.
2. Data acquisition This is the processing stage that accepts data
from external suppliers and injects it into the system.
3. Data creation Internal to the system, data may be generated
and then forwarded to another processing stage.
4. Data processing Any stage that accepts input and generates
output (as well as generating side effects) is called a data process-
ing stage.
ECONOMIC FRAMEWORK OF DATA QUALITY AND THE VALUE PROPOSITION 77
5. Data packaging Any point that information is collated, aggre-

gated, and summarized for reporting purposes is a packaging
stage.
6. Decision making The point where human interaction is
required is called a decision-making stage.
7. Decision implementation This is the stage where the decision
made at a decision-making stage is executed, which may affect
other processing stages or may affect a data delivery stage.
8. Data delivery This is the point where packaged information is
delivered to a known data consumer.
9. Data consumption Because the data consumer is the ultimate
user of processed information, the consumption stage is the exit
stage of the system.
4.2.2 Directed Information Channels
An information channel is essentially a pipeline for the flow of informa-

tion from one processing stage to another. A directed information chan-
nel is attributed with the direction in which data flows. The delivery of
supplier data to an acquisition stage is performed through an informa-
tion channel directed from the supplier to the acquisition stage. It is
preferable to make information channels directed because there may be
bidirectional communication between any two points, and it is possible
to differentiate between data flowing in the different directions.
4.2.3 Information Chain Map
A map of an information chain is represented by the combination of the

processing stages connected by directed information channels. Because
we need to analyze each location in the information chain where a data
quality problem might occur, we need a naming scheme for the sites on
the map. Each processing stage should be labeled with a name and attrib-
uted with its stage type. There may be multiple information channels
between any pair of processing stages, which indicates multiple commu-
nications between those stages. Information channel names may either be
a combination of the source and target processing stages, separated by a
hyphen, along with a distinguishing number (such as "Customer-Hotel
Agent 1"), or (as seen implicitly in Figure 4.1) tagged with the name of
Hotel
DB
\
Query Query Credit
Results Card Info,
Customer ID,
Amount
Dates, Credit Dates, Credit
Card Info Card Info 1
Reservation Credit Card
Customer
System Processing
Approved/
Declined
Confirmation
Processing
FIGURE 4 . 1 Sample information chain
the source and the target processing stage as well as the data items that
are communicated through that channel.
4.2.4 Representing the Information Chain
The information chain can be represented in a pair of data tables as ref-

erence metadata. There are two types of objects that must be repre-
sented: processing stages and communications channels. For each
processing stage, we must keep track of what type of processing stage it
is and a unique name. The stage type refers to an enumerated list of
known processing stages.
create table stages (
stagelD integer,
description varcharClOOO),
stageType integer
);
For communications channels, we need the source and the target of

the channel. The stage identifiers for the source and the targets are indi-
cated.
create table channels (
channelID integer,
description varchar(lOOO),
source integer,
target integer
);
4.2.5 Strategic Data Flow
A strategic data flow represents the stages used for the decision-making
process, as shown in Figure 4.2.
It is likely
There may be many of these
to have J 1
processing and packaging
more than &1 steps in actual processing.
one data 4> .
suppliei: 2 !
Customer
Presentationy
FIGURE 4 . 2 A strategic data flow

4.2.6 Operational Data Flow
An operational data flow represents the automated repeated process of

transaction processing or result-oriented computing (that is, applications
that return specific requested results). In this case, the system is the oper-
ationalization of a known process (or set of processes), and as long as the
input data are correct, the data that pass through the system are, to some
extent, irrelevant to the process, instead acting as "grease" to the system
mechanism. The operational data flow is shown in Figure 4.3.
4.3 EXAMPLES OF INFORMATION CHAINS
4.3.1 Order Fulfillment
In a traditional direct marketing order fulfillment process, orders are

received, payments are processed, and products are packed and shipped.
The order fulfillment business is one of economy of scale the business
kisiikety
to have
t There may be many of these
pcocessiiig and pidcsîog
mcxtdian cS 1 steps in adal procesaag.
one data
siipf^ o 1
^\
sI
Data Customer
Supply Presentation .
There may be
nniifiple customers
for the same
packaged data.
FIGURE 4 . 3 An operational data flow

thrives when larger volumes of orders can be handled, and it therefore

fits into our category of an operational data flow, since the process is
well defined and tangential to the actual data received.
There are a relatively small number of processing stages. The data
suppliers are typically agents who will outsert (open envelopes) and sort
mail. Alternatively, orders may be taken by telephone. Either way, an
electronic version of orders is forwarded to a data acquisition stage,
which can be called "order reception." The next stage is a data packag-
ing stage, where orders being paid for by credit card are extracted, fol-
lowed by a data delivery stage where those records are passed to the
bank (a data consumer). Payments that are declined are returned as
supplied data, held temporarily (a processing stage), and then resubmit-
ted to the bank. Orders that have been paid for and can be filled are
packaged again and sent to another department shipping prepara-
tion where the actual products purchased are plucked off the shelf
and put into boxes for shipping.
If the ordered items are out of stock, the orders are placed in a back-
order location, and an order to refill stock is generated and sent to the
supplier. When back-ordered items arrive, the back-orders are examined
to see if they can befilled,in which case they are then sent to the picking
and packing process for shipping. Figure 4.4 shows this entire process.
FIGURE 4.4 Order fulfillment

4.3.2 Credit Approval
Any organization that is in the business of extending credit will have a

process that, based on a set of inputs associated with an applicant, will
determine whether the credit request is approved. Basically, this is an
information chain that will ultimately provide a single yes-no answer
based on the input values and some internal determination application
and therefore falls into the class of a strategic data flow.
In the process, data is supplied from both the applicant and exter-
nal agencies that collect data associated with creditworthiness. A data
acquisition stage merges the data from the two sets of sources and for-
wards those records to a data processing stage, where the values of the
records determine an answer. If the answer is yes, the answer is pack-
aged and forwarded to another department, where the applicant is pro-
vided with the details of the approval (a credit card is imprinted with
the applicant's name). An additional data packaging stage prepares a
report detailing the credit application that is forwarded to an external
data consumer, who may be the same data agent that originally sup-
plied the process with data (see Figure 4.5).
4.3.3 Consumer-Directed E-Commerce
Interestingly, Web-based businesses combine aspects of the two data

flows. We can focus on the user-oriented component of Web browsing.
The browsing process is relatively mechanical but involves the user as
data supplier, either through direct data acquisition interfaces or more
directly through the user's clickstream. At the Web server location, the
supplied data is acquired, then fed into a process to determine the next
set of information to deliver to the user's browser. That information is
packaged according to the http protocol and is delivered through the
protocol back to the user's browser. At this point, the user faces a deci-
sion stage, where the choice is made as to the next set of pages to
request from the server. Ultimately, the entire process is geared toward
leading the user to make a decision to purchase a product.
A different aspect of the information chain occurs when the user
decides to purchase a product. At that point the service must acquire
information from the user. At that acquisition stage, the order is broken
into the product order component, which is sent to a processing stage to
determine if the product is in stock, and a payment component, which
CoUected information
6rom all the credit
reporting agencies
is merged at this point
FIGURE 4 . 5 Credit Approval
is injected into a payment processing flow similar to the order fulfill-

ment flow described in Section 4.3.1.
4.4 IMPACTS
We'd like to assume that if there are no data quality problems, either of
the data flows described in Section 4.1 will operate smoothly. It is reason-
able to rely on decisions based on valid data and that an operation system
will function smoothly as long as no invalid data items gum up the works.
Issues appear when the information chains involve low quality data.
The effects of low data quality propagate through the systems, ultimately
leading to poor decision making, tactical difficulties, increased costs.
reduced revenues, and lowered customer satisfaction. Similarly, improve-

ments in data quality can lead to reduced costs, increased revenues,
streamlined and effective decision making, and increased customer satis-
faction.
Impacts can be divided into soft impacts, which are clearly evident
but still hard to measure, and hard impacts, whose effects can be esti-
mated and measured. Ultimately, the level of data quality rolls up to the
company's bottom line: Allowing low levels of data quality to remain
will lower profits, while improving data quality should increase profits.
4.4.1 Hard Impacts
Hard impacts are those whose effects can be estimated and/or mea-
sured. These include the following.
Customer attrition
Costs attributed to error detection
Costs attributed to error rework
Costs attributed to prevention of errors
Costs associated with customer service
Costs associated with fixing customer problems
Time delays in operation
Costs attributable to delays in processing
4.4.2 Soft Impacts
Soft impacts are those that are evident and clearly have an effect on pro-
ductivity but are difficult to measure. These include the following.
Difficulty in decision making
Costs associated with enterprise-wide data inconsistency
Organizational mistrust
Lowered ability to effectively compete
Data ownership conflicts
Lowered employee satisfaction
4.5 ECONOMIC MEASURES
Simply put, how do we measure a cost or a benefit? Since we are trying

to figure out the total economic benefit of improved data quality as
compared to the economic detriment of poor data quality, each impact
must relate to some quantification that goes directly to a company's
bottom line. Again, for simplicity's sake, we boil down these measures
into these categories.
1. Cost increase This measures the degree to which poor data
quality increases the cost of doing business.
2. Revenue decrease This measures how low data quality affects
current revenues.
3. Cost decrease This measures how an improvement in data
quality can reduce costs.
4. Revenue increase This measures how improving data quality
increases revenues.
5. Delay This measures whether there is a slowdown in produc-
tivity.
6. Speedup This measures the degree to which a process's cycle
time can be reduced.
7. Increased satisfaction This measures whether customer satis-
faction, employee satisfaction, or shareholder satisfaction is
increased.
8. Decreased satisfaction This measures whether customer satis-
faction, employee satisfaction, or shareholder satisfaction is
decreased.
In each of these categories it is possible to measure with precision
the actual economic impact, but because some impacts are difficult to
tie down to precise dollar amounts, they must be estimated at orders of
magnitude.
4.6 IMPACT DOMAINS
Low data quality has an impact on the operational domain, the tactical
domain, and the strategic domain. Within each domain, the different
kinds of cost measures and their effect on the economic model must be
evaluated. Note that in all three domains relying on incorrect or unfit
data will have a noticeable impact.
4.6.1 Operational Impacts
The operational domain covers the aspects of a system for processing

information and the costs of maintaining the operation of that system.
Operational issues typically are characterized as short-term issues, dis-
crete in scope and reactive by nature. These impacts can be character-
ized as "immediately reactive."
As an example, fixing a broken pipe is operational because the action
is in reaction to some event (the bursting of the pipe), it is limited in scope
(does not represent a major investment in capital improvement but is
meant to treat a problem), and it can be fixed using a short-term solution.
Operational issues revolve around "how to do" or "when to do" some-
thing, and they are usually delegated to operations employees. Data qual-
ity problems that escape detection within the operational domain
manifest themselves internally to the operational system (in other words,
they gum up the works) and frequently find their way to customers.
4.6.2 Tactical Impacts
The tactical domain covers the aspects of "what to do" as opposed to

how or when. Tactical issues are typically medium-term decisions made
to address system problems before they arise. In other words, we can
refer to these impacts as "reactive proactivity." Continuing with our
pipe example, the decision to replace the plumbing infrastructure
because of a burst pipe is a tactical decision. These kinds of decisions
are usually made by those responsible for making sure things run
smoothly, such as middle-level managers.
4.6.3 Strategic Impacts
The strategic domain stresses the decisions that affect the longer term.
Strategic issues are proactive, less precise decisions that address "where
to be" along a long time period. The burden of strategic decisions falls
to the senior executives of an organization.
4.7 OPERATIONAL IMPACTS
The nature of the operational domain makes it easy to focus on hard

impacts associated with low data quality. Data quality problems may be
propagated from a supplier of data at the acquisition stage, can be
introduced during any number of internal data creation stages, or can
be introduced at the time of data packaging. We can associate these
problems into cost categories: detection, correction, rollback, rework,
prevention, and warranty costs.
As errors propagate out of the enterprise to the customer base,
there is a risk of customer dissatisfaction, leading to lost revenues. We
can classify risk areas for lost revenue as: spin, reduction, attrition, and
blockading.
4.7.1 Detection
Detection costs are those incurred when a data quality problem pro-
vokes a system error or processing failure, and a separate process must
be invoked to track down the problem. Error detection only happens
when the system has the ability to recognize that an error has occurred.
Sometimes this is implied by a total system failure, such as an incor-
rectly provided divisor of 0 that causes a system interrupt for dividing
by zero. Sometimes this is implied by an abnormal end during transac-
tion processing because of an invalid data record.
The cost of error detection is mostly associated with three activi-
ties: determining where the failure occurred, determining what caused
the system failure, and determining the seriousness of the problem. This
cost is mostly attributable to employee activity, although there are also
costs associated with the purchase and maintenance of diagnostic tools.
4.7.2 Correction
Correction costs are associated with the actual correction of a problem

as well as the restarting of any failed processes or activities. The amount
of time associated with the activity that failed and extraneous employee
activity are all rolled up into correction costs.
The critical point of correction is the earliest location in the informa-
tion chain where the effect of fixing bad data is the expected execution of
the process. Correction involves figuring out what the incorrect item
should have been and then searching for the critical point of correction.
Correction may require a modification to data, a modification to process-
ing (software or operations), or both. The cost of correction encompasses
all these activities.
4.7.3 Rollback
When work that had been performed needs to be undone, rollback

costs are incurred. Rollbacks may be a straightforward undoing of a
number of transactions or a complicated full restoration from backup.
4.7.4 Rework
When a processing stage must be repeated because of an error and the

required correction, there are rework costs. Rework represents all work
that was additionally performed before the successful run took place.
We expect that all system processes will run correctly each time
they are invoked. When a process must be restarted to complete a suc-
cessful run, the work that was already performed must be accounted to
rework.
4.7.5 Prevention
Prevention costs arise when a new activity is designed, implemented,

and integrated to identify data quality problems and to take the neces-
sary actions to prevent operational failure due to unexpected data prob-
lems.
Frequently, prevention design and implementation costs can be
amortized over a period of time as long as the prevention system is still
in effect.
4.7.6 Warranty
Data quality problems that affect customers incur costs associated with
bothfixingthe problem as well as compensation to customers for dam-
ages. These are warranty costs. Any risks and costs associated with
legal action are also rolled up as warranty costs.
4.7.7 Spin
Spin control is a reaction to bad events that "makes them better." As

errors propagate to the customers, more and more complaints will filter
back to the organization. This in turn creates an increased need for cus-
tomer service improving the customer service systems, increases in
personnel, telephone costs, and so on. Since customer service needs are
directly tied to the quality of the product, a decrease in data quality will
cause an increase in the costs of customer service.
A very simple, yet measurable, example of this is in telephone
charges. Let's assume that a company has set up a toll-free telephone
number for customer support. Because it is a toll-free number, the com-
pany pays the costs on a per-minute basis. If the average on-hold time
for a customer increases from 5 to 10 minutes, that extra 5 minutes per
person translates into an increased per-call cost, and as the number of
complaints increases, so does the average on-hold time.
4.7.8 Reduction
Reduction occurs when a customer chooses to do less business with an

organization because of its data quality problem. When a customer
loses faith in a company's ability to properly conduct business, he or she
will avoid the company until its performance improves or until the next
revenue-reduction impact attrition is established.
For example, let's say we hire a billing agency to outsource our cus-
tomer billing. This agency bills us on a monthly basis based on the vol-
ume of service that the agency provides. If we begin to see errors in the
agency's monthly invoices, we would naturally doubt its ability to do a
good job.
4.7.9 Attrition
Attrition occurs when a customer's reaction to poor data quality results

in the customer no longer patronizing the business. This can be reflected
in the business-to-business world, as illustrated in the example in Sec-

tion 4.7.7, or in the consumer world.
It is not unusual to see errors on service or utility bills, requiring
frequent calls to customer service departments. As the customer service
requirements increase, this should be a signal that there may be a data
quality problem that is escaping the confines of the organization.
4.7.10 Blockading
Blockading is the result of customer dissatisfaction so pervading that it

alienates potential customers of the organization or causes current cus-
tomers to leave. Occasional data quality problems that are exposed to
customers may be tolerated overall, but the inability to resolve them
will result in customer attrition. These former customers will also dis-
courage others from doing business with the company.
This class of impact is prevalent in many places, including com-
plaints to a better business bureau. Many good examples of this can be
gleaned by reading Netnews postings complaining of billing errors or
service interruptions. Very often a consumer reports a bad experience
with some company and advises others to not patronize it. The speed
and breadth of distribution of information on the Internet can only
magnify this ability to spread negative opinions.
4.8 TACTICAL AND STRATEGIC IMPACTS
Enterprises collect data about the operation and effectiveness of the

organization. This information can be relevant to the internal opera-
tions as well as external activity. These data are often used as input to
decision-making processes. When information is incorrect or suspect,
that can lead to delays in decision making or the preemption of a deci-
sion process. Delays in decision making at a senior level lead to delays
in production called idling as well as increased difficulty in get-
ting things done. On the other hand, a high level of data quality will
enhance the decision-making process and prevent delays.
Other aspects of the tactical and strategic impact arenas include
lost opportunities, organizational mistrust, and cross-business unit mis-
alignment. There are also tactical maintenance costs associated with
low data quality. We can classify these costs as acquisition overhead,
decay, and infrastructure costs.
4.8.1 Delay
If data are not accessible, or the timely availability of the data is con-
strained, the decision-making process becomes delayed. A delay in
making a decision will spread to the operational arena as well, causing
productivity delays.
4.8.2 Preemption
If data are deemed untrustworthy, managers may decide to defer the

decision making until the right information is available or choose to not
make a decision at all. If the data are untrustworthy and the manage-
ment is not aware of any problems, a spurious decision based on faulty
data may be made, and it may ultimately affect the bottom line.
4.8.3 Idling
Idling takes place when delays in the decision-making process extend

the amount of time needed to complete a project. When productive
team members are waiting for executive decisions to be made, they can-
not take the steps that must be taken to implement those decisions.
4.8.4 Increased Difficulty
Increased difficulty occurs when the information needed to assess a

process, create a data warehouse, or reengineer a process is incorrect or
unavailable. Additionally, increased difficulty can be encountered when
team members implement incorrect or suboptimal decisions.
4.8.5 Lost Opportunities
As in the operational realm, data quality impacts of reduction, attrition,

and blockading can have a serious impact on strategic initiatives.
4.8.6 Organizational Mistrust
Because managers encounter inconsistencies in the data they solicit to

make decisions, they often choose to implement their own information
acquisition and maintenance systems, frequently using the same data

from the same data sources. This decision leads to redundant work
when multiple teams build and maintain the same data. This also leads
to inconsistency across the organization as entropy sets in.
The redundancy of data ownership eventually exacerbates territo-
riality in the company. The bunker mentality expressed in terms of "my
data is better than your data" or the bottlenecks created when an
employee no longer knows whom to contact to get a report all lead to
decreased productivity and increased costs.
This kind of mistrust also can explode past the corporate bound-
ary. A company's inability to set and execute business strategies leads to
shareholder mistrust as well, resulting in decreased valuation of the
business as a whole.
4.8.7 Misalignment
The issue of organizational mistrust compromises the ability to keep

different business units aligned. Delays in decision making and silo
mentalities divert executive management attention and allow a decen-
tralization of organization and a decrease in corporate agility as the
mistrust grows.
4.8.8 Acquisition Overhead
This is the cost associated with modifying or manipulating supplied

data in order to make it usable within the system. This cost is incurred
in two situations. The first is when supplied data do not conform to the
specific needs internally and must be modified, merged, joined, or oth-
erwise manipulated to prepare it for processing. The second is when the
quality of the data supplied does not meet the internal expectations and
additional processing is performed to bring the level of quality up to a
usable point.
4.8.9 Decay
Information that is kept and never used or updated is subject to a

degree of decay. Any activity that is used to keep stagnating information
up to date, as well as the overhead of maintaining the information (stor-

age, backup, etc.) is classified as a cost of decay.
4.8.10 Infrastructure
The costs of developing and maintaining an enterprise-wide shared data

repository, as well as creating and maintaining redundant databases,
are associated with infrastructure. This includes the construction of any
reference databases or data warehouses, distribution data marts, infor-
mation packaging, information delivery, and information display.
4.9 PUTTING IT ALL TOGETHER THE DATA QUALITY SCORECARD
Now that we have all the pieces for our economic model, let's look at
the actual steps involved in building it.
1. Map the information chain to understand how information flows
within the organization.
2. Interview employees to determine what people are doing with
respect to data quality issues.
3. Interview customers to understand the impacts on customers.
4. Isolate flawed data by reviewing the information chain and locat-
ing the areas where data quality problems are manifested.
5. Identify the impact domain associated with each instance of poor
data quality.
6. Characterize the economic impact based on the ultimate effects
of the bad data.
7. Aggregate the totals to determine the actual economic impact.
8. Identify opportunities for improvement.
The result is what we can call a data quality scorecard, shown in
Figure 4.6. This scorecard summarizes the overall cost associated with
low data quality and can be used as a tool to find the best opportunities
for improvement.
4.9.1 Mapping the Information Chain
The first step in the analysis is mapping the information chain/data

flows. Before assessing the impacts, the locations of the sources of those
Use employee^
interviews to
understand what
people are doing
widi respect to
data quality^
Map the
înformation chain ^ Isolate flawed Identifythe impaa\ / Characterize the >
to understand how /data by reviewing the\ Aggregate the total
domain associated \ / economic impact
information flows information chain with each instance j""H associated with each to determine the
within the and finding data of poor data / \ instance of low data actual economic
organization .quality problemsv V quality y \^ quality y impact
Conduct
customer
interviews to
^ understand customer ^
impacts
FIGURE 4.6 Steps in the Data Quality Scorecard
impacts within the system must be pinpointed. The information chain is

a map composed of the processing stages connected by directed infor-
mation channels. Each stage is annotated with its designation from Sec-
tion 4.2.1.
4.9.2 Identifying the Data Flow
Once we have an information chain, the next step is to determine what

data is being used within the system and at which information channel
source and target points the data pass through. If possible, the record or
message structure should be detailed, so that we can directly associate
any error conditions with the specific data set in which the error occurs.
4.9.3 Employee Interviews
To understand the impact of flawed data within the organization,

employees are interviewed. For each stage in the information chain,
interview the people doing the actual work to determine the amount of
time spent associated with each area of impact. The time for all employ-
ees can be aggregated into one time value.
4.9.4 Customer Interviews
To assess the impact due to decreased customer revenue, current and

former customers are interviewed to determine the reasons for any
decrease in business or attrition and blockading.
4.9.5 Isolating Flawed Data
With the results from the interviews in hand, it is time to start annotat-
ing the information chain. At each point where a data set is sent,
received, or manipulated, any locations of a source of a data flaw are
noted, along with a list of the activities attributable to those flaws.
4.9.6 Identifying the Impact Domain
With an information chain annotated with the list of both data flaws
and the activities associated with each of those flaws, it is time to start
attributing the flaws and activities to impact domains. For each source
of low data quality, the impact domains are selected, and each activity
is classified according to the classifications described in Sections 4.7
and 4.8.
4.9.7 Characterizing the Economic Impact
We can now build a matrix associated with each data quality problem.
The first axis identifies the problem and its location in the information
chain. The second axis represents the activities associated with each
problem. The third axis denotes the impact areas for each activity. In
each cell in this matrix, we insert the estimated cost associated with that
impact, using the economic measures from Section 4.5. If no estimate
can be made, an indication of the order of magnitude of the impact
should be used.
Note that this matrix does not distinguish between hard and soft
impacts. The values assigned to each cell can represent actual dollar val-
ues or coded indications of level of impact. Figure 4.7 shows the data
quality scorecard matrix.
Data Information
Reference Quality Chain
ID Problem Location Activity Impact Cost
1 Malformed
credit card Credit Card
Numbers Node 5 Processing Detection $ 12,000.00
Contact
Customer Correction $ 7,000.00 1
Rework $ 20,000.00 1
2 Invalid Direct
addresses NodeO Marketing Detection $
Correction $ 20,000.00 1
Reduced Acquisition
Reach Overhead $ 4,500.00 1
Lost
Opportunity $ 9,000.00
3 Incorrect Shipping
pick lists Node? Processing Detection $ 25,000.00
Correction $ 21,000.00 1
Customer
Service Warranty $ 43,000.00
Spin $ 12,000.00
Attrition $ 50,000.00
FIGURE 4.7 The Data Quality Scorecard matrix
4.9.8 Aggregating the Total
The matrix described in Section 4.9.7 can be superimposed on a spread-

sheet from which an aggregation model can be built. The costs can be
tallied and summarized in different ways and can be used as input to the
next stage, improvement.
4.9.9 Identify Opportunities for Improvement
The last component of this framework is using the model to look for the
biggest "points of pain." Having categorized the location and the
impacts of the different data quality problems, the next logical step is to
find the best opportunities for improvement, where the biggest value
can be gotten with the smallest investment.
4.10 ADJUSTING THE MODEL FOR SOLUTION COSTS
The final piece of the puzzle is computing a return on investment for

improvement projects. Since there are costs associated with any
improvement project, it must first be shown that the cost of the project
is justified overall. We do this by adjusting our spreadsheet model to
include the costs of the improvement projects but offset by the value
yielded by the improvement project. The result is an environment for
calculating return on investment (ROI) or break-even points for
improvement project implementation.
For any suggested improvement, the cost of designing and imple-
menting the improvement is added to the model, along with a time
frame for the implementation of the improvement. Each improvement
must correspond to the elimination of at least one cost impact. ROI is
calculated based on the decrease in costs (or alternatively, increase in
revenues) versus the cost associated with the improvement project. This
analysis completes the economic framework.
4.11 EXAMPLE
As an example, consider our information chain example of order fulfill-

ment trom Section 4.3.1. One problem frequently occurs during the
payment processing stage, when credit card numbers and their expira-
tion dates are forwarded, along with the charges, to the merchant bank.
There are two common, significant errors: invalid credit card numbers
and invalid expiration dates. In either case, the charge will not go
through successfully, with the result that (1) the payment cannot be
processed and (2) the fulfillment center must analyze the mistake to
determine the source of the problem.
Both of these activities have cost impact. So to determine the bene-

fit of inserting a vaHdation appHcation before the payments are sent to
the bank to check for those data quality violations, we must examine
the actual impacts and how the improvement affects the impacts.
In this case there are detection, correction, rollback, and rework
impacts. The detection occurs at the merchant bank site, as well as inter-
nally once the errors have been found. The correction may involve actu-
ally contacting the customer to get the correct information. The rollback
is associated with any revenue booking policies that need adjustment
because the payments did not actually go through. The rework is associ-
ated with any fulfillment processing performed on the unpaid orders.
Instituting the prevention apphcation will have an effect on the
detection, rollback, and rework impacts. The detection now takes place
earlier in the process than after the data is sent to a data consumer (the
merchant bank). Since the detection can be done before any other pro-
cessing is performed, we can eliminate the rollback and rework costs,
since they are preempted through early detection. On the other hand,
the correction impact remains the same if the card number or expira-
tion date is incorrect, it is still necessary to contact the customer for the
correct information.
Hopefully, our analysis will have told us how much time is being
spent on rollback and rework associated with this set of data flaws. Our
conclusion must be based on the cost of development and implementa-
tion of an error detection application, the time frame in which it can be
implemented, and the break-even point based on the costs eliminated
by the new application.
4.12 SUMMARY
In this chapter, we developed a framework for establishing the value

proposition behind a data quality improvement process. Our main
thrust is that there are real economic impacts of data quality problems,
and the first step in addressing these problems is to build a data quality
scorecard evaluating the cost of low data quality (COLDQ).
First, we looked at the factors indicating the existence of an eco-
nomic impact, including frequent system failures, service interruptions,
high employee turnover, customer attrition, and increased customer ser-
vice requirements, among others. Using these indicators, we can begin
to build the economic model.
The model itself consists of an information chain describing the

flow of data throughout a system, which is later attributed with details
about potential sources of data quality problems. An information chain
consists of a set of processing stages connected by information chan-
nels. These building blocks allow us to describe both operational data
flows and strategic data flows. We saw examples of different kinds of
data flows.
The next step in building our framework is the delineation of
impacts and impact domains. We characterize hard impacts as those that
can be easily measured and soft impacts as those that are demonstrable
but hard to measure. These impacts correlate to one of a number of eco-
nomic measures, representing increase or decrease in cost, increase or
decrease in revenue, increase or decrease in productivity, or increase or
decrease in satisfaction. The impact domains include the operational,
tactical, or strategic domains.
The last part of this chapter focused on putting the framework
together into a real scorecard, which itemizes all areas of the data flow
where poor data quality can affect the bottom line. This involved map-
ping the information chain, understanding the points of impact, charac-
terizing the actual cost, and identifying opportunities for improvement.
The value statement for each improvement can then be calculated in
terms of a return on investment that can map directly to known areas of
economic impact.
5
DIMENSIONS OF DATA QUALITY
The definition of poor data quality is similar to Justice Potter Stewart's

definition of obscenity: We know it when we see it. If we truly want to
improve data quality, however, we must find a way to measure it, and
the first step in measuring something is to define what that something is.
In this chapter, we try to define that "something" by listing the many
dimensions of data quality.
Good data quality is frequently defined in terms of "fitness for use."
Yet, it is difficult to delineate fitness when there are no metrics against
which to measure it. Therefore, before we discuss how to improve data
quality, let's first look at ways to measure it. The assessment of any data
set's levels of data quality whether data in a data warehouse or a
stream of related messages in an inline processing system can be done
in the context of what are referred to as the dimensions of data quality. It
is through the process of classifying requirements and setting measure-
ment goals that data quality can be improved, and the action of defining
this set of data quality dimensions is the beginning of this process.
The concept of data quality dimensions has been explored by
Richard Wang and the Total Data Quality Management group at the
Massachusetts Institute of Technology and is included in the book Data
Quality for the Information Age by Thomas Redman. These dimen-
sions can be used to identify users' data quality requirements, to delin-
eate a product's data quality feature requirements, to measure the levels
of data quality, and to identify the gaps and opportunities for data qual-
ity improvement.
The dimensions can be grouped by categories, distinguishing, for
example, aspects of data models from the aspects of data presentation.
101
For our purposes, we use the categories that include these collections of
data quality dimensions.
Data models
Data values
Information domains
Data presentation
Information policy
We will use an example data application for illustrative purposes.
First we will discuss the application of this data set, and the rest of the
chapter will examine what issues come into play when building the data
management system for this application.
5.1 SAMPLE DATA APPLICATION
The sample application will be a customer sales database consisting of

two parts: current customers and potential customers. For each cus-
tomer, name and contact information will be recorded. Sales informa-
tion about current customers will be kept on file.
5.2 DATA QUALITY OF DATA MODELS
Any information-based application is built on top of a logical informa-

tion framework called a data model. The data model describes the
objects represented within the application, the attributes those objects
take on, and the relationships between different objects within the
application. Therefore, before talking about the levels of quality of the
values in a data set, we must first talk about the suitability of the data
model to represent the information that the user needs. The crux of the
issue is understanding that while an application may mimic the real
world, a "data model" is just that a model.
Because of this, these data quality dimensions relate to ways of
characterizing and measuring the data model's correspondence to the
real world set of objects being modeled. In other words, adhering to the
issues described in this section drives the definition of data quality
requirements on the conceptual view of a data set, although most users
are usually only vaguely aware, if at all, of these requirements.
DIMENSIONS OF DATA QUALITY 103
The data model determines the ultimate usability of the informa-

tion being represented. Assuming the data are accurate and there are
appropriate mechanisms for capturing and presenting information, the
data model must support the users' needs transparently. The data qual-
ity dimensions of data models clarify what we look for when discussing
the data quality aspects of the representation of data.
5.2.1 Clarity of Definition
Clarity of definition refers to the nomenclature assigned to tables, fields,

and relations within the system. When describing the components of
the data model, the meanings of the names assigned to tables and fields
should clearly reflect the sets of information they represent. This clarity
should be reflected as unambiguous labels, as well as distinctive naming
of similar attributes.
In our example, we may have an attribute in the customer table
called "telephone number." Today, however, a telephone number may
have many different meanings as part of a customer account. It could be
a home number, an office number, a mobile telephone number, a per-
sonal toll-free number, a voice-mail system, or a fax number, among oth-
ers. In our case, we might only be interested in the telephone numbers
where the customer can be reached during telemarketing hours. For
business customers, this may be during "business hours," while for resi-
dential customers, this may be during "evening hours." In our case, we
would have an attribute named "marketing telephone number," which
could then be distinguished from the other contact telephone numbers.
5.2.2 Comprehensiveness
When designing a data model, it is important to determine a scope that

accommodates all the information pertinent to current users, as well as
that of future users. Comprehensiveness is a measure of how well that
scope has been covered.
There are two aspects to this dimension. The first is whether enough
information is being modeled to adapt to future uses of the application,
which implies that the users have thought about both their current and
future needs. The second aspect is whether enough information is being
modeled to support all the applications that might draw from that data
set, which implies that all stakeholders in the application suite have had
their say in the design of the model. If users are sharing information that
serves different purposes, there may be other comprehensiveness require-
ments. Is the model comprehensive enough to allow the users to distin-
guish data based on their independent needs .^
In our example, we are using both contact information and sales fig-
ures when representing our current customers. But with that same infor-
mation, the biUing department can also run its applications, although it
may be more important to the billing department that there be a set of
attributes indicating if the product has been shipped, if the customer has
been billed, and if the customer has paid the bill. Therefore, the data
model must be comprehensive enough to support both sets of require-
ments as well as enable the collection and support of extracting only the
information that each data consumer needs.
5.2.3 Flexibility
A data model's flexibility reflects the capacity to change in reaction to new

user requirements. Flexibility is similar to comprehensiveness, the differ-
ence being that comprehensiveness addresses what is planned a priori,
whereas flexibility addresses those situations that crop up but were not
part of the original plan. Restrictions on flexibility include situations
where the form of a data attribute's value carries extra information about
the attribute or where denormalization is built into the data model.
We can see an attribute carrying extra information in our example
if we designate an attribute for an account number for our customers. If
this attribute held seven-digit account numbers, where the first digit of
the account number indicates the region in which the customer lives,
that would be a restriction on flexibility. This would allow for 10
regions. If one region is spHt, making 11 regions, the account numbers
can no longer maintain the region indicator in the first digit.
5.2.4 Robustness
Robustness represents the ability to reflect changes in the modeled

world without excessive changes to the data model. Robustness
includes the foresight to build a model that can adapt to changes, as
well as the definition of attribute types and domains to hold the possible
values that each attribute might contain in the future. Robustness also
involves defining attributes in ways that adapt to changing values with-
out having to constantly update the values.
In our example, we might want to keep track of how many years a
customer has been associated with our organization. A nonrobust way
to do this is with an attribute containing the number of years that the
person has been a customer. Unfortunately, for each customer this
attribute will need to be updated annually. A more robust way to main-
tain a customer's duration is to store the date of initial contact. That
way, the number of years that the customer has been retained can be
computed correctly at any time without having to change the attribute.
The Year 2000 problem (Y2K), for example, evolved because of a lack
of robustness in many data models: a date attribute that has four digits to
hold the year is more robust than a date attribute with only two digits.
5.2.5 Essentialness
On the other hand, a data model should not include extra information,
except for specific needs like planned redundancy or the facilitation of
analytical applications. Extraneous information requires the expense of
acquisition and storage, and unused data, by nature of its being
ignored, will have an entropic tendency to low data quality levels. In
addition, redundant information creates the problem of maintaining
data consistency across multiple copies.
Another potential problem with unessential attributes is the "over-
loading" effect. With applications that have been in production, it
becomes very hard to modify the underlying data model without causing a
lot of stress in the application code. For this reason, when a new attribute
needs to be added, a behavioral tendency is to look for an attribute that is
infrequently used and then overload the use of that attribute with values
for the new attribute. This typically is manifested in program code with
conditional statements with tests to make sure that the overloaded
attribute is treated in the right manner (see Figure 5.1).
These kinds of conditionals are one basis for hidden business rules
that get buried in program code and/or are passed along as "lore"
within the information technology groups. The issue becomes a prob-
lem once the application has been in production for many years and the
original application implementers are long gone.
Current Total Yearly |

Name Customer Sales
John Smith G $1^00.00

Jane Jones N
Robert Kalmer M $200.00
Brian Kahn N
Ellen Fitzpatrick N
Jason Balder G $1,564.00
Kathy Knowles 1 P $10.00
Jacob Jansen 1 N
Errol Smythe G $1,295.00
Note that in dus attribute.

instead of just using *** and
^'N* to indicate whether the
party is a customer or not, if
the value is not **N,* die
attribute is used to indicate
the customer classtftcadon
as either "G* for good, "M**
for mediocre, or "P" for poot
Tins attribute is overloaded.

actually representing two
different attributes.
FIGURE 5.1 An overloaded attribute
5.2.6 Attribute Granularity
Granularity refers to the number of objects used to represent one notion.

If a data field can take on more values, it is said to be of a fine granular-
ity. If it can take on fewer values, it is said to be of coarse granularity. In
a table, more attributes may be used to convey finer granularity. While
fine granularity can provide more detailed information because of the
extra information being stored, there are added costs for storage, and
there is a greater possibility for introducing errors.
In our example, consider a table called "Customer Sales." Customer
sales may be measured by the day, week, month, quarter, or year. Repre-
senting customer sales by quarter, the data model may look like this.
DIMENSIONS OF DATA QUAUTY 107
Customer Sales = (customer id, product, Q l , Q2, Q3, Q4)

To gain a finer granularity, we can represent customer sales by
month instead.
Customer Sales = (customer id, product, Jan, Feb, Mar,
Apr, May, Jun, Jul, Aug, Sep, Oct, Nov, Dec)
5.2.7 Precision of Domains
As opposed to granularity of attributes, precision refers to the degree of

granularity that may be applied to an attribute's value. While the number
of attributes in a set of attributes characterizes attribute granularity, preci-
sion refers to the number of values that can be taken by a single attribute.
The precision is partially determined by the data type of the attribute, but
is also predicated on subtyping that is applied to the domain.
In our example, consider when we maintain a tally of the total
sales revenues generated from a certain customer. If we use a packed
decimal or fixed-size sales column along with a country code and a cur-
rency code indicating the country and currency, we may allow accept-
able precision when the currency is U.S. dollars but will not allow an
acceptable precision when the currency is Turkish lira.
Another case would be if we create a ZIP code field in the address
with five places. This may be sufficient for pre ZIP+4 United States
addresses, but it will not be able to accommodate Canadian or British
postal codes.
5.2.8 Homogeneity
Overloaded attributes were introduced in Section 5.2.5 with respect to

the overloaded use of nonessential attributes. Here we see the overload-
ing of value classes within a used attribute. Often, while a data model is
intended to maintain a single class of entities, the use of the model
evolves to overload the data set to maintain multiple classes of entities
within the same database.
In our example, this might occur when we want to be able to dis-
tinguish between different classes of customers, such as high-yield cus-
tomers who generate greater than $1,000 a month in sales versus
maintenance customers whose monthly sales are less than $200. In this
case, we have two different kinds of customers that are being main-
tained within a single attribute.
Usually, this subclassing is evidenced by "footprints" in the accom-
panying application code. These conditional statements and extra sup-
port code are representative of business rules that are actually embedded
in the data and are only unlocked through program execution. As this
happens more and more, the application code will become more and
more convoluted and will require some sort of reverse engineering to
uncover the business rules. The evolution of subclassing of attribute val-
ues will eventually necessitate the insertion of new attributes by the data-
base administrator to allow distinction between entities in each subclass.
5.2.9 Naturalness
Naturalness is a qualification of the relationship between what is being

modeled and the model's representation, meaning that each represented
attribute should pair with a natural object in the world being modeled.
Also, each attribute should represent a single fact about the object and
pull its values from a meaningfully natural (in other words, not artifi-
cially restricted) domain.
This is not to say that restricted domains are not a good idea. On
the contrary, natural value domains may be defined as a subset of a
"base" type. In our example, consider the account number (used above
in the flexibility dimension) to illustrate this dimension. Having assigned
the first digit of the seven-digit account number to represent the sales
region, we have an unnatural representation of account number. Should
a new sales region be opened, bringing the total to more than 10 regions,
the representation will no longer be sufficient to represent all possible
accounts. Additionally, should more than 999,999 accounts be opened
in any region, there will also be a problem. These problems, due to the
fact that one attribute is overloaded for information, can be fixed by
having a more natural representation, breaking out the region from the
account number and keeping it as a separate attribute.
5.2.10 Identifiability
In relational database systems, as well as in most data set collections,

where there is a presumption that each entity (used in the generic sense.
DIMENSIONS OF DATA QUAUTY 109
not the relational database sense) is unique, there must be some

absolute means for distinguishing between any pair of entities. If not,
there is a possibility that there are multiple records representing the
same individual entity. Therefore, each entity type must allow for
unique identification.
Identifiability is maintained via the presence of a primary key to the
data set. A primary key is a set of (one or more) attributes that, when
composed, form a unique reference into the data set. Because identifia-
bility is contingent on the uniqueness of the primary key, this dimension
drives an obvious data quality requirement that all primary key values be
unique, which, surprisingly, is not always the case. In our example, we
might assign a new attribute called "customer id" to each customer, and
we will make sure that any newly assigned customer id is unique.
Data sets whose models are not built with a specific primary key
must be checked for some kind of identifiability (perhaps through a
combination of attributes, if no single attribute will work). In Chapter
13 we will discuss how to determine a unique primary key if one is not
defined from the beginning. Once a primary key is discovered and the
user chooses to maintain that set of attributes, the primary key con-
straint must be established as a business rule and maintained as the data
set evolves.
5.2.11 Obtainability
Obtainability is a dimension that qualifies whether the information to

be modeled can be collected and stored. This obtainability may be char-
acterized by a measure of the ease and feasibility of collection and stor-
age, as well as more obtuse issues such as legality of information
collection or storage. Building a model to represent values that are diffi-
cult (or impossible) to obtain will probably result in having attributes
that will remain mostly empty.
In our example, an attribute totaling each customer's assets might
be useful when trying to determine the customer's ability to afford our
products, but it is unlikely that any customer will willingly provide that
information, and without a significant investment on our part, it will be
relatively hard to obtain. As a good example of a legal constraint on
obtainability, consider a Human Resources department database on
hiring and employee information. An attribute representing candidate
age is probably an unobtainable attribute.
5.2.12 Relevance
Obviously, it is important that the information being stored is relevant.

An irrelevant attribute is one that, if it were removed, would have no
effect on either the application as it is being run today or on any of the
planned features of the application.
In our example, maintaining a list of addresses associated with an
account is important, but if there is no additional attribute that indicates
the reason for each address's association with the account, then those
extra addresses are irrelevant. It is conceivable that with some additional
attributes, or with better definition clarity, what appears to be irrelevant
can be made useful. For example, if we have two address attributes but
one is called "billing address" and the other is called "shipping address,"
there is some distinction between the two addresses and therefore some
relevance.
5.2.13 Simplicity
The simplicity dimension refers to the complexity in the data model.

Complexity in a model, whether it is due to extra attributes or a compli-
cated relationship between attributes, can lead to errors. Complexity in
the data model manifests itself over time in terms of complications in
the associated application code. Another way of looking at complexity
is looking at how easy it is to understand the model. Simplicity is hard
to quantify, but, fortunately, its opposite is not so problematical
(although it's not that easy either!).
An example of complexity in a data set is if the account type
attribute is set to a value within one range, there must be a set of associ-
ated addresses for delivery of a product, but if the account type falls
within a different range, alternate delivery instructions must be speci-
fied. A "model normalization" might break the account types into more
homogeneous object classes and then specify the rules for attributes
within each object class separately.
5.2.14 Semantic Consistency
Semantic consistency refers to consistency of definitions among attrib-

utes within a data model as well as similarly named attributes in differ-
ent data sets. Semantic consistency is a dimension that characterizes the

degree to which the meanings and the names of objects within the data
set are consistent. One aspect of semantic consistency involves the
meanings of similarly named attributes in different data sets. The mean-
ings of these attribute names should be distinguished, or the attributes
should be assigned different names.
In our example, we are maintaining two tables, one for current
customers and one for potential customers. In each table, there is an
attribute for the name of the party. If both tables have that attribute
named "customer name," we could claim there is a semantic consis-
tency problem, since the potential customer database really does not
contain customers but only potential customers. This is a subtle differ-
ence, but the meanings are different and the distinction should be
noted.
Another example would be if that same customer name attribute
were meant to represent the name of a single party. Sometimes compos-
ite names appear in data sets, such as "DBA" names ("Doing Business
As") or trust accounts. What does it mean to have a composite name
entered in that field? Semantic consistency would require that if the
attribute represents a single party's name, then composite names are not
to be allowed. In fact, this particular issue is a very common occurrence
in customer and account databases (see Figure 5.2).
Another aspect of semantic consistency involves definitions of rela-
tions between tables. Data models contain many tables, and as applica-
tions mature, the size of the encompassing database grows. When this
happens, new relations between tables appear, whether planned or by
chance. Foreign key relations between tables based on a set of attributes
belonging to the two tables are very common. If two attributes are
assumed to be related, the definition of that relation should be made clear.
5.2.15 Structural Consistency
Structural consistency refers to the consistency in the representation of

similar attribute values, both within the same data set and across the
data models associated with related tables. Structural consistency is a
characterization of the care taken by the database modelers, adminis-
trators, and stewards in guaranteeing "strong-typing" of similar attrib-
utes. This dimension can be measured across two axes. The first regards
how different attributes of the same virtual type are represented, and
Fresumably, diese duee

attrUmtes bdd the same
vahiesy even t h o t ^ the
atcribiite names cUEfei:
CUSTOMER ADDRESS OTY STATE ZIP CUSTOMER Accr CREDrrCARD ORDER LIMIT
IACCINUM rrEM QTY TAX TOTAL
Example 1: The same attribute values

in different tables but named with
different names.
Two customer
names are listed.
^^îSs^^^^^^^mSmf^&^^^^^n^^S
Is a trust fund John Smith 123 Main Street, New York, NY 10011
a customer? John and Mary Smith 123 Main St., New York, NY 10011
Smith Family Trust, John Smith l i t 123 Main St, NY NY 10011
Three customer 123 Main Street, Apt. 4, NY NY
Mary Cummings Smith
names are listed*
> < John and Mary Smith UGMA John Jn 123 Main Street, Apartment 4, New York 10011
Arediesetwo Mary Smith 123 Main Street, Apt. 4, NY NY
custCHners the *
same person?
jcample 2: This attri bute is supposed

t 0 refer to customer rlame, but in these
r ecords, we see that t he field often
r epresents more than one custoinet
FIGURE 5 . 8 Breaks from Semantic Consistency
the second regards the degree to which there are different representa-
tions for the same value.
The first axis can be shown by this example: If our model has attrib-
utes for date of first contact and date of last sale, it is preferable to use the
same representation of dates in both attributes, both in structure (month,
day and year) and in format (American dates vs European dates). An
example of the second axis is measured by enumerating the different ways
that dates are maintained within the system. We will examine these issues
in the discussion of domains and mappings in Chapter 7.
5.3 DATA QUALITY OF DATA VALUES
Because low levels of quality of data values are the most obvious to
observers, when most people think of data quality, they think of these
most easily understood data quality dimensions. Low levels of data
value quality are likely to be recognized by both the users (as well as
customers!) and are most likely to lead the user to conclusions about
the reliability (or lack thereof) of a data set. In an environment where
data are not only being used to serve customers but also as input to
automatic knowledge discovery systems (that is, data mining), it is
important to provide high levels of data quality for the data values.
Relying on bad data for decision-making purposes leads to poor strate-
gic decisions, and conclusions drawn from rules derived from incorrect
data can have disastrous effects.
Data value quality centers around accuracy of data values, com-
pleteness of the data sets, consistency of the data values, and timeliness
of information. Most data quality tools are designed to help improve
the quality of data quality values. In Chapter 8, we build a framework
for describing conformance to dimensions of data value quality as a set
of business rules that can be applied to the data set and used to measure
levels of data value quality.
5.3.1 Accuracy
Data accuracy refers to the degree with which data values agree with an
identified source of correct information. There are different sources of
correct information: a database of record, a similar, corroborative set of
data values from another table, dynamically computed values, the
result of a manual workflow, or irate customers. Inaccurate values don't
just cause confusion when examining a database bad data values
result in increased costs. When inaccuracies reach the customers, costs
can increase due to increased pressure on customer service centers,
searches for the inaccuracies, and the necessity to rework the process.
In our example data set, an inaccurate shipping address will result
in errors delivering products to the customers. The repercussions may
be great: A customer will delay payment, cancel the order, and even
cease to be a customer.
5.3.2 Null Values
A null value is a missing value. However, a value that is missing may

provide more information than one might think because there may be
different reasons that it is missing. A null value might actually represent
an unavailable value, an attribute that is not appUcable for this entity,
or no value in the attribute's domain that correctly classifies this entity.
Of course, the value may actually be missing!
Even though databases may provide a default representation for
the null value, there may be times when a specific internal representa-
tion of a null value is needed. An example of a poor design for null
value representation is the use of 99/99/99 as a null date.
5.3.3 Completeness
Completeness refers to the expectation that certain attributes are

expected to have assigned values in a data set. Completeness rules can
be assigned to a data set in three levels of constraints.
1. Mandatory attributes that require a value
2. Optional attributes, which may have a value
3. Inapplicable attributes (such as maiden name for a single male),
which may not have a value
We will see that completeness can be prescribed on a single attribute
or can be dependent on the values of other attributes within a record or
message. We can also discuss completeness with respect to a single at-
tribute across all records or with respect to a record.
In our example, we decide that a record is considered incomplete if
the daytime telephone number attribute is missing. There are two impli-
cations: No entry in the "daytime telephone" column is empty and no
record may be inserted into the database if the daytime telephone
attribute is missing.
5.3.4 Consistency
Consistency can be curiously simple or dangerously complex. In its

most basic form, consistency refers to data values in one data set being
consistent with values in another data set. But what does consistency
really mean? If we follow a strict definition, then two data values drawn
from separate data sets may be consistent with each other, yet both can
be incorrect. Even more complicated is the notion of consistency with a
set of predefined constraints. We may declare some data set to be the
"database of record," although what guarantees that the database of
record is of high quality.^
More formal consistency constraints can be encapsulated as a set
of rules that specify consistency relationships between values of attrib-
utes, either across a record or message, or along all values of a single
attribute. These consistency rules can be applied to one or more dimen-
sions of a table or even across tables.
In our example, we can express one consistency constraint for all
values of a ZIP code attribute by indicating that each value must con-
form to the U.S. Postal Service structural definition. A second consis-
tency constraint declares that in every record, the ZIP code attribute's
value must be consistent with the city attribute's value, validated
through a lookup table. A third consistency constraint specifies that if
the ZIP code represents an area within a qualified geographic region,
the account specified by the account number field must be associated
with a salesman whose territory includes that geographic region.
The first consistency constraint applies to a single attribute. The
second applies to a relationship between two attributes within the same
record. The third constraint applies to values in different tables. Consis-
tency constraints can be arbitrarily complex as shown by these three
examples and they frequently reflect business rules inherent in the
applications using the data.
5.3.5 Currency/Timeliness
Currency refers to the degree to which information is current with the

world that it models. Currency can measure how up to date informa-
tion is and whether it is correct despite possible time-related changes.
Timeliness refers to the time expectation for accessibility of infor-
mation. Timeliness can be measured as the time between when informa-
tion is expected and when it is readily available for use.
In our example, we like to maintain customer addresses, but
because we live in a mobile society, many of our customers move each
year, leaving our address data slightly less than current. An example of
a timeliness issue deals with publishing our product price list on a Web
site. To make sure that we do not charge a customer the wrong price,
we need to guarantee that the time lag between a product price change
and the new price's appearance on the Web site is minimized!
5.4 DATA QUALITY OF DATA DOMAINS
A data domain, which will be discussed in greater detail in Chapter 7, is

a collection of values that are related somehow, by definition, and repre-
sent a recognized authority for data sourcing. Data domains are restric-
tions on larger sets of values. A data domain can either be descriptive,
which means that the set of values can be derived from some prescription
applied to the larger set of values, or enumerated, in which all the values
are explicitly listed.
A mapping exists between two domains when we express a rela-
tionship that associates any selected value in the first domain with one
or more values in the second domain. As with domains, mappings may
be descriptive or enumerated.
An example of a descriptive domain is the format definition for
valid Social Security numbers. An example of an enumerated domain is
a list of currencies accepted for payment for our products.
Domains and mappings are ubiquitous in databases. It is through
domain discovery and analysis that denormalized tables can be normal-
ized, and many data quality rules and business rules can be expressed in
terms of relations between domains and mappings. The data quality
dimensions associated with data domains include enterprise agreement
of usage, stewardship, and ubiquity.
5.4.1 Enterprise Agreement of Usage
The notion of abstracting information into a data domain implies that

there are enough users of the same set of data that it makes sense to
manage that data set separately as a resource instead of having separate
groups manage their own versions. The dimension of enterprise agree-
ment of usage measures the degree to which different organizations
conform to the usage of the enterprise data domain of record instead of
relying on their own data set.
In our example, a usage agreement that we might make is to decide
to use a shared ZIP code database to fill in the ZIP code field of cus-
tomer addresses.
5.4.2 Stewardship
A dimension of data quality is the degree to which responsibility has

been assigned for the stewardship of information domains. Since a data
domain represents a collection of data values that are recognized as a
data source of record, it is advisable to appoint a steward to be respon-
sible for the upkeep and maintenance of the data domains. The degree
of stewardship should increase as the agreement of usage across the
enterprise grows.
5.4.3 Ubiquity
As a data quality-oriented organization matures, the agreement of usage

will move from a small set of "early adopters" to gradually encompass
more and more of the enterprise. Ubiquity measures the degree to which
different departments in an organization use shared reference data.
5.5 DATA QUALITY OF DATA PRESENTATION
Data quaUty does not only apply to the way that information is repre-
sented and stored. On the contrary, there are dimensions of data quality
that are related to the way that information is presented to the users
and the way that information is collected from the users. Typically, we
would like to measure the quality of labels, which are used for naming
or identifying items in a presentation; classification categories, which
indicate specific attributes within a category; and quantities, which
indicate the result of measurement or magnitude of quantitative values.
We also want to look at formats, which are mappings from data to
a set of symbols meant to convey information. The format for repre-
senting information depends on the application. As an example, in our
application, we might want to report on sales activity within certain
geographical ranges. We can take a dry approach and deliver a report of
numbers of sales by product within each region, sorted by dollars, or
we might take a different approach in providing a map of the region
with color-coding for each product with intensities applied to indicate
the sales ranges.
Formats and presentations can take many different attributes.
Color, intensity, icons, fonts, scales, positioning, and so forth can all
augment the presentation of text or numbers stored in a data set. Even

though data may be stored in a traditional format, the presentation to
the user is one that gives a muhidimensional view of what information
is inherent in the data.
5.5.1 Appropriateness
Appropriateness is the dimension we use to categorize how well the for-

mat and presentation of the data match the users needs. In our example,
there is a difference between a high-level monthly sales report that is
supplied to senior management and the daily product manifests that are
handed to the shipping department for product packaging.
A data presentation is not appropriate if its intended users never
look at it. There is an apocryphal story that revolves around a com-
pany's enterprise-wide system renovation. Each day, thousands of
reams of paper reports were generated, and an entire department was
dedicated to the collation and dispersal of these reports. Because the
company wanted to reduce the amount of generated reports, the ques-
tion arose as to how to determine which reports could be eliminated.
Because of the size of the system, it was difficult to see which
reports depended on which other reports, so no decision was made until
eventually this idea was proposed: Delay the daily delivery of the
printed reports by one hour and wait to see who complained. That
would be a clear indication that those who did not complain did not
consider the reports critical, and these would be the first candidates for
elimination.
5.5.2 Correct Interpretation
A good presentation provides the user with everything required for the
correct interpretation of information. When there is any possibility of
ambiguity, a key or legend should be included.
In our example, consider an attribute that represents priority of
customer service calls, with the domain being integer values from 0 to
10. While it may have been clear to the data modeler that 0 represents
the highest priority and 10 the lowest, the presentation of that attribute
in the original integer form may confuse the user if he or she is not
familiar with the direction of the priority scale. The presentation of the
information, therefore, should not display the integer value of the prior-
ity, which might be confusing. Instead, providing an iconic format, such
as incremental intensities of the color red, to indicate the degree of
importance of each particular call. This is shown in Figure 5.3.
Here are four

examples of different
kinds of representations
of the same data
attribute. Different
representations
convey different
amounts of
In this presentation,
the shaded graphic
indicates critical Priority 2
priority with darker
shading.
N
N Priority
Low-priority level (= 2) Priority
FIGURE 5 . 3 Differences in representation allow different interpretation

5.5.3 Flexibility
Flexibility in presentation describes the ability of the system to adapt to

changes in both the represented information and in user requirements
for presentation of information. For example, a system that displays
different countries' currencies may need to have the screen presentation
change to allow for more significant digits for prices to be displayed
when there is a steep devaluation in one county's currency.
5.5.4 Format Precision
The degree of precision of the presentation of an attribute's value

should reasonably match the degree of precision of the value being dis-
played. The user should be able to see any value the attribute may take
and also be able to distinguish different values.
In many older systems, there was a limit on the number of signifi-
cant digits that could be shown forfloating-pointvalues. In these sys-
tems, the screen display of a value with more significant digits than the
allowed limit showed either a truncated or rounded value, neither of
which represented the actual value. If the value were too large or too
small, the value might be either displayed incorrectly or not appear at all.
5.5.5 Portability
In an environment that makes use of different kinds of systems and

applications, a portable interface is important so that as applications
are migrated from one platform to another, the presentation of data is
familiar to the users. Also, when dealing with a system designed for
international use, the use of international standards as well as univer-
sally recognized icons is a sign of a system designed with presentation
portability in mind.
5.5.6 Representation Consistency
This dimension refers to whether instances of data are represented in a

format that is consistent with the domain of values and with other simi-
lar attribute values. For example, the display of time in a nonmilitary
(12-hour) format may be confusing if all other instances of times in the

system are displayed in the 24-hour military format.
5.5.7 Representation of Null Values
When the null value (or absence of a value) is required for an attribute,
there should be a recognizable form for presenting that null value that
does not conflict with any valid values. This means that for a numerical
field, if the value is missing, it is not an indication that it may be repre-
sented to the user as the value 0, since the presence of any number there
may have different meaning than the absence of the value. Also, if there
are ways of distinguishing the different kinds of null values (see Section
5.3.2), then there should also be different ways of presenting those null
values.
5.5.8 Use of Storage
Over the past few years there has been an incredible reduction in the
cost of disk storage to the point where it seems silly to think about con-
serving disk space when building and using a database. Yet, just as the
interstate highway system encouraged travel by automobile, the high
availability of inexpensive disk space encourages our penchant for col-
lecting and storing data.
It is important to remember that even though disk space is inex-
pensive, it is not unlimited. A dimension of data quality, therefore, is in
the evaluation of storage use. This is not to say that the only issue is to
squeeze out every last bit. Instead, it is in investigating how effectively
the storage requirements are offset by other needs, such as performance
or ease of use. For example, the traditional relational database is
assumed to be in normal form, but in some analytical databases, data-
bases are specifically denormalized to improve performance when
accessing the data.
5.6 DATA QUALITY OF INFORMATION POLICY
5.6.1 Accessibility
The dimension of accessibility refers to the degree of ease of access to

the information. This includes both how easy it is to access the informa-
tion in the system and whether all the information can be accessed. This
also includes determining if the presentation allows for the display of all
expected information and whether the presentation is in a form that
allows the user to absorb it as well. This dimension refers only to that
information that is allowed to be presented to any selected subset of
users. Privacy and security are separate dimensions of data quality.
In our example, it is important that senior managers have easy
access to all sales data for each of the sales representatives for the past
18 months. This not only means that an 18-month sliding window of
the database must be kept available at all times, but that there should
also be application code to retrieve and display that information.
5.6.2 Metadata
Metadata is data about the data in the system. The dimension of data
quality policy regarding metadata revolves around whether there is an
enterprise-wide metadata framework (which differs from a repository).
Is it required to maintain metadata? Where is it stored, and under
whose authority.^ Metadata is particularly interesting, and we cover it
in Chapter 11.
5.6.3 Privacy
If there is a privacy issue associated with any data set, there should be a
way to safeguard that information to maintain security. Privacy is an
issue of selective display of information based on internally managed
permissions. It involves the ways unauthorized users are prevented
from accessing data and ensures that data are secured from unautho-
rized viewing. Privacy is a policy issue that may extend from the way
that data is stored and encrypted to the means of transference and
whether the information is allowed to be viewed in a nonsecure loca-
tion (such as on a laptop while riding on a train).
5.6.4 Redundancy
Data redundancy refers to the acquisition and storage of multiple

copies of equivalent data values. Planned redundancy is desirable in
that it provides fault-tolerance and may improve the accessibility of
information. It may even ease political issues of ownership. But redun-
dancy becomes unwieldy when organizations have a dozen or more
copies of the same data. This leads to the opportunity for copies to
become unsynchronized and for data to become stale. Unplanned
redundancy increases costs and increases the opportunities for consis-
tency flaws to creep into the enterprise system.
5.6.5 Security
The dimension of security is the protection of data from harm, unau-

thorized modifications, or unwanted destruction. Security is similar to
privacy, except that privacy deals with protecting the entities being
modeled by the system, whereas security protects the data itself.
5.6.6 Unit Cost
The costs incurred to obtain values, maintain levels of data quality,

store data, and so on all comprise the cost of maintaining information.
The cost of building data quality into an information product must be
weighed against the cost of not having data quality. This includes both
tangibles costs, such as the cost of rework and failed applications due to
bad input, and intangible costs, such as low customer satisfaction,
decrease in reputation, and loss of revenues.
5.7 SUMMARY: IMPORTANCE OF THE DIMENSIONS

OF DATA QUALITY
In this chapter, we have looked at the dimensions of data quality that

correspond to data models, data values, data domains, data presenta-
tion, and data policy. Understanding the notion of data quality dimen-
sions gives us a starting point to a set of variables that we can start to
measure, probe, and attempt to improve.
An interesting characteristic is that the different dimensions of data

quality take on different levels of importance to different organizations.
For some companies, just ensuring that data values are correct may be
the most important issue, while other companies care more about the
way that their information is presented. The critical point is that before
we can improve the quality of data, we must first choose those dimen-
sions that we care about most. It is at that point that we can begin to
define our expectations and then measure how the data meet those
expectations. Only then can we begin to improve data quality overall.
The list of dimensions of data quality is extensive, yet it is never
really finished. As we continue collecting information and trying to
make use of that information, we will find more and more ways in
which that information may yield unexpected surprises. Maybe you
already know an area of data quality that could be added to this list!
6
STATISTICAL PROCESS CONTROL AND
THE IMPROVEMENT CYCLE
In Chapter 5, we looked at the different dimensions of data quaUty,

with the understanding that in any system, there are a number of spe-
cific aspects of data quality that may be important to the users. In this
chapter, we see how we can use those well-defined dimensions and
specifically quantify the levels of data quality. We establish the tools
used for initiating the measurement and for determining at a gross level
the degree of conformance to data quality standards. In Chapters 7 and
8, we will define a framework for defining data quality rules, and in
Chapter 9 we will discuss ways to measure the data quality dimensions
discussed. Together, the next four chapters constitute the measurement
gauges we use in the evaluation and measurement of data quality.
In the early 1920s, Walter Shewhart at Bell Laboratories performed
a number of sampling studies that led to the development of a quality
tool known as Statistical Process Control. By evaluating the occurrence
of faults, defects, and errors in the manufacturing process, Shewhart dis-
covered that, just as in nature, there are all sorts of variations that can
occur during a manufacturing process. He determined that by studying
the different kinds of variations (that is, the ones that generate unusable
end products) and evaluating the root causes behind them, the occur-
rences of poor quality can be identified and the processes causing those
irregularities can be improved.
These notions don't just apply to product manufacture. They also
apply to data quality, and in this chapter we look at the use of Statistical
Process Control (SPC) as a way to provide a context for a continuous
data quality improvement cycle. SPC is a process of instituting measure-
ments during a manufacturing process to both control quality and
125
detect variations in quality as items are being produced, instead of find-

ing them during inspection after production.
In this chapter, we first explore the notions of variation and control,
followed by a discussion of the tools used for evaluating SPC. We look at
the construction of control charts and the use of the Pareto Principle for
identifying the variables that most affect the system. We then explore
control charts in greater detail, looking at some of the different kinds of
charts that can be used for analyzing SPC in the context of data quality.
6.1 VARIATION AND CONTROL
Variation occurs in many contexts, but in any quality program, we must

be able to distinguish between those variations that are meaningless and
those that have an adverse effect on the stability of the system. In other
words, how can we differentiate between normal, expected variations
and those that are business critical? And given the fact that different
business environments have different characteristics, it may be possible
that the same set of variations may be meaningless in one area and very
critical in another!
To bring these questions into focus, we must look at the different
kinds of variation within a system and the concept of a process in con-
trol. In any quality process, the goal is to narrow the possibility for
unexpected and unwanted variation in the manufacturing process. In
the data quality sense, since we consider the provision of information as
a manufacturing process, we can use the notion of quality control as a
means for monitoring data quality throughout a system. Statistical
Process Control is a tool used to monitor control, but before we explore
SPC, we must first understand the nature of variations and control.
6.1.1 Causes of Variation
Shewhart's studies focused on distinguishing between expected varia-

tions that occurred within a manufacturing process and any significant
fluctuations that indicated a problem with the manufacturing process.
He isolated two main causes of variation.
1. ChancCy or common, causes. These are minor fluctuations or
small variations in the end product that do not have to be fixed.
STATISTICAL PROCESS CONTROL AND IMPROVEMENT CYCLE 127
2. Assignable, or special, causes. These are specific causes of varia-

tion, resulting in significant fluctuation in the level of quality.
While specific variations from common causes are not predictable
in their own right, their occurrences are likely to form a pattern, and
Shewhart's observation was that they formed a normal distribution.
Because of this, we expect that when we measure variation in a process,
there are limits to the number of variations due to chance causes. The
implication is that when we see fluctuation in the process that exceeds
those limits, it is probably due to a special cause, which should then be
investigated.
An example might be tracking the on-time performance of a rail-
road line. The train's departure and arrival times can be recorded daily
and variations from the scheduled times measured. We would expect
that from day to day, the train will likely be a little early, on time, or a
little late. The times that the train is slightly early or late are due to com-
mon causes. But one day, a large electrical storm knocked out power
from the third rail, causing systemwide delays. On that day, the train
arrived an hour late, due to an assignable cause.
As another example, let's consider a simple information collection
process and look at the different kinds of variations that might occur.
Say we have several people transcribing names and addresses from
hard-copy lists into a computer sales database. We can expect that,
overall, most of the transcribers will make some kind of typographical
error, perhaps substituting one letter for another. These errors are all
due to common causes. Now let's say all of a sudden the "e" key on one
of the keyboards breaks, creating consistent errors from that computer.
This is an example of errors due to a special cause.
6.1.2 Statistical Control
According to the ANSI/ISO/ASQC standard A3534-1993 (Statistics

Vocabulary and Symbols)^ the state of statistical control is the state in
which the observed sampling results can be attributed to a system of
chance causes that does not appear to change with time. A process is in
control (that is, it is stable) if each of its quality measures is in a state of
statistical control.
What this means is that, having selected a set of variables or attrib-
utes of a process to measure based on random sampling, we expect that
if the system is in control, there will be a normal distribution of varia-

tions and that the specific occurrences of variations will be random. If
we observe that this is not the case, there must be some special cause to
which this pattern or variation can be attributed, which then must be
further explored.
6.2 CONTROL CHART
To determine whether an information process is in control, we must

sample data to see how well it conforms to our expectations of data
quality. We can sample data at predesignated points or at different
points in the information chain. In addition, samples must be taken
over a period of time to see whether there are any significant changes.
The data samples can be integrated into a simple tool to record data
quality measures over a certain time period that will highlight the dif-
ferences between chance causes and assignable causes of variation. This
tool is called a control chart, which is a graphical representation of the
variations produced from a process.
Simply, a control chart plots the values of a time series or a sample
series along with upper and/or lower control limits. A central line can
be shown to display the typical (or mean) behavior of the system and to
detect trends toward either of the control limits. Control limits are plot-
ted lines above and below the central line to bound the space in which
expected variations will occur.
Control limits are not defined by the customer but instead are
determined by observing behavior over the series. Because the common
causes form a normal distribution, wefirstneed to determine how that
distribution is reflected within the normal expectations. A process is
said to be stable as long as all the points fall between the control limits.
Figure 6.1 shows an example of a control chart.
6.3 THE PARETO PRINCIPLE
In any systems with causes and effects, the bulk of the effects are caused
by a small percentage of the causes. This concept, called the Pareto
Principle, has been integrated into common parlance as the "80-20
rule" 80 percent of the effect is caused by 20 percent of the causes.
This rule is often used to establish the degree of effort that must be
UCL
Center line
LCL
FIGURE 6 . 1 Control chart
expended on a particular project. If the rule of thumb is that 80 percent

of the benefit can be achieved with 20 percent of the work necessary for
completion, then the project will go forward, at least until the 80 per-
cent benefit has been achieved.
In fact, the Pareto Principle has a more interesting application in
that we use Pareto analysis to determine what aspects of a system (or in
our case, data quality) are to be incorporated into the SPC process. A
Pareto chart is a bar chart representing the measurement of specific
aspects of a system. The presentation of the chart is based on cumula-
tive frequency measurements of particular metrics, ordered from the
greatest to the least frequency. The chart highlights the areas responsi-
ble for the greatest percentage of a problem and the variables involved
in those areas (see Figure 6.2).
Another interesting corollary to the Pareto Principle is that as the
larger problems are solved, there is a diminished opportunity for subse-
quent improvements. In other words, once we have attacked and con-
quered the first three or four problems, it is not likely that we will
achieve significantly more improvements from attacking any additional
problems. This is actually quite reassuring, since it means that we can
limit the amount of effort to be expended on improving a particular
area!
4,?^
D
-i-'j&
FIGURE 6 . 2 Pareto chart
Pareto analysis is used in performance improvement. A computer

program that is not running up to speed is subjected to a series of profil-
ing processes that gauge the time spent in each individual functional
component. The professional performance technician then looks at the
function in which the most time was spent, with an eye toward improv-
ing the performance of that function. Let's say we had these data points.
Function Total Seconds
Foo 56
Bar 26
Baz 8
Boz 6
Raz 4
Faz 2
Together, these six functions account for 100 percent of the run-
time of the program, totaling 102 seconds. If we can speed up function
Foo by a factor of 2, we will have reduced the runtime of the entire
application by 28 percent (half of the time of function Foo), making the
total time now 74 seconds. A subsequent improvement in the function
Bar of a factor of 2 will result in only an additional 13 percent improve-
ment over the original runtime (actually, the effect is 18 percent of the
current runtime, due to the previous reduction in runtime from improv-

ing Foo).
If we then focus on improving the runtime of function Baz by a fac-
tor of 2, the best speedup we can achieve is now a mere 4 seconds, which
will only slightly improve our performance from 61 seconds to 57 sec-
onds. The same improvement in function Boz only reduces the speed by
another 3 seconds. As you can see, the same amount of effort expended
on making improvements results in a rapidly decreasing benefit.
By performing a Pareto analysis, we can use the results to focus
attention on the areas that are contributing the most to the problem.
The variables that contribute to these areas become the variables or
attributes that are to be incorporated into the control chart.
6.4 BUILDING A CONTROL CHART
Our next step is to build a control chart. The control chart is made up
of data points consisting of individual or aggregated measures associ-
ated with a periodic sample enhanced with the center line and the upper
and lower control limits.
These are the steps to building a control chart for measuring data
quality.
1. Select one or more data quality dimensions to be charted. Use the
Pareto analysis we discussed in Section 6.3 to determine the variables or
attributes that most closely represent the measured problem, since trying
to track down the most grievous offenders is a good place to start.
2. If the goal is to find the source of particular problems, make sure
to determine what the right variables are for charting. For example, if
the dimension being charted is timeliness, consider making the charted
variable the "number of minutes late," instead of "time arrived." When
trying to determine variables, keep in mind that the result of charting
should help find the source and diagnosis of any problems.
3. Determine the proper location within the information chain to
attach the measurement probe. This choice should reflect the following
characteristics.
a. It should be early enough in the information processing
chain that detection and correction of a problem at that point
can prevent incorrectness further along the data flow.
b. It should be in a location in the information chain that is

easily accessed and retooled, so as not to cause too much chaos
in implementing the charting process.
c. It should not be in a place such that observation of the
sample can modify the data being observed.
4. Decide which kind of control chart is to be used.
a. A variables chart measures individual measurable charac-
teristics. A variables chart will provide a lot of information
about each item being produced.
b. An attributes chart measures the percentage or number of
items that vary from the expected. An attributes chart provides
summary information about the entire process, focusing on
cumulative effects rather than individual effects.
5. Choose a center line and control limits for the chart. The center
line can either be the average of past measurements, the average of data
that has not yet been measured or collected, or a predefined expected
standard. The upper control limit (UCL) is set at 3 standard deviations
(+ 3a) above the center line, and the lower control limit (LCL) is set at 3
standard deviations (- 3a) below the center line.
6. Choose the sample. The sample may consist of measuring indi-
vidual data values or measuring a collection of data values for the pur-
pose of summarization. It is important that the sample be taken at a
point in the process or a point in time where it has a significant effect.
7. Choose a method for collecting and logging the sample data.
This can range from asking people to read gauges and recording the
answers to having an integrated mechanism for measuring and logging
sample results.
8. Plot the chart and calculate the center line and control limits
based on history.
6.5 KINDS OF CONTROL CHARTS
There are many different varieties of control charts.^ Since our goal is to
measure nonconformance with data quality expectation, we will con-
centrate on particular control chart attributes for measuring noncon-
1. For a detailed list of different control chart types, see Juran's Quality Hand-
book, 5th edition, edited by Joseph M. Juran and A. Blanton Godfrey (New York:
McGraw-Hill. 1999).
STATISTICAL PRCXiESS CONTROL AND IMPROVEMENT CYCLE 133
formity. Our statement of the data quality requirements will be using

the rules and assertions system we develop in Chapters 7 and 8. Our
sample measurements will be based on defining the granularity of the
data item being observed (record vs data attribute), defining a set of
data quality rules, and then testing the data items against those rules.
Each sample will consists of a number of measured items.
6.5.1 Percentage Nonconforming
Thefirsttype is a "Control Chart for Percentage Nonconforming" and

is also known as a p chart. A p chart is an attributes chart whose data
points represent a percentage of the data items that do not conform to
our requirements.
The distribution of this data set is a binomial distribution if we
assume that the process is constant. Most data points should fall within
3 standard deviations of the mean. For binomial variables, the standard
deviation is computed as
Gp
^Jpg-p)
where p is the probability of occurrence, and n is the sample size.
To set up a /? chart, a small sample size is collected over a short
period of time (in most cases, 25 to 30 time points will be enough) and
the average P is computed by counting the number of nonconforming
items in each sample, totalling the number of items in each sample
group, and dividing the total number of nonconforming items by the
total number of sampled items. For p charts, the control limits are cal-
culated using the binomial variable standard deviation; the UCL is com-
puted as P + 3ap, and the LCL is computed as P - 3ap. If the LCL is
computed to be a negative number, we just use 0 as the LCL.
6.5.2 Number Nonconforming
In this chart, instead of plotting the percentage of nonconforming data

objects, we will plot the number of nonconforming items. In the p chart,
the percentage p is equal to the number of nonconforming items divided
by the number of observed items, n. Clearly, the number of nonconform-
ing items is equal to np, and therefore, this chart is called an np chart.
134 ENTERPRISE KN0WLECX3E MANAGEMENT
For an np chart, the UCL is computed as P + 3 VwP(l-P), and the

LCL is computed as wP - 3 VP(1---P) where P is the average of the
number of nonconforming items, and n is the size of the sample.
6.5.3 Number of Nonconformities
It is not out of the realm of possibility that each data item being
observed may have more than one error! In this case, we may not just
want to chart the number of nonconforming data items but the total of
all nonconformities. This kind of attributes chart is called a c chart, and
the UCL is calculated as C + 3 VC. The LCL is calculated as C - 3 VQ
where C is the average number of nonconformities over all the samples.
6.5.4 Number of Nonconformities per Item
If our samples consist of multiple observed data errors, then we might

want to look at not only the number of nonconformities but the num-
ber of nonconformities per item. This chart is called a u chart, and the
UCL is computed as 17 + 3 V[7/. The LCL is computed as U - 3 Û/n,
where U is the average number of nonconformities, and n is the number
of items.
6.5.5 Defining the Control Limits
Thus far, we have discussed the calculations of the upper and lower con-
trol limits as a function of the statistical distribution of points in the data
set. This is not to say that we can only define these limits statistically.
In reality, as quality overseers, it is our duty to specify the accept-
able limits for data quality. For example, when it comes to the acceptable
level of incorrect values in certain kinds of databases, we can specify that
there is no tolerance for error. In this case, the UCL for errors would be
0. In many cases of examining data quality, there is no need for a lower
control limit either. Ultimately, it is up to the users to determine their tol-
erance for expected variations and errors and use that as a guideline for
setting the control limits.
6.6 EXAMPLE: INVALID RECORDS
In this example, each day a number of records are passed through an

automated data validation system, where each record is compared
against a number of data validation rules. If the record fails any of the
rules, it is tagged as an invalid record and a count is taken. This process
was repeated for a period of 24 days, yielding the following table.
Number of Number of
Day Records Processed Bad Records Bad Ratio
1 10,000 300 0.03

2 10,000 600 0.06
3 10,000 532 0.0532
4 10,000 476 0.0476
5 10,000 620 0.062
6 10,000 546 0.0546
7 10,000 665 0.0665
8 10,000 331 0.0331
9 10,000 337 0.0337
10 10,000 328 0.0328
11 10,000 345 0.0345
12 10,000 358 0.0358
13 10,000 403 0.0403
14 10,000 341 0.0341
15 10,000 347 0.0347
16 10,000 395 0.0395
17 10,000 342 0.0342
18 10,000 334 0.0334
19 10,000 346 0.0346
20 10,000 347 0.0347
21 10,000 378 0.0378
22 10,000 365 0.0365
23 10,000 351 0.0351
24 10,000 432 0.0432
Over the period of time, the overall average ratio of bad records
was computed to be 0.0409, which we use as the center line. The UCL
and LCL were computed in accordance with the computation for the p
chart, to be 0.0469 and 0.0349, respectively. The corresponding control

chart is shown in Figure 6.3.
Since we are trying to limit the error percentage to below a certain
point, we can essentially ignore the lower control limit; the fewer
errors, the better. As we can see, early on in the history, the process was
not in control because there were a number of days (days 2 through 8)
in which the upper control limit for errors was exceeded. At that point,
the number of errors each day begins to move into a more predictable
pattern, even sometimes moving below the lower control limit. Near the
end of the measurement period, the process displays errors well within
the acceptable limits.
0.07
0.06
0.05
0.04
0.03
0.02
0.01
FIGURE 6.3 Control chart for invalid records

6.7 THE GOAL OF STATISTICAL PROCESS CONTROL
Statistical process control is a tool that makes use of measurements of

certain aspects of quality of a process or a product over time to gain
insight into the differences between expected, common variations and
unexpected, special variations. The use of the control chart to represent
the behavior of a process over time is not just used as a means for locat-
ing anomalous events; the ultimate goals of SPC are stability and pre-
dictability.
Let's look at stability first. By definition, when there are a large
number of data points in the control chart outside the control limits, it
means that the process is very unstable. This instability is more than
just points plotted against a handful of parallel lines on a graph it
reflects different causes working at cross-purposes to affect the way the
system acts. It is the role of the quality specialist to analyze the results of
the SPC process to determine whether the instability is due to common
or special causes.
As we identify the special causes associated with each set of out-of-
control data points, we gradually improve the process, making it more
stable. This will be reflected in the control graph moving forward
because as the data points move closer to the mean, they will also begin
to fall within the control limits. This exhibits the stabilizing effect that
SPC can have on a process.
Another effect is that as the points fall closer together, the standard
deviation becomes smaller as well, and since the control limits are
defined as a function of the standard deviation, they will also begin to
move closer to the center line. This is a different kind of stability, one
that focuses on a tightening band within which we expect to see defec-
tive information the narrower the band between the UCL and the
LCL, the fewer expected errors! This kind of stability of a system also
implies predictability: If for the past 30 days, there were fewer than 10
errors every day, we can expect that tomorrow there will also be fewer
than 10 errors.
This notion of predictability is a significant benefit of SPC. By
using the SPC tools to understand the nature of problems within a sys-
tem and the selected variables measured to help locate the source of the
problem and by eliminating the problems, we gradually bring a system
where we can predict day-to-day behavior and confirm that anomalous
activity is due to special causes.
6.8 INTERPRETING A CONTROL CHART
Our next step in the SPC process is interpreting a control chart. Now
that we have collected and plotted the data points, how can we make
sense out of the resulting control chart? When a process is stable, we
can expect that all the points in the control chart will reflect a natural
pattern. The data points on the chart should be randomly distributed
above and below the center line, and the chart should have these char-
acteristics.
Most of the points are close to the center line.
Some of the points are near the UCL and LCL.
There may be some points above the UCL or below the LCL.
The distribution of points on the chart should not have any non-
random clustering or trending.
In the interpretation of control charts, our goal is to determine
whether a process is stable, and if it is not stable, to find and eliminate
special causes. So, what do we look for in a control chart?
6.8.1 Unnatural Patterns
Thefirstthing to look for is any departure from what we expect to see.

Any apparent patterns that belie the expected randomness in the chart
should be a signal for further investigation. Here are some examples.
Many points that lie outside of control limits. This clearly indi-
cates that the system is out of control. Note that when the control
limits are user-defined, there is a much greater possibility of this
happening than if we rely on the equations prescribed for calculat-
ing the control limits.
Unnatural clusters of points. Clusters most likely represent pat-
terns in which special causes lurk.
Shifts in levels seen in the control chart. In other words, is there a
sequence of points within one standard deviation, followed by a
sequence of points between one and two standard deviations?
Any trends up or down probably indicate some deterioration in
quality.
6.8.2 Zone Tests
Another aspect of unnaturalness in the distribution of data points can be

uncovered using what is called a zone test. A zone is an area of the chart
where there are unlikely distributions of data points on the chart, such as
2 or 3 successive points outside 2 standard deviations, 4 or 5 successive
points outside 1 standard deviation, or 8 successive points on the same
side of the center line. All these occurrences are equally likely to occur
when the process is stable, and if any appear in a way that is not consis-
tent with our expectations, this is an indicator of a special cause.
6.8.3 Rebalancing
After the root cause of a problem has been identified and corrected, we
can claim that at least one aspect of an out-of-control situation has been
resolved. In this case, it may be interesting to recalculate the points and
control limits on the control chart, ignoring the data points associated
with the identified cause. This should help strengthen the control limit
calculations and point out other locations to explore for special causes.
6.8.4 Refactoring the Data
Let's say we collected several days' worth of data on the number of

errors that occurred in the data each day as a function of the total num-
ber of records with errors in them. By aggregating the errors by record
instead of by error we may have obscured the fact that many records
failed more than one data validation test. As a result, the appearance of
a special cause may have been overlooked.
Instead, we can rechart the data points by collecting the data as the
number of records that failed due to a specific validation test failing. We
might find out that a large number of records are erroneous due to fail-
ing more than one test or that several of the erroneous records have
failed one test during the first half of the measurement period and failed
a different test during the second half.
By separating data points by attribute or remeasuring based on a
finer granularity, we may be able to identify occurrences of variations
due to special causes that the standard charting method fails to high-
light.
6.9 FINDING SPECIAL CAUSES
We have collected our data, built our control chart, plotted data points
and control limits, and analyzed the chart for anomalous behavior. The
last step in the SPC process is to identify the special causes that are
echoed in the control chart.
Hopefully, we will have selected the areas of measurement in a way
that will point us in the right direction. While we discuss this process in
greater detail in Chapter 15, we can briefly introduce it here.
Assuming that we have translated our data quality expectations
into a set of data quality rules, we can use those rules for validating data
records. If we log the number of times a record is erroneous due to fail-
ing a particular test, we can use those logs to plot the daily conformance
for each specific rule.
At the end of the measurement period, we can construct a control
chart consolidating data from each of the data quality rules. Because
each rule describes a specific aspect of the users' data quality require-
ments, the problem of identifying a special cause reduces to determining
which of the data quality rules accounted for the anomalous behavior.
This provides a starting point for the root cause analysis process
described in Chapter 15.
6.10 MAINTAINING CONTROL
In Section 6.7, we discussed the goals of statistical process control as

being stability and predictability. Once a process has been brought
under control, it is beneficial to continue making use of the SPC process
to make sure that the process remains under control.
As long as the data points continue to fall between the control lim-
its, the process is stable. Attempts can be made to improve the process
on a continuous process, either by making the control limits closer or
by introducing new variables or attributes to be measured.
8.11 SUMMARY
In this chapter we discussed the use of an analysis tool called Statistical

Process Control (SPC) as a method for measuring and charting the con-
formance of information to a set of data quality rules. The SPC method-
STATISTICAL PRCXIESS CONTROL AND IMPROVEMENT CYCLE 141
ology is based on the analysis of variation in a system and that some

variations are due to chance or common causes, whereas others are due
to special causes. It is those due to special causes that can be highlighted
by SPC, since the occurrence of variations due to common causes form
a normal distribution.
Because of this, we can build a chart called a control chart, which
integrates data taken from a sample with control limits (both upper and
lower). When the data points appear above the upper or below the
lower control limit, it is an indication of a special cause. Other unnat-
ural patterns also indicate the appearance of variations due to a special
cause.
We looked at different kinds of control charts that can be used to
analyze conformance to a set of data quality rules. We also looked at
the issues involved in interpreting a control chart. Finally, we intro-
duced the notion of root-cause analysis, which is discussed in greater
length in Chapter 15, and maintaining control, which consists of con-
tinuing to collect and chart data for the control charts.
7
DOMAINS, MAPPINGS, AND
ENTERPRISE REFERENCE DATA
Because a data type is assigned to each data attribute in a database table,

it draws its values from a specific value set. The same goes for informa-
tion embedded in a transmitted message: The value in each field in the
message should conform to the expected type for that field. In a way, it is
taken for granted that any value stored in a field is taken from a value
class (or set) that has some structural (or syntactic) rules as well as some
explicit connotative (or semantic) rules that govern the correctness or
validity of those values. Either way, these expectations actually boil
down into a set of restrictions on the values that the attribute may take.
The syntactic rules cover restrictions on the form of the value. A
field that has an integer data type cannot take a value that has alphabetic
characters. The semantic rules are additional restrictions on the set of
valid values for an attribute that are expressed as a subset of the allowed
structural values. When we can define an explicit set of restrictions on a
set of values within a type, we call that a domain.
A simple example is a U.S. Social Security number. A Social Secu-
rity number has a data type (character(ll)), but it also has both struc-
tural and semantic restrictions. The structural restriction is on the form:
three digits (0-9) followed by a hyphen (-), followed by two digits, a
hyphen, then four digits. The semantic restrictions specify rules about
the number itself. The first three digits denote the state (or area) where
the application for the number was filed. The next two digits are called
the group number, and they are issued in a particular order, namely odd
numbers from 01 through 09, followed by even numbers from 10
though 98. After those groups have been allocated come even numbers
from 02 through 08, followed by odd groups numbers from 11 through
143
99. Each month, the Social Security Administration (SSA) publishes

high group numbers for each area. The final four digits, called the serial
number, are allocated consecutively for each group. And, according to
the SSA, "Alleged Social Security numbers containing area numbers
other than those found on that table are impossible."
These rules define what is and is not a vaHd Social Security number.
A program can be written to determine whether a proffered Social Secu-
rity number is valid or not, but, as we will see, a framework can be
developed that will allow us to express the validation of a Social Secu-
rity number as a business rule that can be captured and documented
and from which a validation script can be automatically generated.
There are two major benefits to formalizing business and data
quality rules with respect to domains. Thefirstis knowing the means of
value restriction for afieldcan help in generating a test for validation.
Knowing an attribute's domain is useful for prescriptive validation of
data quality of data values. In other words, if we know what values the
field cannot have, we can make sure that any value that is inserted is
permissible. It is even better if we have a formal shorthand for describ-
ing those sets, if that formality can be turned into a validation test.
The second benefit is that accumulating metadata about the sets of
values that are used throughout the enterprise adds to collective knowl-
edge about how information is used and shared. By abstracting the sets
of values used and documenting their use and subscribers, we begin to
comprehend enterprise reference data, and then we can build a central-
ized enterprise reference data and metadata repository.
In this chapter, we explore the ideas revolving around data types
and how data types are related to the notion of sets. We then describe
our definition of data domains, both descriptive and enumerated. Next,
we discuss the relations between domains, how those relations exist in
databases, and the power of abstracting these mappings as reference
metadata. Finally, we propose a publish/subscribe model for the man-
agement and use of enterprise reference data.
7.1 DATA TYPES
We begin our exploration with a discussion of data types. What is a

data type, how is it used, and what are the mechanics of predefined data
types?
DOMAINS. MAPPINGS, AND ENTERPRISE REFERENCE DATA 145
7.1.1 Apples and Oranges
We are often reminded when we make questionable comparisons that we

should not compare apples and oranges. What does this really mean?
Usually, it implies that the basis for comparison of the two objects is
flawed and the comparison is basically meaningless. Formally, the classifi-
cation "apple" and "orange" are types, and it is meaningless to compare
objects of different types.
This is true in the data world as well, where each object is associ-
ated with a data type. In many computer programming languages,
intertype comparisons are illegal, whereas in others, there is implicit-
type conversion if it is possible. In the data world, depending on the sys-
tem, intertype comparisons will either be flagged as being illegal or
worse yet, ignored, yielding a garbage answer.
A data type provides a complete specification of the values that can
be stored in an attribute, such as the kind of value it holds, the rules for
allocating physical storage space for the value, and the operations that
may be applied to the values. For example, integers are whole numbers,
they take up limited space if they can be stored in a long word, and they
are subject to both arithmetic and conditional operators. Character
strings, on the other hand, consist of sequences of symbols, may require
explicit space allocations, and are not typically subject to arithmetic
operations.
It is useful to look at the possible data types that occur frequently
in database systems. In the next sections, we will build up an overview
of data types, starting with simple data types and building up to more
complex ones.
7.1.2 Base Types
The following are base data types.

Whole Number Types Whole number types are for storing inte-
gers. Integer types may be categorized within ranges, based on the
amount of storage required.
Character and String Types The character type is used to hold
alphanumeric and symbol characters. Character types are either
fixed length or variable length, and there are frequently options
for specifying national character sets (for example, Kanji).
Decimal Decimal types refer to exact numeric representation,

based on a precision, which is the total number of digits on both
sides of the decimal point, and a scale, which is the number of dig-
its to the right of the decimal point. This is sometimes referred to
as numeric.
Floating Point Floating point numbers are approximate repre-
sentations of real numbers. The precision of floating point num-
bers is dependent on representation and the system that is used.
Dates or Times A date or time (or datestamp or timestamp) is
used to store dates or times or combinations thereof.
Binary Types A binary object can hold data that may not con-
form to other predefined types, such as graphics, executable code,
or representations of data structure objects.
7.2 OPERATIONS
We will build on the notion of data types in a moment, but first, let's
look at the operations that are valid between data values within a data
type and between data values in different data types.
7.2.1 Arithmetic Operations
The operations that can be applied to the numeric data types (integers,
decimals, floating point, etc.) are the standard arithmetic operations.
Addition (+)
Subtraction (-)
Multiplication (*)
Division (/)
Modulo (%)
Note that division of integers may by definition yield an integer
result. Some systems may add it other numeric operations.
Floor (returns the largest integer less than a real)
Ceiling (returns the next integer greater than a real)
DOMAINS. MAPPINGS. AND ENTERPRISE REFERENCE DATA 147
7.2.2 Conditional Operations
These are the operators used for making comparisons.

less than (<)
less than or equal to (<=)
greater than (>)
greater than or equal to (>=)
equal to (==)
not equal to (!=)
7.2.3 Logical Operators
These are the operators used for forming logical expressions.

AND
OR
NOT
7.2.4 String Operations
These are operations that can be performed on character strings.

Concatenation (composing two strings together)
Ipad, rpad (pad strings with blanks)
Itrim, rtrim (removes a given substring from the left or right of a
string)
Lower, upper (conversion to lower case or upper case)
Length (returns the length of the string)
Substring (returns a substring of a given string starting at a spe-
cific location for a specific number of characters)
7.2.5 Aggregate Operations
Aggregate functions are those that can be used to summarize informa-

tion across a set of values. These include the following.
SUM, which sums up the values in a set
AVG, which computes the average of the values in a set
148 ENTERPRISE KN0WLECX3E MANAGEMENT
COUNT, which gives the number of values (there is also a DIS-

TINCT classifier for this operator)
MAX, which returns the highest value
MIN, which returns the lowest value
7.2.6 Date and Time Operations
Dates are usually stored in a special format, but the conditional opera-
tors listed should work on dates. In addition, there are special versions
of some of the arithmetic operators.
Addition (add a number of days to a date to get a new date)
Subtraction (subtract a number of days from a date to get a new date)
7.2.7 Conversion Operations
When dealing with strictly typed systems, operations between values

with different types are not allowed. In the real world, though, we fre-
quently have occasion to try to manipulate data values with different
types. In order to accommodate this, there are conversion operators
that change the data type (and possibly representation) of a data value.
Sometimes these conversions are implicit (such as multiplying a
real value by an integer value to yield a real value the integer is
implicitly converted). Sometimes these conversions are explicit, such as
transforming a string representation of a date to a date representation.
7.3 DOMAINS
In this section, we look at the ways to collect data values that can take
on intuitive meanings. When it becomes clear that a single collection of
values is used for the same meaning throughout different data reposito-
ries in the enterprise, a special status should be assigned to that collec-
tion as a data domain that can be shared by the users in the enterprise.
7.3.1 Data Types and Sets
A set is a collection of items that can be perceived as a whole, where

each item in the collection is distinguishable from all the others, and
where it can be clearly determined whether an item is a member of the

collection. A data type is a way of classifying data values into sets: the
set of whole numbers, the set of character strings of length 5, or the set
of decimal numbers that have 10 significant digits, with 3 of them fol-
lowing the decimal point.
Data types differ from sets in that each data type has a set of opera-
tions on values that have that data type. This concept is explicit in object-
oriented programming languages, where classes, which are really abstract
data types coupled with intrinsically defined operations, are defined.
Alternatively, there are certain operations that can be applied to
sets. Set union represents the combination of the values of two sets into
one larger set. Set intersection as applied to two sets represents those
values that reside in both sets. The difference between two sets repre-
sents those values that are in the first set but not in the second. There
are also set comparison operators. We can compare two sets to see if
one is a subset of the other, if two sets are equal to each other, and if two
sets are not equal to each other.
7.3.2 The Descriptive Power of Sets and Domains
We can think about sets in two different ways: (1) through enumera-
tion, listing all the elements in the set, and (2) through description. One
nice thing about sets is that all the values in the set can frequently be
described in a few short terms. A very simple example is the positive
whole number the integral values that are greater than 0. In that one
small phrase we can categorize an infinite number of values that intu-
itively belong to the same collection.
When you think about it, the use of databases is very much driven
by the idea of sets. All the records in a table in a database represent a
set. An SQL SELECT query into a database is a definition of a subset of
the records in the database.
In turn, when we look at the relationship between data attributes
and data types, we can see that many attributes draw their values from
a subset of the values allowable under the data type. For example, an
attribute called STATE may have been defined to be CHAR(2). While
there are 676 distinct two-character strings, the values populating that
attribute are limited to the 62 two-letter United States state and posses-
sion abbreviations. This subset restriction can be explicitly stated as
"All values in the attribute STATE must also belong to the set of recog-
nized USPS state and possession abbreviations."
We call a set of values that is described as a restriction or subset of

the values allowed within a base type a domain. There are two different
kinds of domains: enumerated domains, which are likely to draw focus
because of their relationship to preexisting business set definitions, and
descriptive domains, which are defined through constructive rules such
as "All three character strings where the first character is A and the
third character is a digit."
Because domains represent additional restrictions on the values
within a data type, they play an important role in data quality. Every
explicitly stated restriction can be automatically transformed into a vali-
dation test. All the domains discussed in this book have these properties.
A base type
A name with intuitive meaning
An enumeration list
A description rule
7.3.3 Enumerated Domains
An enumerated domain is useful when there are a finite number of ele-

ments in the set or the set members cannot be described easily. An enu-
merated domain is a list of values all associated with a particular base
type. Domain values are unique within a domain.
To continue our example from the previous section, we can define
a domain entitled USPS State Abbreviations with a base type of
CHAR(2) and an enumeration list of {AL, AK, AS, AZ, AR, CA, CO,
CT, DE, DC, FM, FL, GA, GU, HI, ID, IL, IN, lA, KS, KY, LA, ME,
MH, MD, MA, MI, MN, MS, MO, MT, NE, NV, NH, NJ, NM, NY,
NC, ND, MP, OH, OK, OR, PW, PA, PR, RI, SC, SD, TN, TX, UT, VT,
VA, VI, WA, WV, WI, WY, AE, AP, AA}.
7.3.4 Enumerated Domains
It is worthwhile to look at a number of examples of enumerated domains

to appreciate how frequently they are used. We can even describe classes
of enumerated domains without explicitly listing all the values.
Geographic domains: comprehensive listings of countries, regions,
continents, states, provinces, cities, and so on
Code values
Colors
Employment categories and titles
National Holiday dates
Catalog items
Telephone area codes
Product suppliers
Currency Codes
7.3.5 Descriptive Domains
A descriptive domain is defined using a base type and a descriptive rule.

The rule specifies a restriction on the values within that data type, using
operations that are allowed on values within that data type. A descrip-
tive domain that is based on an integer data type can be defined using
the standard integer arithmetic and conditional operations.
A descriptive domain can be used to describe data values that have
an explicit format. Descriptive domains can also refer to other domains.
Descriptive domains are often (unknowingly) used when there is some
embedded information in the data value (see Figure 7.1).
7.3.6 Examples of Descriptive Domains
Here are some well-known examples of descriptive domains.

ZIP + 4 Code Has a base type of CHARACTER(IO) where
characters 0-4 must be numeric digits, the fifth character is a
hyphen, and the last four digits must be numeric digits.
Corporate E-mail Address An alphanumeric string concatenated
with an "at" sign (@) concatenated with an alphanumeric string
followed by the string ".com."
Male Name The composition of two VARCHAR strings, sepa-
rated by a blank, where the first is taken from the malefibrstnames
domain, and the second is taken from the last names domain.
Passing Grade An integer value that is greater than or equal to
65 and less than or equal to 100.
Valid Password A character string whose length is at least 4 and
not greater than 9 and has at least one character that is not alpha-
betic.
ZIP + 4 Code DDDDD-DDDD
Telephone Number 1-(DDD)DDD-DDDD
DDDD DDDD DDDD DDDD
and the digits, when alternately multiplied by 1 or 2, and

then truncated to the units position, when summed, is
divisible by 10. This is embodied in this psuedocode:
multiplier := 1; sum := 0;
for i := length(cardnumber) downto 1 do
begin
Credit Card Number char := cardnumber[i];
if digittype(char) then
begin
product := digitvalue(char) * multiplier;
sum := sum + (product div 10) + (product mod 10);
multiplier := 3 - multiplier; // l->2, 2->l
end
end;
if (sum mod 10) = 0 then ok
E-mail Address <Stringl >@<String2>.COM
Alphanumeric string, whose length is greater

ATM PIN
than or equal to 4 and less than or equal to 8
Bank Account
CC-DDDDDDDD
Number
FIGURE 7.1 Examples of some data domains
7.3.7 Operations on Domains
Our domains are restrictions on value sets, and so we allow the same
operations for domains that we allow for sets. Figure 7.2 shows how
domains are similar to sets.
Union The union of domain A and domain B yields a new
domain that contains all the unique values of domain A and
domain B.
Intersection The intersection of domain A and domain B yields
a new domain that contains the values that are in both domain A
and domain B.
Difference The difference of domain A and domain B yields a new

domain that contains those values in domain A but not in domain B.
Membership A value ^ is a member of domain A if â is a value
in domain A.
Subset A domain A is a subset of domain B if all the values in
domain A are in domain 6.
Equality A domain A is equal to domain B if all the values in
domain A are in domain B and all the values in domain B are in
domain A.
Inequality A domain A is not equal to domain B if there is a
value in domain A that is not in domain B or there is a value in
domain B that is not in domain A.
( Domain A ]
A union B a is a member of A
A intersect B B is a subset of A
A is equal to B if
(A intersect B) = A and
(A intersect B) = B
A is not equal to B if
Difference (A-B) (A-B) is not empty
FIGURE 7 . 2 Domains are like sets

7.3.8 Composed Domains
A composed domain is a domain of values constructed through the

composition of values selected from two or more domains.
7.3.9 Knowledge Management and Cataloging Domains
Thefirststeps toward enterprise knowledge management are conceptu-

ally small ones, but they can be some of the more important ones.
Identify those values that can be characterized as domains.
Extract them and store them in a central repository.
Manage them as explicit content.
Let's look at each one of these in turn. Identifying domains can be
done two ways. Thefirstis through expert interaction asking the users
if there are any sets of values that are commonly used. The second way is
through automatic means by exhaustively analyzing the values that reside
in enterprise data systems. We will explore these in Chapter 13.
Extracting domains and putting them into a central repository is
relatively straightforward as long as the mechanics for representing
domains and managing them is available. We will look at one way to do
this in greater detail in Chapter 17.
Managing domains as explicit content is easier said than done. The
reason is more political than technical: Convincing stakeholders to
release their authority over a "staked-out claim" will probably be the
biggest hurdle. An important factor is convincing senior management
to take an active role.
7.4 MAPPINGS
Now that we have domains, we can start to see how relationships

between domains become important. While relational database systems
take relations across sets of values as tuples in a database table, there
are more concrete relationships that appear frequently in an enterprise.
Recognizing both the existence and the business value of one of these
relationships is the next step in creating a knowledge-based enterprise.
7.4.1 What Is a Mapping?
The natural progression in discussing data values, data types, and

domains is looking at relationships between data values that have already
been assigned into different domains. Simply put, a mapping is a relation-
ship between two domains. More formally, a mapping between domain A
and domain B is a set of pairs of values {a, b] where ^ is a member of
domain A and bisz member of domain B such that there is an intuitive
understanding of the relationship between the values a and b.
A familiar example of a mapping is the relationship between ZIP
code and city. Every ZIP code belongs to a named area covered by a
small post office or postal zone (see Figure 7.3).
7.4.2 Characteristics of Mappings
As with domains, all mappings have some intuitive meaning, which

we can capture in its name. In addition, each mapping will have these
features.
A source domain
A target domain
An enumeration list
A mapping rule
FIGURE 7.3 Definition of a mapping

The domain of a mapping itself is the same as the target domain.

Mappings may be one-to-one, in which for a value in domain X,
there is one and only one value mapped to domain Y; one-to-many, in
which for a value in domain X, there may be one or more values
mapped in domain Y; many-to-one, in which there may be one or more
values in domain X mapped to one value in domain Y; and many-to-
many, in which there may be one or more values in domain X mapped
to one or more values in domain Y.
If you recall mathematical terms, a one-to-one mapping can also be
referred to as a function, and the other mappings can also be referred to
as relations. By identifying a relation as a mapping, and by classifying
the mapping and managing it as knowledge, we can clearly define vali-
dation rules for relationships between values in different domains. Fig-
ure 7.4 shows the four types of mapping.
7.4.3 Enumerated Mappings
A simple form of a mapping is the enumerated list of pairs. Per the defi-
nition in Section 7.4.1, a mapping between domain A and domain B is a
One-to-One Many-to-One
One-to-Many Many-to-Many
FIGURE 7 . 4 Different kinds of mappings

set of pairs of values [a, b] such that <a is a member of domain A and b is
a member of domain B. For an enumerated domain, the collection of [a,
b] pairs is listed explicitly. Note that all the a values must be validated
as belonging to domain A, and all the b values must be validated as
belonging to domain B.
In Section 7.3.3, we encountered the domain of USPS state abbre-
viations. Presuming that we have another domain called USPS State
Names, which has base data type VARCHAR(30), we can define a map-
ping from state abbreviations to state names, as shown in Figure 7.5.
(AA, Armed ForcesThe Americas) (MO, Missouri)

(AE, Armed ForcesEurope) (MS, Mississippi)
(AP, Armed ForcesPacific) (MT, Montana)
(AL, Alabama ) (NC, North Carolina)
(AK, Alaska ) (NE, Nebraska)
(AS, American Samoa) (NH, New Hampshire)
(AR, Arkansas) (NJ, New Jersey)
(AZ, Arizonia) (NM, New Mexico)
(CA, California) (NV, Nevada)
(CO, Colorado) (NY, New York)
(CT, Connecticut) (ND, North Dakota)
(DC, District of Columbia) (MP, Northern Mariana Islands)
(DE, Delaware) (OH, Ohio)
(FL, Florida) (OK, Oklahoma)
(FM, Federated states of Micronesia) (OR, Oregon)
(GA, Georgia) (PA, Pennsylvania)
(GU, Guam) (PR, Puerto Rico)
(HI, Hawaii) (PW, Palau)
(lA, Iowa) (RI, Rhode Island)
(ID, Idaho) (SC, South Carolina)
1 (IL, Illinois) (SD, South Dakota)
(IN, Indiana) (TN, Tennessee)
(KS, Kansas) (TX, Texas)
(KY, Kentucky) (UT, Utah)
(LA, Louisiana) (VA, Virginia) !
(MA, Massachusetts) (VI, Virgin Islands)
j (MD, Maryland) (VT, Vermont) '
(ME, Maine) (WA, Washington)
(MH, Marshall Islands) (WI, Wisconsin)
(MI, Michigan) (WV, West Virginia )
(MN, Minnesota) (WY, Wyoming)
FIGURE 7.5
1 58 ENTERPRISE KNOWLEDGE MANAGEMENT
7.4.4 Examples of Enumerated Mappings
Here are some examples of enumerated mappings.

Geographic codes to geographic regions
Part number to part description
Radio station call letters to radio frequency
Job title to maximum salary
Currency code to country
7.4.5 Descriptive Mappings
A descriptive mapping is a mapping that is defined using a rule, similar

to the way that descriptive domains use rules. A descriptive mapping is
specified with a source domain, a target domain, and a rule for deriving
the mapped value from the source value. Because the rule specifically
maps a single value to a single value, a descriptive mapping must be a
one-to-one mapping. Descriptive mappings are used when there is a
simple way to describe the relationship.
7.4.6 Examples of Descriptive Mappings
These are some frequently encountered descriptive domains.

Numeric scores to grades
Currency to Currency Conversion
Telephone call duration to charge
7.4.7 Composed Mappings
A composed mapping is a mapping that incorporates the use of a com-

posed domain.
7.4.8 Operations on Mappings
The following are types of operations on mappings.

Membership A value pair [a, b) is a member of mapping M(Ay
B)iiaisa value in domain A andfcis a value in domain B.
DOMAINS, MAPPINGS. AND ENTERPRISE REFERENCE DATA 1 59
Subset A mapping M(A, B) is a subset of mapping N(C,D) if all

the value pairs in mapping M are in mapping N.
Equality A mapping M(Ay B) is equal to mapping N(Q D) if all
the value pairs in mapping M are in mapping N and all the value
pairs in mapping N are in mapping M.
Inequality A mapping M(A, B) is not equal to mapping N(Q D)
if there is a value pair in mapping M that is not in mapping N or
there is a value pair in mapping N that is not in mapping M.
7.5 EXAMPLE: SOCIAL SECURITY NUMBERS
To round out the examples, let's go back to the example discussed in the
beginning of the chapter the U.S. Social Security number. Our earlier
discussion focused on the structure of a valid Social Security number.
Now we'll look at constructing a domain definition for valid Social
Security numbers.
Our first step is to note that the Social Security number is com-
posed of five parts: the area denotation, a hyphen, the group number, a
hyphen, and the serial number. We can start our domain definition by
describing a Social Security number as a CHARACTER(ll) string that
is a composition of thesefivedomains, ordered by simplicity.
7.5.1 Hyphen
Simply enough, the hyphen domain is a one-character domain consist-

ing solely of the "-" character.
7.5.2 Geographical Code
The issued geographic code^ domain GC has a base type of CHARAC-

TER(3), and based on information collected from the Social Security
Administration, is defined using this expression.^
1. The term geographical code is misleading, since according to the SSA, it is not
meant to be any kind of usable geographical information. It is a relic of the precomputer
filing and indexing system used as a bookkeeping device.
2. There are other number ranges that are noted as being "new areas allocated but
not yet issued." We are using this to mean that for testing to see if a Social Security num-
ber is valid, it must have one of the issued geographical codes.
(CGC >= ' 0 0 1 ' ) AND CGC <= ' 5 8 6 ' ) ) OR

CCGC >= ' 7 0 0 ' ) AND (GC <= ' 7 2 8 ' ) )
These ranges are based on the information in Figure 7.6, which is

from the SSA Web site, indicating the mapping between geographic
codes and U.S. location.
7.5.3 Group Number
The group number domain is actually more complicated, since the

number is restricted based on both the order of issuance and the
monthly report of high group numbers by geographic code, which can
be accessed via the Internet at http://www.ssa.gov/foia/highgroup.htm.
The monthly issuance report is actually a mapping from geo-
graphic codes to high numbers. To properly create the rule for the
group number, we must introduce a new operation on a CHARAC-
TER(2) strings of digits X.
ODD(X) i s TRUE i f CONVERT(X, INTEGER) MOD 2 = 1
ODD(X) i s FALSE i f CONVERT(X, INTEGER) MOD 2 = 0
The defining rule for group number (GN) is conditional on the

value mapped by the monthly issuance map HGN for the geographic
code (GC) and reflects the convoluted logic described earlier in this
chapter.
CODD(HGNCGC)) AND (HGNCGC) < '10') AND (ODDCGN) AND GN <
10'))
OR
(NOT ODDCHGN(GC)) AND (HGNCGC) >= '1') AND ((ODD(GN) AND
GN < '10') OR (NOT ODDCGN) AND GN < HGNCGC)))
OR
CNOT ODDCHGNCGC)) AND CHGNCGC) < '10') AND CCGN < '10')
OR CNOT ODDCGN))
OR
CODDCHGNCGC)) AND CHGNCGC) >= '10') AND CGN < HGNCGC))
7.5.4 Serial Number
The serial number domain is a CHARACTER(4) string where each

character must be a digit between 0 and 9, and the number must be
between *000r and'9999'.
DOMAINS, MAPPINGS. AND ENTERPRISE REFERENCE DATA 161
7.5.5 Putting It Together
We canfinallydescribe the domain of valid Social Security numbers as a

composite domain. Social Security Number (SSN) has the base type
CHARACTER(ll), where SUBSTR(SSN, 0, 3) belongs to the geo-
graphic code domain, SUBSTR(SSN, 4, 1) belongs to the hyphen
domain, SUBSTR(SSN, 5, 2) belongs to the group number domain,
SUBSTR(SSN, 7, 1) belongs to the hyphen domain, and SUBSTR(SSN,
8, 4) belongs to the serial number domain.
7.6 DOMAINS, MAPPINGS, AND METADATA
We have concentrated on a formal definition of domains and mappings

because they are the basis for consolidating enterprise knowledge and
experience in one knowledge repository. Once we begin to identify
known collections of data and assign meanings to them, we begin to get
a better understanding of what information is being used in the organi-
zation, who is using it, and how that information is being used.
7.6.1 Domains and Mappings as Reference Data
We differentiate between different flavors of information based on their

derivation. Transactional data is generated from standard online trans-
action processing systems (OLTPs). Reference data is information used
for reference, no matter where it came from. Reference data sets contain
frequently used information, but the information is moderately static. It
is updated infrequently compared to OLTP systems.
Domains and mappings are types of reference data. While the val-
ues that make up the domains and the mappings may have actually
been derived from transactional data, once the data sets are categorized
as domains and the relationships as mappings, and especially when we
ascribe some meaning to the collections, those sets can be moved into
the reference data arena.
As data sets cross the border into enterprise reference data, there
will be a convergence of data users of the reference data. It is this
process that leads to an increase in the three dimensions of data domain
data quality: agreement of usage, stewardship, and ubiquity, discussed
in Section 7.7.
001-003 ...New Hampshire 440-448 ...Oklahoma High Group

449-467 ...Texas Number Table
004-007 ...Maine
008-009 ...Vermont 627-645
010-034 ...Massachusetts 468-477 ...Minnesota
478-485 ...Iowa (downloaded from
035-039 ...Rhode Island
486-500 ...Missouri Social Security
040-049 ...Connecticut
501-502 ...North Dakota Agency)
050-134 ...New York
135-158 ...New Jersey 503-504 ...South Dakota
159-211 ...Pennsylvania 505-508 ...Nebraska
212-220 ...Maryland 509-515 ...Kansas
221-222 ...Delaware 516-517 ...Montana
223-231 ...Virginia 518-519 ...Idaho
691-699* 520 ...Wyoming
232-236 ...West Virginia 521-524 ...Colorado
232 ...North CaroHna 650-653*
237-246 525,585 ...New Mexico
681-690* 648-649
247-251 ...South Carolina 526-527 ...Arizona
654-658 600-601
252-260 ...Georgia 528-529 ...Utah
646-647
261-267 ...Florida 530 ...Nevada
589-595 680
268-302 ...Ohio 531-539 ...Washington
303-317. .Indiana 540-544 ...Oregon
318-361. ..Illinois 545-573 ...California
362-386. ..Michigan 602-626
387-399. ..Wisconsin 574 ...Alaska
400-407. ..Kentucky 575-576 ...Hawaii
408-415. ..Tennessee 750-751*
lS(>-l(>y 577-579 ...District of Columbia
416-424. ..Alabama 580 ...Virgin Islands
425-428. ..Mississippi 580-584 ...Puerto Rico
587 596-599
588* 586 ...Guam
752-755* 586 ...American Samoa
429-432. ..Arkansas 586 ...Philippine Islands
676-679* 700-728 ...Railroad Board**
433-439. ..Louisiana
(>S^-(>(^y
FIGURE 7.8 Social Security Number

DOMAINS, MAPPINGS. AND ENTERPRISE REFERENCE DATA 1 63
7.6.2 Domains and Mappings as Metadata
Another way of looking at domains and mappings is that they belong to

a metacoUection of information: the class of data sets that have some
meaning to at least one set of users. The fact that a data domain or a
mapping between domains has been identified carries more information
than the knowledge about the data set or sets alone. Rather, the knowl-
edge of the existence of a new domain should be available to anyone in
the enterprise (modulo any usage or privacy restrictions), as well as the
ability to browse the domain or mapping.
This is an example of metadata that adds to the enterprise knowl-
edge base. We can manage a repository of domain names and meanings
that can be accessed from across the enterprise. In the next section, we
will see how this metadata can be used.
7.7 THE PUBLISH/SUBSCRIBE MODEL OF REFERENCE

DATA PROVISION
The natural conclusion of our discussion on the identification and isola-

tion of domains and mappings is putting into place a framework that
allows that information to be shared across the enterprise. Before this
can be done successfully, these four pillars must be in place.
1. Senior management must support a centralized reference data
system complete with assigned stewardship responsibilities.
2. There must be a usage agreement between the users and the cen-
tralized reference managers.
3. There must be a means for the efficient publication and distribu-
tion of reference data.
4. There must be a means for users to subscribe to published infor-
mation.
7.7.1 Reference Data and Stewardship
If we know that the organization has overcome any political issues

regarding data ownership (as discussed in Chapter 2), most Ukely senior
management supports a knowledge-based enterprise. In this case, it is
imperative that a data steward be selected and empowered to oversee
the management of the domain set. These are the responsibilities of a

data steward for domains and mappings.
1. Authorization of Access and Validation of Security This
involves authorizing users of domains and mappings, validating
security for access to the domains and mappings, authorizing
modifications to the reference data, and providing access to the
metadata for the purposes of browsing.
2. Support the User Community If there are any issues regarding
the use or availability of reference data, the steward must resolve
them.
3. Maintenance of Data This involves scheduling any periodic
updates, making sure that the resources required for data provi-
sion are available and working, and acquiring any technology
required to maintain the data. Any issues of data aging or reten-
tion are handled within the scope of maintenance.
4. Responsibility for Data Quality This includes defining the data
quality rules associated with the domains and mappings and
making use of any technology required to assess and maintain a
high level of data quality within each domain or mapping.
5. Distribution of Information This involves enabling the capabil-
ity to disseminate reference data in an efficient manner, which
may include replication strategies or data delivery systems using
message integration software.
6. Management of Business Rules This incorporates the documen-
tation of all domain and mapping metadata, the documentation
of all associated business rules, and identification and documen-
tation of the way the reference data is being used and by which
users.
7. Management of Metadata This involves developing and
approving business naming standards, data definitions, aliases,
and the documentation of the information about the data
domains and mappings.
8. Source Management This involves managing the sources of ref-
erence information, whether it means internal data providers or
external data providers. This integrates with the preceding data
quality responsibility.
7.7.2 Agreement of Usage
The agreement of usage is essentially a contract between the users of a

data domain (or mapping) and the responsible data stewards. It speci-
fies what reference data sets are to be used, which users are authorized
to use them, and a service and support arrangement. In addition, a level
of service is negotiated based on the users requirements and priorities.
The agreement of usage is the method for the data stewards, as rep-
resentatives of the knowledge-based enterprise, to measure and catego-
rize the value of stored reference data. Usage agreements are also
managed as metadata.
7.7.3 Publication of Reference Data
The pivotal point of the accumulation of reference data managed under

the watchful eye of a data steward is when the information itself can be
distributed to enterprise users. In order to benefit from the management
of reference data, there must be a scalable mechanism for publication of
reference data.
What do we mean by "publication".^ In most cases, it represents
the channel through which requested information is delivered. This can
be via standard database queries, specialized database views that are
materialized on demand, or network information agents that propagate
replicated copies of information through the enterprise.
The requirements for a publication system are predicated on these
conditions.
The system must supply information to all subscribers within a
prearranged level of service and timeliness.
The system must guarantee that the information is consistent
across the enterprise.
The delivery mechanism must be scalable as the number of sub-
scribers increases.
The delivery mechanism must be robust in the presence of differ-
ent kinds of system platforms and data architectures.
7.7.4 Subscription-Based Distribution
What do we mean by "subscribers"? In our system, a subscriber is any

user (or group of users) that has contracted an agreement of usage for a
particular set of reference information and has been authenticated to
access that information. While a subscriber is allowed access to the
information, the means for that access is defined in accordance with the
data publication framework described in Section 7.7.3.
How can these ideas be actualized? Since many organizations
migrate their information systems toward a multitiered architecture, the
notion of a "broker" is gaining in popularity. A broker is an application
agent that sits between the source of information and the destinations
that have requested information (see Figure 7.7). Brokers are a useful
paradigm because the existence of a mediator for data requests provides
a means for providing support at all negotiated levels of service, for
guaranteeing a level of consistency across the network, for providing
scalability as the system grows, and for providing seamless connectivity
in the presence of different hardware systems.
7.7.5 Privacy Using Anonymous Publish/Subscribe
Sometimes the knowledge of who is reading which information is also

subject to a degree of privacy. Because the broker acts as intermediary, it
is possible to configure an enterprise information distribution system
using a more refined concept known as anonymous publish/subscribe
that can provide a level of privacy. In this model, it is the broker that
completely manages the relationship between requestor and data
source. While the data stewards may know that a number of sub-
scribers exist, it is only really important for them to know about the
contracted levels of service for those subscribers. Alternatively, the sub-
scribers only care about getting the information but they do not care
where it comes from.
To this end, the data publisher can post information into a blind
data channel that is managed by one or more brokers. In turn, the bro-
ker can feed that information into blind data channels that are directed
at the subscribers. It is possible to configure this system so that the pub-
lishers are unaware of the subscribers and vice versa thus, the name
anonymous publish/subscribe.
Data
n
-cm
o
Data
Information Broker
manages subscription
and distribution
requests to
Data
clients

Data
FIGURE 7 . 7 The information broker
7.8 SUMMARY: DOMAINS. MAPPINGS. AND REFERENCE DATA
In this chapter, we looked at a more refined notion of data typing called

a domain, which represents restrictions on sets of data values within a
data type. Relations between domains are expressed as mappings.
Domains can either be represented as enumerations of values or rules.
Together, domains and mappings are a source of reference information,
and as more commonality in usage between domains and mappings is
discovered across the enterprise, the more relevant the used domains
become as enterprise reference data.
As domains and mappings are consolidated from different data sets,
they can be captured and managed not only as knowledge content but
also as metadata. Data stewards are assigned to manage the reference
data and are responsible for service, support, maintenance, data quality,
and source management, among other responsibilities.
When a catalog of domains and mappings is available for all users,
they can use the reference data to contract an agreement with the infor-
mation suppliers or data stewards. Information is disseminated using a
publish/subscribe mechanism, which can be implemented using inter-
mediary agents processes known as brokers. If there is any sensitivity in
reference to the users themselves or the specific reference data, an
anonymous publish/subscribe mechanism can be implemented.
8
DATA QUALITY ASSERTIONS
AND BUSINESS RULES
In this chapter, we explore the formal way to describe what we expect

to see in terms of data quality. Through the definition of data quality
rules, we can derive a means to both assess the current state of and mea-
sure the progress in improvement of the data's quality.
Remember this basic principle: For any rule, given a set of data
items and a data quality rule base, the data set can always be grouped
into (1) items that conform to the rule and (2) items that do not con-
form to the rule. We will see more about this principle in Chapter 13
when we look at the automated discovery of data quality rules and in
Chapter 17, where we explore the transformation of these rules into
implementable objects.
8.1 DATA QUALITY ASSERTIONS
Our discussion of data quality rules focuses on four kinds of assertions.

Definitions
Proscriptive assertions
Prescriptive assertions
Conditional assertions
Each class in its own way defines some aspect of how information
is extracted from data and how the business operations depend on the
quality of the data.
169
8.1.1 Definitions
Definitions describe metainformation about the data that is under

scrutiny. The set of definitions forms the basis for the remainder of the
rules. Definitions include the domain and mapping definitions, described
in Chapter 7, as well as null value definitions.
8.1.2 Proscriptive Assertions
Proscriptive assertions define what is not allowed to happen in a data

set. For example, there may be a rule that an attribute may not have a
null value. These are the kind of rules that can be used to determine
what is wrong with data. We use these rules for specifying and quantify-
ing the levels of data quality in a data set. In other words, we can use
these rules to quantify a level of data quality. Proscriptive assertions
may appear in either a positive or negative form either describing a
restriction on what is or is not allowed in a field.
8.1.3 Prescriptive Assertions
Prescriptive assertions define what is supposed to happen in a data set.

An example is a rule that indicates that attribute A's values are com-
posed of the sum of attribute B and attribute C. These are rules that can
represent application logic for data population as well as data extrac-
tion, transformation, and migration. These rules can also be used for
message passing and content-based data flow.
8.1.4 Conditional Assertions
Conditional assertions specify that an action must take place if a condi-

tion is true. These describe what should happen if conditions are met.
An example is if the value of attribute A is greater than 10, then
attribute B may be blank. Conditional assertions essentially add logic to
the rules' definitions.
DATA QUALITY ASSERTIONS AND BUSINESS RULES 171
8.2 DATA QUALITY ASSERTIONS AS BUSINESS RULES
Today, businesses are increasingly distinguished by their ability to accu-

mulate, process, and disseminate information. A goal for increasing
competitiveness is increased automation and the use of straight-through
processing, which depends on discrete data interchange definitions and
high-quality data.
Successful businesses will gain greater scalability and higher opera-
tional margins by achieving a high level of automated information pro-
cessing. The success factors can only be reached by building an
enterprise whose business rules dictate the guidelines for the validation
of information quality. Since high information throughput can be
achieved if nothing "gums up the works," maintaining high levels of
data quality affects the proper operation of the business. Therefore, any
rule that decreases the number of exceptions in an operation stream is a
business rule. In other words, in a knowledge-oriented business, busi-
ness rules are data quality rules.
8.3 THE NINE CLASSES OF DATA QUALITY RULES
Data quality rules can be separated into nine classes.

1. Null value rules
2. Value rules
3. Domain membership rules
4. Domain Mappings
5. Relation rules
6. Table, cross-table, and cross-message assertions
7. In-process directives
8. Operational directives
9. Other rules
8.4 NULL VALUE RULES
Null value rules specify whether a data field may or may not contain
null values. A null value is essentially the absence of a value, although
there are different kinds of null values. Consequently, we will work
with one method for defining and characterizing null values and two
kinds of null value rules. Thefirstasserts that null values are allowed to
be present in a field, and the second asserts that null values are not
allowed in a field.
8.4.1 Null Value Specification
Our goal for null value specifications is to isolate the difference between
a legitimate null value and a missing value. Since a data record may at
times allow certain fields to contain null values, we provide these deeper
characterizations.
1. No value There is no value for this field a true null.
2. Unavailable There is a value for this field, but for some reason
it has been omitted. Using the unavailable characterization
implies that at some point the value will be available and the field
should be completed.
3. Not applicable This indicates that in this instance, there is no
applicable value.
4. Not classified There is a value for this field, but it does not con-
form to a predefined set of domain values for that field.
5. Unknown The fact that there is a value is established, but that
value is not known.
Considering that we allow more than one kind of null value, we
also need to allow different actual representations, since there is usually
only one system-defined null value. Therefore, any null value specifica-
tion must include both the kind of null along with an optional assigned
representation. Here are some examples.
Use "U"* for unknown
Use "X" for unavailable
Use "N/A" for not applicable
Of course, in building our rule set, it is worthwhile to assign a han-
dle to any specific null value specifications. In this way other rules can
refer to null values by their handle, which increases readability.
Define X for unknown as "X''
Define GETDATE for unavailable as "fiill in date''
Define U for unknown as " ? "
This way, we can have different representations for the different
kinds of null values in a way that allows flexibility in defining the null
value rules. Figure 8.1 gives examples of each null value specification.
ATA QUALITY ASSERTIONS AND BUSINESS RULES 173
Type of Null Description Example
In a field for "mobile phone," if there is no

mobile phone number associated with the
No value There is no value for this field.
customer, this is left null, since there is no
value for it.
There is a value for this field, In a field for "mobile phone," there is a
Unavailable but for some reason it has been mobile phone number associated with the
omitted. customer, but it has not been filled in.
In a product orders when the buyer is a

In this instance, there is no
; Not applicable not-for-profit agency, their tax field is not
applicable value.
applicable and will be null.
There is a value for this field, In an order for a sweater, where the colors
but it does not conform to a are limited to red, blue, or black, if the buyer
Not classified
predefined set of domain values requested a color that was not available, the
for that field. field might be left blank.
In a field for "mobile phone," it is known that

The fact that there is a value is
there is a mobile phone number associated
Unknown established, but that value is not
with the customer; but the number has not
known.
yet been acquired.
FIGURE 8.1 Different kind of nulls
8.4.2 Null Value Rule
The null value rule is a prescriptive rule that specifies that null values
may be present in the indicated field. A null value rule will specify the
kinds of null values that may be present and a representation (if any)
used for those null values.
A null value rule may allow traditional null values (such as system
nulls, empty fields, or blanks), generic null values as defined in null
value specifications, or a detailed Ust of specific null value representa-
tions. Note that if we only allow certain kinds of null values, this will
most likely mean that we want to restrict the appearance of the tradi-
tional nulls!
Here are some examples.
Rule ANuUs:
Attribute A allowed nulls {GETDATE, U, X}
In this example, we only allow certain kinds of null values, as
described in Section 8.4.1.
RuleBNuUs:
Attribute B allowed nulls
With the null value rules, the resulting validation depends on the
types of null values allowed. If any nulls are allowed, then there is really
nothing to do whether or not a value is in the field, the field is con-
formant. But if only certain kinds of nulls are allowed, then the valida-
tion for that rule includes checking to make sure that if any other null
values appear (such as the system null or blanks), the record is marked
as violating that rule.
8.4.3 Non-Null value rules
A non-null value rule, as its name implies, is a proscriptive rule that

specifies which kinds of null values are not allowed. If the rule indicates
that no nulls are allowed, then the rule is violated if the system null or
blanks appear in the field. The non-null value rule can also specify that
certain predefined null representations are not allowed, in which case
any appearance of those defined null values constitutes a violation.
Here is an example.
Rule BnoNuUs:
Attribute B nulls not allowed
8.5 VALUE MANIPULATION OPERATORS AND FUNCTIONS
In order to further define rules, we need a battery of functions that can

express manipulations or conditions on values. Table 8.1 lists these
functions and their purposes.
8.6 VALUE RULES
Rules covering particular values that may be assigned to a designated

field are referred to as value rules. The classifying name for rules that
constrain field values are called Attribute Value Restriction rules, which
limit the set of valid values that can be assigned. Value rules are pre-
scriptive, but are used in the proscriptive sense when looking for viola-
tions.
ATA QUALITY ASSERTIONS AND BUSINESS RULES 175
TABLE 8.1
Manipulation Operators and Functions
Function Name Arguments Result Description
PLUS X, Y: arithmetic values X+Y Addition

MINUS X, Y: arithmetic values X-Y Subtraction
TIMES X, Y: arithmetic values X*Y Multiplication
OVER X, Y: arithmetic values X/Y Division
MODULO X, Y: arithmetic values Integer remainder after division Modulo
SUBSTR S: String, X, Y: integer The substring of S starting at Takes a substring
position X and continuing for of specific start and
Y characters length
INDEX SI, S2: Strings The position in SI Searches for

where S2 appears location of S2 in SI
LENGTH SI: String The length of SI
< X,Y: comparable values Returns TRUE if X Less than
is less than Y
<= X, Y: comparable values Returns TRUE if X is Less than or equal to
less than or equal to Y
> X, Y: comparable values Returns TRUE if X is Greater than
greater than Y
>= X, Y: comparable values Returns TRUE if X is greater Greater than or
than or equal to Y equal to
== X, Y: comparable values Returns TRUE if X is equal to Y Equals
1= X, Y: comparable values Returns TRUE if X is not Not Equals
equal to Y
AND LI, L2: two logical values TRUE if both values are Logical AND
TRUE; FALSE otherwise
OR LI, L2: two logical values TRUE if either value (or both) Logical OR
is TRUE; FALSE if both are
FALSE
NOT LI: a logical value TRUE if LI is FALSE: False Logical N O T
if LI is TRUE
8.6.1 Attribute Value Restriction Rules
All attributes are associated with some preassigned data type, such as
string or integer. While the data type provides a level of data validation,
it is limited in its extent, since there may be many values that are data
type conformant, yet are inappropriate for a particular field.
One example would be an attribute GRADE associated with a stu-
dent course database. The data type for the field would be CHAR(l) (a
character string of length 1), but we can presume that the only valid val-
ues for GRADE are {A, B, C, D, F}. An entry for this field of the letter J
would be completely valid for the data type. To further restrict the set of
valid values, we can add an attribute value restriction rule for that field.
Rule GradeRestrict:
Restrict GRADE: (value >= "A*^ AND value <= "F") AND
(value != "E*')
This rule contains all the restrictions necessary for the field; The
value is between A and F but cannot be E. All value restriction rules
indicate a narrowing of the set of values valid for the attribute and are
usually expressed using range operations combined with the logical
connectors "AND," "OR," and so on.
These kinds of rules exist in many places, and they are the ones
most frequently embedded in application code. Exposing these rules
can be very useful, especially if there is a possibility that there are con-
tradictions between the restrictions. If more than one rule is associated
with an attribute, both rules must hold true, the equivalent of creating a
new rule composed of both restrictions connected using the logical
AND operator. A contradiction exists when two rules exist that are
applied to the same field, yet it is not possible for both to always hold.
Here is an example of two contradicting rules.
Rule StartRestrictl:
Restrict STARTDATE: value > "June 21,1987" AND value <
"December 1,1990"
Rule StartRestrict2:
Restrict STARTDATE: value < "February 1,1986"
In thefirstrule, the date must fall within a certain range, while the sec-
ond rule says that the date must be earlier than a different date. Combining
the rules yields a restriction that can't be true: The date is before February
1,1986 but is also between June 21,1987 and December 1,1990.
These contradictions may not be discovered in standard use because
application code is executed in a sequential manner, and the time depen-
dence masks out what may be a harmful oversight. Once the rules are
enumerated, an automated process can check for contradictions.
8.7 DOMAIN MEMBERSHIP RULES
We described domains and mappings in Chapter 7. Now we examine the

rules affecting domains. As with the null rules, we have both definition
specifications as well as conformance rules. There are two kinds of defin-
ition rules, one prescriptive rule, and two proscriptive membership rules.
8.7.1 Syntactic (Descriptive] Domain Definition
In Chapter 7, we discussed the difference between descriptive domains

and enumerated domains. A descriptive domain uses a "formula" to
determine whether the value is in the domain. Our examples included
the U.S. Postal Service ZIP + 4 code, which can be described as having a
base type of CHARACTER(IO), where characters 0-4 must be numeric
digits, the fifth character is a hyphen, and the last four digits must be
numeric digits.
Here is that rule expressed more formally, using the variable val to
represent the value.
Define ZIP4 as: substr(vai, 0, 5) >= 00000 AND substr(val, 0, 5)
<= 99999 AND substr(val, 5,1) = "-" AND substr(val, 6,4)
>= 0000 AND substr(val, 6,4) <= 9999
We refer to these definitions as syntactic because there is a distinct
syntax that can be applied to each value to determine whether it
belongs to the domain.
8.7.2 Enumerated Domain Definition
In contrast, some domains are defined as a list of vaUd values. Our

examples in Chapter 7 included United States state names, national hol-
iday dates, or catalog items. Here is a formal representation of one of
these domains.
Define NationalHolidays as:
{"New Year's Day,"
"Martin Luther King Day,''
"President's Day,"
"Memorial Day,"
"Labor Day,'^
"Veteran's Day,"
"Thanksgiving,"
"Christmas"}
8.7.3 Domain Assignment
In some data environments, data values that appear in one field always
belong to the same domain, although it may not be clear a priori what
the complete set of values will be. For example, we may define a domain
of "Current Customers," but values are added to this domain dynami-
cally as new customers are added to the database. But once a value
appears in the field, it will always be considered as a domain member.
This becomes especially useful when other data consumers rely on the
values in that domain. Because of this, we have a rule that specifies that
allfieldvalues define a domain.
Domain assignment is a prescriptive rule. A domain assignment
rule, when applied to a datafield,specifies that all values that appear in
thatfield(or hst offields)are automatically assigned to be members of a
named domain. This domain is added to the set of domains and can be
used in other domain rules.
Our formal representation includes a domain name and a list of
database table field names whose values are propagated into the
domain. The domain is constructed from the union of all the values in
each of the table columns. Here is an example.
Define Domain CurrencyCodes from {Countries.Currency,
Orders.Currency}
In this example, the domain CurrencyCodes is created (and
updated) from the values that appear in both the Currency column in
the Countries table, and the Currency column in the Orders table. This
rule may have been configured assuming that the right currency codes
would be correctly set in the Countries table, but in reality, the curren-
cies that are used in customer orders are those that are actually used, so
we'll include the currencies listed in the Orders table.
8.7.4 Domain Membership
We now have three ways to define domains. We are therefore ready to

restrict field values to named domains. A domain membership rule is a
proscriptive rule that says all values that appear in a specified field must
belong to a named domain.
As an example, the addresses of our U.S. customers in our sales
database must have states that are drawn from the US_STATES
domain. Our formal rule must include the field name that belongs to a
domain and the domain to which it belongs:
Rule StatesMembership:
SALES.CUSTOMER_STATE Taken from Domain US_STATES
8.7.5 Domain Nonmembership
Conversely, there may be times that we want to make sure that no

attribute's value is taken from a specific domain. We can specify this
using the proscriptive domain nonmembership rule. For example, if we
have divided our sales database into regional tables, we may want to
make sure that no states in the customer addresses in our Eastern region
table belong to WESTERN.STATES:
Rule StatesSelect:
EASTERN_SALES.CUSTOMER^STATE NOT Taken from
Domain WESTERN_STATES
Formally, a domain nonmembership rule must include a field name
and a domain from which the field's values may not be taken (see Fig-
ure 8.2).
8.8 DOMAIN MAPPINGS AND RELATIONS

ON FINITE DEFINED DOMAINS
As we discussed in Chapter 7, domain mappings relate the values in a

source domain to values in a target domain. Like domains, we have five
kinds of mapping rules two are for definition, one is a prescriptive
rule, and two proscriptive rules.
All domain mapping definitions must specify the source domain
and the target domain. Base types are valid for domain names, although
Rule Type Description Rule Format
Syntactic Domain Definition of a rule specifying those

DEFINE <domam-name> AS <condition>
Definition values valid within the domain
Enumerated Domain Definition of a rule specifying those

DEFINE <domam-name> AS (<value-list>)
Definition values that belong to the domain
A rule that indicates that an attribute's

Domain Assignment DEFINE Domain <domam-name>FROM {<attribute-list>}
values define the domain
A rule that indicates that an attribute's

Domain Membership <attribute'name> TAKEN FROM <domam-name> i
values must be taken from the domain
Domain A rule that indicates that an attribute's

<attribute-name> NOT TAKEN FROM <domain-name>
Nonmembership values may not belong to the domain
FIGURE 8.2 Domain rules
true domains are preferred. A mapping can be used both as a validation

reference (do the two values belong to the mapping?) and as a func-
tional reference (what is the mapped result of value X?).
8.8.1 Functional Domain Mapping
One way to define a domain mapping is through the use of a functional

map between source values and target values. In this rule, a function is
used to describe how to map from the source value to a target value.
This functional mapping may incorporate some conditional logic, and
the function must be able to provide a result for every value in the
source domain.
An example from Chapter 7 is the mapping from numeric score to
alphabet grade. The functional mapping can be described like this.
Define Mapping ALPHAGRADE(val):INTEGER TO CHAR(l):
{
"A^' if val > 90 AND val <= 100;
"B" if val > 80 AND val <= 90;
"C^' if val > 70 AND val <= 80;
"D'' if val > 65 AND val <= 70;
"F'' otherwise
}
What would happen if the value being mapped is an integer but is

not less than or equal to 100? An invalid input to this functional spec
would provide an answer, but most likely not one that we expect. Of
course, we would probably have already defined two domains to help
us in this rule, one domain for valid numeric scores, and one for the
alphabetic grade:
Define Domain VALIDSCORE as: val >= 0 AND val <= 100
Define Domain GRADESfrom{"A", "B", " C , D", "F"}
The functional mapping becomes:
Define Mapping ALPHAGRADE(val): VALIDSCORE TO
GRADES:
{
"A" if val > 90 AND val <= 100;
"B" if val > 80 AND val <= 90;
"C if val > 70 AND val <= 80;
"D" if val > 65 AND val <= 70;
"F" otherwise
}
The benefit of basing the definition on predefined domains is that it
adds a level of encapsulation in which business logic can be automati-
cally embedded.
8.8.2 Domain Mapping Enumeration
A domain mapping can also be defined using an explicit enumeration of

mapping pairs. In Chapter 7, we discussed USPS state abbreviations.
Here we show the mapping between USPS state abbreviations to the
actual full state names.
Define Mapping STATE_ABBREV_TO_NAMES:STATEABBREVS
TO STATENAMES:
{{"AL", "ALABAMA"},
{"AK", "ALASKA"},
{"AS", "AMERICAN SAMOA"},
{"AZ", "ARIZONA"},
{"AR", "ARKANSAS"},
{"CA", "CALIFORNIA"},
{"CO", "COLORADO"},
{"CT", "CONNECnCUT"},
{"DE", "DELAWARE"},
{"DC", "DISTRICT OF COLUMBIA"},
{"FM", "FEDERATED STATES OF MICRONESIA"},
{"FL", "FLORIDA"},
{"GA", "GEORGIA"},
{"GU", "GUAM"},
{"HI", "HAWAH"},
{"ID", "IDAHO"},
{"IL", "ILLINOIS"},
{"IN", "INDIANA"},
{"lA", "IOWA"},
{"KS", "KANSAS"},
{"KY", "KENTUCKY"},
{"LA", "LOUISIANA"},
{"ME", "MAINE"},
{"MH", "MARSHALL ISLANDS"},
{"MD", "MARYLAND"},
{"MA", "MASSACHUSETTS"},
{"MI", "MICHIGAN"},
{"MN", "MINNESOTA"},
{"MS", "MISSISSIPPI"},
{"MO", "MISSOURI"},
{"MT", "MONTANA"},
{"NE", "NEBRASKA"},
{"NV", "NEVADA"},
{"NH", "NEW HAMPSHIRE"},
{"NJ", "NEW JERSEY"},
{"NM", "NEW MEXICO"},
{"NY", "NEW YORK"},
{"NC", "NORTH CAROLINA"},
{"ND", "NORTH DAKOTA"},
{"MP", "NORTHERN MARIANA ISLANDS"},
{"OH", "OHIO"},
{"OK", "OKLAHOMA"},
{"OR", "OREGON"},
{"PW", "PALAU"},
{"PA", "PENNSYLVANIA"},
{"PR", "PUERTO RICO"},
{"RI", "RHODE ISLAND"},
{"SC", "SOUTH CAROLINA"},

{"SD", "SOUTH DAKOTA"},
{"TN", "TENNESSEE"},
{"TX", "TEXAS"},
{"UT", "UTAH"},
{"VT", "VERMONT"},
{"VA", "VIRGINIA"},
{"VI", "VIRGIN ISLANDS"},
{"WA", "WASHINGTON"},
{"WV", "WEST VIRGINIA"},
{"WI", "WISCONSIN"},
{"WY", "WYOMING"},
{"AE", "ARMED FORCES EUROPE, THE MIDDLE EAST, AND
CANADA"},
{"AP", "ARMED FORCES PACMC"},
{"AA", "ARMED FORCES AMERICAS"}
}
8.8.3 Mapping Assignment
Again, similar to domains, we may have mapping relationships that

exist within the data environment, but the existence of a pair in the sys-
tem automatically adds it to the mapping. It is slightly more complex
with mappings, since there must be some pivotal means for linking the
source value to the target value. In our case, we will allow the mapping
to be pivoted on a pair of data fields linking records then choosing the
two fields holding the source and target domain values. Mapping
assignment is a prescriptive rule, since it is through the data relations
described in the rule that the mapping is defined.
As an example consider the mapping between country names and
currency codes. In one table, each record with a country name is accom-
panied by a country code; in the second table, each record with a coun-
try code is accompanied by a currency code. We can use this definition.
Define Mapping COUNTRY_TO_CURRENCY:
COUNTRY_NAMES TO CURRENCY_CODES:
{
With Pivot Geography.CountryCode with Orders.CountryCode,
Map Geography.CountryName to Orders.CurrencyCode
}
In this definition, we have established the relation between two

tables based on a single pivot value (that is, the one that appears in both
the CountryCode fields of tables Geography and Orders). Using this
relation, we assert that an implicit mapping exists between the Coun-
tryName field of the Geography table and the CurrencyCode field of the
Orders table. This mapping can now be embedded in the metadata
repository and can be used in other rules.
8.8.4 Mapping Membership
Mapping membership is a proscriptive rule that asserts that the relation

between two attributes or fields is restricted based on a named map-
ping. More formally, if one specifies that two fields, A and B, are mem-
bers of domain mapping Z, where Z is a mapping between domain X
and domain Y, it means that any value in field A must belong to domain
X, any value in field B must belong to domain Y, and the pair of values
from fields A and B, (Uyb) must belong to the mapping Z.
Mapping membership is formally specified by indicating a map-
ping and two fields that belong to that mapping, making sure to indi-
cate which is the source field and which is the target.
Rule OrdersCountries:
(Orders.CountryName, Orders.CountryCode) Belong to Mapping
COUNTRIES TO CODES
8.8.5 Mapping Nonmembership
Conversely, mapping nonmembership asserts that a pair of values may

never belong to a named mapping. This rule is a proscriptive rule that
says that if one specifies that two fields, A and B, are not members of
domain mapping Z, where Z is a mapping between domain X and
domain Yy it means that any value in field A must belong to domain X,
any value in field B must belong to domain Y, and the pair of values
from fields A and B, (ayb) may not belong to the mapping Z.
Note that in this rule, we indicate that the fields must belong to the
domains. Violation of the rule only occurs if the two values belong to
their respective domains but the pair does not belong to the mapping. If
one of the fields contains a value that is not a member of its specified
domain, the rule is violated.
Mapping nonmembership is formally specified by indicating a

mapping and two fields that do not belong to that mapping, making
sure to indicate which is the sourcefieldand which is the target. Assume
that we have a mapping called DISCONTINUEDJTEMS between
product line names and model numbers, which represents all product
name and model combinations that are no longer sold. An example of a
rule forcing nonmembership could be used to make sure that a cus-
tomer order does not refer to a discontinued item.
Rule Discontinued:
(Orders.ProductLine, Orders.ModelNumber) Does not Belong to
Mapping DISCONTINUED_ITEMS
Figure 8.3 details the types of rules.
8.8.6 Mappings as Functions
Mappings can also be used to provide the resulting mapped value

belonging to the second domain from a submitted value belonging to
the first domain. In other words, a mapping can be used as a relation. If
the mapping is one-to-one, then it is a function. If the mapping is one-
to-many, then the result of applying the mapping is a randomly selected
value from the set of values mapped in the second domain.
Functional Mapping DEFINE <mapping'name>:<sourc-domam> TO <target

A list of functions describing mapping
Definition dotnam> AS <function-speC'list>
Enumerated Mapping Definition of a rule specifying those DEFINE <mapping-name>:<source-domain> TO <target

Definition value pairs that belong to the mapping domam> AS {<value-pair-list>}
Define Mapping <niapping-name>: <source-domam> TO

A rule that indicates that two attributes'
<target domam> AS:
Mapping Assignment values define the mapping via a
{Pivot Ktablel .column> with <table2.coluntn>
"hooked** third pivot attribute
Map <source attributo to target attribute>}
A rule that indicates that two attributes' (<source-attribute-name>:<target-attribute-name>)

Mapping Membership
values must conform to the mapping BELONG TO <mapping-name>
Mapping A rule that indicates that two attributes* (<source-attribute-name>:<target-attribute-name>)

Nonmembership values may not conform to the mapping DOES NOT BELONG TO <mapping-name>
FIGURE 8.3 Mapping rules

8.9 RELATION RULES
The relation rules refer to data quality rules that apply to more than one
data field. We have two versions of these rules: intra-record and inter-
record. The intra-record version is limited to fields within a single
record or message, and the inter-record version allows referral to fields
in other records, messages, or tables.
There are four kinds of relation rules. Three kinds complete-
ness, exemption, and consistency rules are proscriptive rules. One
kind derivation is a prescriptive rule that describes how some
fields are filled in. Completeness rules describe when a record is fully
complete. Exemption rules describe times that a rule may be missing
data. Consistency rules indicate the consistency relationship between
fields.
8.9.1 Completeness Rule
Completeness is one of the dimensions of data quality discussed in

Chapter 5, and, therefore, it is useful to have a rule overseeing it. Com-
pleteness governs the proper population of a record's fields, possibly
depending on other fields' values. Completeness rules are conditional
assertions.
A completeness rule specifies a condition followed by a list of
attributes that must have values. For example, a sales order must have a
biUing address if the total of the bill is greater than $0.00.
Rule OrdersComplete:
IF (Orders.Total > 0.0), Complete With
{Orders.Billing_Street,
Orders.Billing_City,
Orders.Billing_State,
Orders.Billing_ZIP}
All completeness rules must have a condition and a list of attribute
names. If the condition evaluates to true, then each attribute enumer-
ated in the list must have a non-null value. Completeness conditions
may refer to more than one attribute, and these conditional clauses are
grouped together using the logical operators described in Table 8.1.
8.9.2 Exemption Rule
There are times when some attributes are not required to have values,
depending on other values in the record. An exemption rule specifies a
condition and a list of attributes, and if the condition evaluates to true,
then those attributes in the list are allowed to have null values. Exemp-
tion rules, like completeness rules, are conditional assertions.
For example, a catalog order might require the customer to specify
a color or size of an article of clothing, but neither is required if the cus-
tomer orders a nonclothing product.
Rule NotClothing:
IF (OrdersJtem^Ciass != "CLOTHING") Exempt
{Orders.Coior,
Orders.Size
1
If a null value rule has been specified for any of the attributes in the
list, then the null value representations for those attributes may appear
as well as traditional nulls.
8.9.3 Consistency Rule
Consistency is another of the dimensions of data quality described in

Chapter 5. Consistency refers to maintaining a relationship between two
(or more) attributes based on the actual values of the attributes. Consis-
tency rules are conditional assertions, expressed as "if-then" statements
that include both a condition and a consequent. Our consistency rule
indicates that if a particular condition holds true, then a following conse-
quent must also be true. For example, if a human resources data man-
ager wants to make sure that the record for each employee has a salary
that is within the range for that employee's title, we might have a set of
consistency rules:
Rule Salary 1:
IF (Employees.title = "Staff Member") Then (Employees.Salary >=
20000 AND Employees.Salary < 30000)
Rule Salary!:
IF (Employees.title = **Senior Staff Member") Then
(Employees.Salary >= 30000 AND Employees.Salary < 40000)
Rule Salary3:
IF (Employees.title = "Manager") Then (Employees.Salary >=
40000 AND Employees.Salary < 50000)
Rule Salary4:
IF (Employees.title == ''Senior Manager'*) Then (Employees.Salary
>= 50000 AND Employees.Salary < 60000)
For each consistency rule, there is a condition and a consequent.
The condition may consist of clauses referring to more than one
attribute, and the consequent may also refer to more than one attribute.
8.9.4 Derivation Rule
A derivation rule is the prescriptive form of the consistency rule. In a

derivation rule, if a condition evaluates to true, then a consequent dic-
tates how another attribute's value is defined. A derivation rule defines a
data dependency between the attributes in the condition and the
attribute in the consequent.
As an example, if we want to fill in the total amount charged for a
catalog order, that total would be computed as the sum of the product
of the number ordered and the price, plus the computed sales tax of
5 percent.
Rule DerivedTotal:
IF (Orders.NumberOrdered > 0) Then {
Orders.Total = (Orders.NumberOrdered * Orders.Price) * 1.05
}
Derivation rules can expose "virtual attributes." Such attributes
don't really need to have their values filled in because the value can be
derived as long as the dependent attributes have been filled.
8.10 Table, Cross-Table, and Cross-Message Assertions
So far we have mostly focused on rules that apply between the values of
attributes within the same record. In fact, there are rules that apply
between records within a table, between records in different tables, or
between fields in different messages.
The rules in this section are important in terms of traditional rela-
tional database theory. The most powerful of these rules is the func-
tional dependency rule, which indicates the relationship between the

values of attributes across all records in a table. The other rules refer to
what is called referential integrity.
8.10.1 Functional Dependency Rule
In database jargon, a functional dependency between two columns X

and Y, means that for any two records Rl and R2 in the table, iffieldX
of record Rl contains value x and field X of record R2 contains the
same value x, then if field Y of record Rl contains the value y, then field
Y of record R2 must contain the value y. In other words, attribute Y is
said to be determined by attribute X.
In Section 8.9 we looked at consistency rules applied to records
that defined restrictions on some attribute values based on the values of
other specific attributes. These consistency rules are applied at the
record level, which means that if the condition does not apply, there are
no constraints on the target attributes' values. Functional dependency
rules, on the other hand, describe the same inter-record constraints on
all records in a table.
Essentially, a functional dependency rule is like a consistency rule
that covers all records in the table. Based on this understanding, we can
see that aside from the standard database use of functional dependence
to ensure normalization, functional dependencies may actually repre-
sent a governing rule about what is represented in the data in other
words, a business rule!
Functional dependencies can hold between sets of attributes as
well. A functional dependency is likely to represent either an implicit
domain mapping or a business rule embedded in the data. We denote a
functional dependency by specifying two sets of fields: the defining set
and the determined set. For example, if we say that an employee's salary
is determined by her title and her years of service, and this rule holds
true for all employees, then this is a functional dependence that is
expressed like this.
Rule SalaryDetermination:
{Empioyees.Title, Employees.Years} Determines {Employees.Salary}
Functional dependencies are particularly powerful because they
can be manipulated using a series of inference rules.
1. Reflexivity If the set of values in attribute X is a proper super-

set of the values in attribute Y, then X determines Y.
2. Augmentation If X determines Y, then {X, Z} determines {Y, Z}.
3. Decomposition If X determines {Y, Z}, then X determines Y.
4. Transitivity If X determines Y, and Y determines Z, then X
determines Z.
5. Union If X determines Y, and X determines Z, then X deter-
mines {Y, Z}.
6. Pseudotransitivity If X determines Y, and {W, Y} determines Z,
then {W, X} determines Z.
These inference rules clarify the embedded business rules in a data-
base table. We can also use them to assist in the discovery of domain
mappings.
8.10.2 Key Assertions
A table key is a set of attributes such that, for all records in a table, no
two records have the same set of values for all attributes in that key set.
A primary key is a key set that has been designated as a key that identi-
fies records in the table. A key assertion is a proscriptive rule indicating
that a set of attributes is a key for a specific table.
Rule CustomersKey:
{Customers.FirstName, Customers.LastName, Customers.DD} Key
for Customers
If a key assertion has been specified, no two records in the table
may have the same values for the indicated attributes. Therefore, a key
assertion basically describes the identifiability dimension of data quality
as described in Chapter 5.
8.10.3 Primary Key Assertions
A candidate key is a key set consisting of attributes such that the com-
position of the values in the key set uniquely identify a record in the
table and that none of the keys in the key set may be removed without
destroying the uniqueness property. A primary key is an arbitrarily
selected candidate key used as the only record-level addressing mecha-
nism for that table. We can qualify the key assertion by indicating that
the key is primary, as in this example.
Rule CustomersPrimaryKey:
{Customers.FirstName, Customers.LastName, Customers.ID} Pri-
mary Key for Customers
A primary key defines an entity in a table, and, therefore, the
enforcement of a primary key assertion is that no attribute belonging to
the primary key may have null values. This demonstrates an interesting
point: Some rule definitions may have an impact on other rule defini-
tions. Our framework already has the facility for defining null value
rules, but implicit in a primary key rule are other null value rules! Fig-
ure 8.4 explains both key assertions.
8.10.4 Foreign Key Assertions
A foreign key represents a relationship between two tables. When the

values in field f in table T are chosen from the key values in field g in
table S, field S.g is said to be a foreign key for field T.f. A foreign key
assertion must indicate that a field in one table is a foreign key for a
field in another table. For example, if the part_number field in the
Orders table is taken from the part_numberfieldin the Parts table, then
we would have a rules like this.
Rule PartsForeignKey:
{Parts.part_number} Foreign Key for {Orders.part_number}
Specifiedfirstin the rule is the foreign key, followed by the key tar-
get. Since keys may be composed of multiple attributes, the rule may
specify more than onefieldfor both the foreign key and the key target.
Key Assertion The attribute list forms a key. l<attribute-list>] KEY FOR <table name>
Primary Key The attribute list forms the

l<attHbute-list>l PRIMARY KEY FOR <table name>
Assertion primary key.
FIGURE 8.4
The application of a foreign key assertion is in ensuring what is

called referential integrity. Referential integrity means that if table T
has a foreign key matching the primary key of another table S then
every value of the foreign key in T must be equal to the value of some
primary key in table S, or be completely null. Therefore, a foreign key
assertion specifies a consistency relationship between tables for all
non-null foreign keys that appear in table T, there must exist a primary
key with the same values in table S. For example, every part number
referred to in the ordered parts field must be listed in the parts table (see
Figure 8.5).
8.11 IN-PROCESS RULES
We must not forget that the quality of data also affects the way the data
is used, and if there is a processing stream that uses information, there
may be rules that govern automated processing. Therefore, we also
define a class of rules that apply "in-process." In-process rules are those
that make assertions about information as it passes between interfaces.
It is through the use of in-process rules that we can add "hooks"
for measurement, validation, or the invocation of any external applica-
tion. Before we can specify the rules themselves, we must also have a
way to express the data flow for the process, and we do that using pro-
cessing chains. Given a processing chain, we can insert into that chain
in-process rules.
8.11.1 Processing (or Data-Flow) Chains
A processing chain (also referred to as a data-flow chain) is a high-level

encapsulation of the processing steps that are applied to a flow of informa-
tion. In order to assign in-process rules, there must be steps in that process
The attribute list forms l<table name and attributo) FOREIGN KEY FOR
Foreign Key Assertion
a foreign key. Ktable name and attribute>
FIGURE 8.5 Foreign keys

at which the rules are to be appUed. Therefore, we have a formaUsm for

defining steps in a processing chain. Each processing step is either a source
for data, a target for data, or a processing stage for data. Each step is
declared with a name along with a set of comment fields (which can indi-
cate the actual operations that take place at that step). The flow of infor-
mation is indicated as a directed connection between two processing steps.
As an example, consider a process that reads customer order data
from a file, selects out a subset of orders for credit card processing,
processes payments, forwards the successfully paid orders to a second
data file and forwards the unsuccessfully paid orders to a reconciliation
file. Each step is defined first, followed by the connective statements.
For each connective statement, we also name the connector.
Source OrderFile;
Stage SelectOrders;
Stage ProcessPayments;
Target PaidOrders;
Target Reconciliation;
Selector: OrderFile TO SelectOrders;
OrderPay: SelectOrders TO ProcessPayments;
GoodPayment: ProcessPayments TO PaidOrders;
BadPayment: ProcessPayments TO Reconciliation;
This description is reflected in Figure 8.6,
Given a way to describe both processing stages and the connec-
tions between them, we now have a "hook" on which we can hang in-
process data quality and business rules.
8.11.2 Measurement Directives
A measurement directive associates a set of rules with a location in a

data-flow chain and indicates that at that point in the processing, those
rules are applied. For each rule, the determination is made as to whether
the data is in accordance with the rule or not, and the results are logged.
Measurement directives are used for two stages of data quality:
The first is the current state analysis and isolation of data quality prob-
lems, and the second is for continued data quality assurance. Measure-
ment directives act as probes into the data flow, giving the analyst the
chance to "look under the hood" to look for potential problems or to
just make sure everything is running soundly.
Order Input
Files
f Paid ^
GoodPayment "I Orders J
Type Name
BadPayment
Source Order Files
Stage Select Orders
Stage Process Payments
Target Reconciliation
Target Paid Orders
Channel
Channel Name Source Target
Type
Selector Input Order Files Select Orders
' Channel OrderPay Select Orders Process Payments
Channel GoodPayment Process Payments Paid Orders
Channel BadPayment Process Payments Reconciliation
FIGURE 8 . 6 Representation of an information chain
A measurement directive must specify either a processing stage, a

connector, a source, or a target and a set of rules. For example, if we
want to measure how frequently credit card fields have the wrong num-
ber of digits in the credit card number before the payment processing
stage, we might use this directive.
Rule CCNumbers:
Restrict CreditCardNumber: LENGTH{value) 16
Rule MeasureCCNumbers:
AT OrderPay Measure {CCNumbers}
8.11.3 Trigger Directives
A trigger directive indicates the invocation of a named operation at

some point in the data-flow chain. We may include a condition to trig-
ger the operation, or the operation may take place specifically when a
record or a message hits a certain point in the data-flow chain.
We first must have a means for naming operations that can be
invoked via triggers. All we really need is a declaration of the operation
by its name. For example, consider an operation that sends an e-mail to
your manager.
Operation EmailTheBoss;
When a message arrives at a particular point in the data-flow chain
and an optional condition is fulfilled at that point, the operation is trig-
gered. Therefore, a trigger directive includes either a processing stage,
source, target, or connector, an optional condition, and the name of the
operation to invoke. For example, if we want to e-mail the manager if an
order's value exceeds $10,000, we might have a trigger directive like this.
Stage ProcessOrders;
Trigger at ProcessOrders if (Orders.Total > 10000): {E-mailThe-
Boss};
We can include a list of operations that are invoked in sequential
order instead of just a single operation.
8.12 OPERATIONAL RULES
There are other opportunities for inserting in-process rules that do not
necessarily affect the data-flow chain but rather affect the data being
transferred. In this section we discuss transformation rules, that
describe how data is transformed from one form to another and update
rules that define when information should be updated.
8.12.1 Transformation Rules
A transformation rule describes how data are mapped from a set of

source locations to a destination location. These rules are used to define
a portion of the extraction, transformation, and loading process for
moving data between databases. Each transformation rule, which spec-

ifies how a single attribute is fulfilled, must indicate the sources of data
needed for seeding the ultimate value and the operations that define
how the attribute is ultimately defined.
For example, consider an international payment system that must
accept payments in euros and disburse payments in dollars. The dis-
bursement database field Disbursements.DoUarsPaid is transformed
from the Payments.Euros field in the payments database as a function
of the currency exchange factor stored in the DoUarsToForex mapping.
Transformation: Disbursements.Dollars =
{
Payments.Euros * DollarsToForex(**Euro")
)
A transformation rule specifies the target attribute and a transforma-
tion expression.
8.12.2 Update Rules
Update rules indicate when a data set should be updated, whether it

means the invocation of a new validation or the initiation of a new set
of in-process rules. Update rules are basically trigger rules governed by
timestamps. There are two directives that we need to allow update
rules. The first is the initiation directive, which initiates an operation,
and the second is the restart directive, which restarts an operation.
At <timestamp> Initiate: DWPopuIate
At <timestamp> Restart: MeasureCCNumbers;
Both the initiate and restart directives can apply to operations or to
in-process rules. An update rule, then, is indicated by a time/date-stamp
and a list of initiate or restart directives. For example, if we want to
restart the measurement directives after a certain time every day, we can
use an update rule to indicate this.
At 08:00:00 Restart MeasureCCNumbers;
This is shown in Figure 8.7.
MeasureCCNumbers BigOrder
AT OrderPay Measure IF (orders.total > 1000):
{CCNumbers} E-mailTheBoss
FIGURE 8 . 7 Inserting directives
8.13 OTHER RULES
There are other kinds of rules that are not discussed in this chapter. One
class of rules is the approximate or fuzzy rules, which specify assertions
allowing some degree of uncertainty. Another set is the navigation rules,
which indicate appropriate navigation paths that can be executed by
users at a Web site. Both of these classes will be treated at greater length
in other chapters.
8.14 RULE MANAGEMENT, COMPILATION, AND VALIDATION
Once data quality and business rules have been defined, there are two
more steps. The first is the accumulation and consolidation of all the
rules into one working set, and the second step is confirming that the
rules interact with each other correctly. These, together with the opera-
tion of storing, retrieving and editing rules, are part of the management
of rules.
8.14.1 Rule Management
A benefit of defining rules is that we have described a formality that is

simpler than a programming language and can express significant detail
about the business and data quality rules that govern a data system.
This simple formalism makes it possible for these rules to be defined,
read, and understood by a nontechnician. The ultimate goal is a frame-
work in which the data consumer can be integrated into the data qual-
ity process. In this way, data quality rules are integrated into the
enterprise metadata and can be managed as content that can be pub-
lished directly to all eligible data consumers.
8.14.2 Compilation
Rule compilation is more than just collecting rules. There is a possibility

for interaction between rules, since more than one rule can reference the
same set of data attributes. Another concern is the dimension across
which the rules are applied across a tuple, down a column, through-
out a table, or across more than one table. Rule compilation is the
process of collecting all the rules, ordering the rules, and preparing the
rules for validation.
Rule compilation also identifies the tables, columns, and rows that
are associated with rules. By collecting this information and allowing
users to browse this data, both mapped from rule to data element and
from data element to rule, users have a means for evaluating business
rule dependencies between rows, columns, and tables. This gives users a
valuable knowledge discovery tool.
In a more complex environment, rule compilation actually may
involve translation from the rule format to an executable format. This
may mean translating the set of rules into SQL calls or application code
that can be used for data validation.
8.14.3 Rule Validation
Before rules can be compiled into an executable format, the rules them-
selves must conform to their own set of validity constraints. Specifically,
the assertions about a data set cannot conflict with each other. For
example, we can make these two assertions about our data.
Rule: IF (Customers.Cust^Class != "BUSINESS**) Exempt
{Customers.Business_Phone
}
DATA QUALITY ASSERTIONS AND BUSINESS RULES 1 99
Rule: RequiredBizPhone:
Customers.Business__Phone nulls not allowed
Each of these rules by itself is completely understandable. The first rule
says that if the customer is not a business customer, the business phone
attribute may be empty. The second rule indicates that the business
phone field may not contain nulls. Put together, these two rules are con-
tradictory: The first says that nulls are allowed under some conditions,
while the second asserts that the same attribute may not be null. This is
an example of violating a validity constraint on data quality rules.
Some other kinds of validity rules include the following.
Restriction contradictions, where two restrictions create an asser-
tion that cannot be true
Uncovered consistency ranges, where consistency rules are defined
for only a subset of data values but not all the possible values that
can populate an attribute
Key assertion violations, where an attribute set specified as a key
is allowed to have nulls
8.15 RULE ORDERING
The last issue regarding rules is order. One of the nice things about the
rule system is that if we want to declare that the data conform to the
rules, all the rules and assertions hold true. But how do we test this?
The answer is that we must apply each rule in its own context and vali-
date that there are no violations.
8.15.1 Benefits of a Declarative System
This poses an interesting conundrum. A major benefit of having a rule

set is that we have a declarative means for specifying expectations. But
in order to actually test these rules, some of the rules must be trans-
formed into an executable sequence of tests, which implies an assigned
order of execution. It is because of the declarative nature of rules that in
reality there are no constraints on the execution of the tests, although
there are some exceptions.
The best way to order the rule execution, therefore, is dependent

on the method by which the tests will be performed. If an automated
method is used, then optimizing the runtime might be the overriding
variable. If your junior data analysts are going to do the work, then the
tests might be partitioned by table or by attribute. In other words, any
suitable method is appropriate.
8.15.2 Dependencies
The only exceptions, as just noted, come into play when the rule itself
includes a data dependence or a control dependence. A data depen-
dence occurs when one piece of the rule depends on another part of the
rule. Any of the rules with conditions and domain and mapping assign-
ments have data dependences inherent in them. A control dependence
occurs when an event must take place before the test is allowed. The
trigger, update, and measurement directives imply control dependen-
cies. Some rules may have both control and data dependence issues:
Consider the transformation directives.
A dependence is said to be fulfilled if the no further modifications
may take place to the dependent entity (attribute, table). A rule's depen-
dence is violated if a test for that rule is performed before the associated
dependencies have been fulfilled. The presence of dependences in the rule
set forces some ordering on the execution of rule tests. The goal, then, is to
determine the most efficient order that does not violate any dependence.
8.15.3 Benefits of Enforcing Dependence
Enforcing a valid order of execution in the presence of dependences

provides a very nice side effect: it allows a user to pinpoint the location
of a business rule or data quaHty violation at its earliest point in the
information chain. When a violation occurs, there is an execution trail
that can be used as a "debugging" device for the data steward. Not only
does this provide the capacity to prevent unexpected data from passing
through a quality gate, it also speeds up the problem determination and
discovery process.
A second benefit is that the dependence constraints can help in the
development of a user interface. Since the order of data acquisition can be
determined by the data and control dependences embedded in the rules, a

rule-directed sequence of query screens can be automatically generated.
Because the rules guide the actual data acquisition, we can enforce our
data quality and business rules on information as it enters the system a
significant step in building a high-quality knowledge enterprise.
8.16 SUMMARY
In this chapter, we looked at a formal way to define and describe our

data quality expectations. Maintaining high-quality information levels
allows a business to achieve a high level of automated information pro-
cessing, which in turn allows for greater scalability. Because bad data
can affect this operation, data quality rules are, in effect, business rules.
We then looked at the following nine different classes of data qual-
ity rules.
1. Null value rules Including null specifications, null values, and
non-null values
2. Value rules Restrict the actual values that can populate an
attribute
3. Domain membership rules Covered restricted sets of values with
business meanings from which attributes can draw their values
4. Domain mappings Associate values from one domain with val-
ues from another domain, also with business meanings.
5. Relation rules Including completeness, exemption, consistency,
and derivation rules
6. Table, cross-table, and cross-message assertions Including func-
tional dependencies and key assertions
7. In-process directives Including measurement directives and
trigger directives
8. Operational directives Including transformational rules and
update rules
9. Other rules Covered in other chapters
The use of a rule system has a number of benefits, especially when
it comes to managing business and data quality rules as content. In par-
ticular, there are opportunities for automating a large part of the data
acquisition and validation process, as well as the opportunity to pin-
point problem processes with respect to poor data quality.
9
MEASUREMENT AND CURRENT
STATE ASSESSMENT
The first step in the mechanics of initiating a data quality improvement

program is to perform an assessment of the current state of data quaUty.
This approach provides a means to identify and document those areas
with the greatest need of improvement as well as provide a baseline
against which further improvements can be measured. This process,
which we call current state assessment (CSA), is complementary to the
economic analysis described in Chapter 4 and uses the rules framework
described in Chapters 7 and 8.
The economic model described in Chapter 4 helps us determine the
financial impact of poor data quality and identifies those locations in
the information processing chain where there is a need for improve-
ment. But the financial model does not give us any insight into what
might be causing the poor data quality or any suggestions as to the
kinds of improvements that are necessary. It is at this point in the analy-
sis that the current state assessment adds value.
The CSA gives the analyst the opportunity to select and measure
specific aspects of data quality that might be the source of economic dis-
comfort. The results of this measurement are then used as the input to
the next stage of the economic model, which is evaluating the costs
associated with instituting improvements. This in turn helps in deter-
mining the return on investment associated with each improvement,
which can help focus the management decisions when committing to
particular improvement projects.
The result of the current state assessment is a document similar to
the data quality scorecard, except that instead of measuring financial
impact, we provide data quality levels associated with a set of data quality
203
dimensions measured at each location in the information chain. We can

then correlate the data quality measurements in the CSA to the high-
impact areas in the scorecard. We then use those measurements to sug-
gest target data quality measurements that will provide the needed
improvement in quality to resolve any negative economic impact.
In this chapter, we detail the evolution of a current state assess-
ment. This partially overlaps with the data quality scorecard, but we
will see that at some point the development diverges. What is important
for the CSA is that a choice is made about the kinds of data quality
issues that are selected for analysis.
We start with identifying the data consumers, then working back
to determine the information chain. At this point, we choose a subset of
the dimensions of data quality we discussed in Chapter 5 to measure.
The next step is to assert a set of what we call "sentinel rules" those
rules that can be most clearly articulated within the rules framework
and constrained by what is known about the data at any particular
point in the information flow.
These sentinel rules are then used for measurement. There are two
kinds of measurement: static measurement, which measures the rules
using a snapshot of the information under investigation, and dynamic
measurement, which measure information in transit as it passes
through any particular point in the information chain.
9.1 IDENTIFY EACH DATA CUSTOMER
A key to performing a successful current state assessment is identifying

the data customers that consume the information produced. Because
these customers are most likely to have experience with the quality of
the information, and also because their investment is mostly in the
value of the information produced instead of the information produc-
tion process, they are more likely to provide anecdotal if not verifiable
evidence of low data quality. The data customers are also more likely to
report to the data quality analyst the effects low data quality.
We should note here that not all data consumers are real, live
people many are other processes or external automated customers.
These are often more difficult to identify, but they are just as important
within the "data ecosystem."
Building the information chain requires that we know the ultimate
destination of the data. Whether the consumer is an automated system
MEASUREMENT AND CURRENT STATE ASSESSMENT 205
or not, we first tabulate the set of consumers as endpoints in the infor-

mation chain. Starting with each endpoint, we can trace backward how
the information reached that point.
9.2 MAPPING THE INFORMATION CHAIN
Let's review the steps in building the information chain that we dis-
cussed in Chapter 4.
We define a set of processing stages through which data moves inside a

system. Remember, these stages are not all-inclusive but rather a place
to start building the information chain.
ing stage.
stage.
required is called a decision-making stage.
made at a decision-making stage is executed, possibly affecting
other processing stages or the data delivery stage.
9. Data consumption As the data consumer is the ultimate user of
processed information, the consumption stage is the exit stage of
the system.
9.2.2 Directed Information Channels
Information channels are pipelines indicating the flow of information

from one processing stage to another. A directed information channel is
additionally attributed with the direction in which data flows. The
delivery of supplier data to an acquisition stage is performed through
an information channel directed from the supplier to the acquisition
stage. We like to make our information channels directed because there
may be bidirectional communication taking place between any two
points, and this makes it possible to differentiate between data flowing
in the different directions.
9.2.3 Information Chain Map
A map of an information chain is represented by the combination of the

processing stages connected by directed information channels. Because
we will need to analyze each location in the information chain where a
data quality problem might arise, we need a naming scheme for the sites
on the map. Each processing stage should be labeled with a name and
attributed with its stage type. Information channels are named by com-
posing the source and target processing stage names, separated by a
hyphen.
9.2.4 Degree of Information Load
Each processing stage in the information chain can be characterized by

the amount of information that passes into and out of that stage. We
can count the number of directed channels that enter and leave the
stage.
The number of directed information channels entering the stage is
the degree of information input. A processing stage with a high degree
of information input is called a collector. The number of directed infor-
mation channels leaving the stage is the degree of information output. A
processing stage with a high degree of information output is called a
broadcaster. The sum of the information channels entering and leaving
each processing stage is called the degree of information load. A pro-
cessing stage with a high degree of information load is likely an impor-
tant node in the information chain and is called a critical junction.
9.2.5 Creating the Information Chain
By starting with the endpoints and working backward, each critical

processing point and information channel in the information chain can
be identified and labeled. At each step, we ask the people responsible
for the processing where they get their input and where the output goes.
This way, the information chain is built incrementally and we know
which employees or suppliers are responsible at each point in the data
flow.
When the information chain is in reasonable shape (one can reason-
ably follow the flow of data from its supply or creation point to a set of
consumers), the next stage of analysis can begin. Since the information
chain is used as input to both the economic framework and the current
state assessment, a completed information chain should be published
and archived as enterprise metadata.
g.3 CHOOSE LOCATIONS IN THE INFORMATION CHAIN
When the information chain has been published, the next part of the
current state assessment is to select locations in the information chain
to use as the measurement points. The choice of location can be depen-
dent on a number of aspects.
1. Critical junction Any processing stage with a high degree of
information load is likely to be a site where information from dif-
ferent data sources are merged or manipulated.
2. Collector A collector is likely to be a place where information
is aggregated and prepared for reporting or prepared for storage.
3. Broadcaster A broadcaster is likely to be a processing stage
that prepares information for many consumers and, therefore,
may be a ripe target for measurement.
4. Ease of access While some processing stages may be more
attractive in terms of analytical power, they may not be easily
accessed for information collection.
5. High-profile stages A processing stage that consumes a large
percentage of company resources might provide useful measure-
ment information.
The important point is that a number of locations in the informa-
tion chain be selected, preferably where the same data sets pass through
more than one of the locations. This will add some depth to the mea-
surement, since place-stamping the data along the way while registering
data quality measurements will help determine if there are specific fail-
ure points in the information chain.
9.4 CHOOSE A SUBSET OF THE DQ DIMENSIONS
As we discussed in Chapter 5, there are many dimensions across which

we can measure data quaHty. Because the number of dimensions is rela-
tively large, the next stage of the current state assessment is to select a
subset of the dimensions of data quality specifically for measurement.
Frequently, the kinds of problems that appear will suggest the
dimensions of choice. If not, there are some logical choices to make
with regard to generating a comprehensive CSA. It is a good idea to
select at least one dimension from each of thefiveclasses of data quality
dimensions. The resulting CSA will then provide insight into the quality
of the organization's data models, data values, data domains, presenta-
tion, and information policy.
9.5 IDENTIFY SENTINEL RULES
When we actually measure the levels of data quality, we will be looking

at the degree to which what we examine meets our expectations,
according to our definition of data quality from Chapter 3. We will
define "sentinel rules" rules that particularly capture the essence of
the quality of data as per data consumer expectations. They are called
sentinel rules because they act as a "redflag,"or sentinel, in terms of a
location in an information chain where there might be additional work
associated with low data quality.
Sentinel rules can be associated with the data model, the data val-
ues, the data presentation, and so forth. They manifest themselves in
the way that everyday operations staff members attack data quality
problems, as well as in ways that increase levels of frustration from
using the data.
As a simple example, consider a direct marketing company that
sends out millions of addressed mailings. Some of the addresses on the
mailing list are incorrect and thus undeliverable, and they are returned
to the company. The operation would be to investigate each undelivered
piece of mail to understand why it was returned. The goal would be to

correct the incorrect addresses by looking up the names in a current
address directory. The sentinel rule here is embedded in the reaction to
the data quality incorrectness. The real assertion would be that no mail
would have an incorrect address.
The more formal definition of a sentinel rule is an assertion that
characterizes the minimum expectation of the most important aspect of
data quality at a specific point in the information chain. Sentinel rules
distinguish between assertions about data that are meaningful in the
business environment from those that are irrelevant. Measurements of
the data quality associated with a sentinel rule are an important part of
the current state assessment, since the results of these measurements can
give a sound characterization of the general quality of data overall.
How do we identify a sentinel rule? Since sentinel rules are the for-
mal definition of vague notions of poor data quality, the best way to
identify them is by asking employees to relate their most interesting
anecdotes about data quality. Typically, these stories will revolve
around situations that required the most cleanup work, and if any anec-
dotes reflect multiple occurrences or are related by several different
employees, that's a good indication of an underlying sentinel rule. The
rule itself is abstracted as the opposite of the problem in an anecdote.
For example, if a tale describes how an error in one attribute in a
particular data record caused an overnight batch processing to halt,
requiring a 2 A.M. paging of a support personnel team to determine the
cause of the problem and fix it, the underlying sentinel rule describes
the expectation that the one attribute must have its correct value. Fur-
ther exploration should reveal which values are good and which are
not, giving us a formal definition of the sentinel rule.
9.6 MEASURING DATA QUALITY
The next sections detail some ways to measure data quality in terms of
the dimensions of data quality introduced in Chapter 5. In each section,
we review a dimension of data quality and suggest ways to measure it.
Of course, these are only starting points, and the ways to measure data
quality may be dependent on the specific situation.
9.7 MEASURING DATA QUALITY OF DATA MODELS
First let's examine the dimensions of data quality affecting data models.
In most cases, data quality associated with the data model is likely to be
measured using static metrics. Each section will review the dimension of
data quality and provide suggestions for measuring its level.
9.7.1 Clarity of Definition
Clarity of definition refers to the names and the naming conventions

assigned to tables, fields, and relations within the system. The names
assigned to tables and fields should clearly reflect the meaning contained
in the sets of information they represent. Naming of different fields should
be unambiguous, but similar attributes should share distinctive names.
There are a number of ways to measure clarity of definition:
Measure the degree to which there is a defined organizational
naming convention.
Measure the degree to which the named objects conform to that
convention. This quantifies how well the members of an organiza-
tion participate in the naming convention.
If there is no organizational convention, measure the number of
different names used for attributes holding the same kind of data.
This quantifies the levels of de facto naming standardization.
Assign a score to each table and attribute name quaUfying how
well it describes the data it holds. This quantifies the relationship
between name and meaningfulness.
9.7.2 Comprehensiveness
Comprehensiveness is a measure of how well the data model encom-

passes enough information to accommodate the current needs of all the
users of the data set, as well as broad or extensible enough to provide
for future user needs.
We measure comprehensiveness by looking at how easily the data
model adapts to general use. We can measure the following:
The number of requests over a period of time for adding new
tables to the data model.

attributes to data tables. These two items will quantify how well
laid-out the original model was.
The degree to which auxiliary database systems are being used in
the organization. This quantifies whether the original model is
being abandoned.
9.7.3 Flexibility
A data model's flexibility reflects the capacity to change in reaction to

new user requirements. Flexibility is similar to comprehensiveness, the
difference being that comprehensiveness addresses what is planned a
priori, whereas flexibility addresses those situations that crop up but
were not part of the original plan.
We measure flexibility in ways similar to measuring comprehen-
siveness:
tables to the data model
attributes to data tables
The number of times an attribute is overloaded
The amount of code needed to implement new functionality
The amount of time it takes to implement new functionality
9.7.4 Robustness
Robustness represents the ability to reflect changes in the modeled

world without excessive changes to the data model. Robustness also
involves defining attributes in ways that adapt to changing values with-
out having to constantly update the values.
Robustness can be measured as a function of the changes made to
already existing tables and attributes. We can get a quantitative mea-
sure of robustness by counting the number of times within a specified
period that there have been changes made to the data model aside from
adding new tables or columns. Examples include changing the size or
type of an attribute, changing relations between tables, modifying keys,
and so forth.
9.7.5 Essentialness
Other than for specific needs such as planned redundancy or the facili-
tation of analytical applications, a data model should not include extra
information. Extra information requires the expense of acquisition and
storage and tends to decrease data quality levels. Also, there may be a
lot of data attributes for which space has been allocated, yet these
attributes are not filled. Statistics regarding the frequency at which data
tables and attributes are referenced is a good source for measuring
essentialness:
Here are some ideas for ways to measure essentialness.
Count the number of data elements, tables, and attributes that are
never read.
Count the number of data elements, tables, and attributes that are
never written.
Count the number of redundant copies of the same data in the
enterprise.
Granularity refers to the number of objects used to represent one

notion. It is defined on a scale from "fine," where a data field can take
on more values, to "coarse," where it can take on fewer values. In a
table, we can use more attributes to represent finer granularity. Fine
granularity provides more detailed information, but there are added
costs for storage and an increased possibility for introducing errors.
We can measure the granularity these ways:
Count the number of attributes allocated for representing a spe-
cific component of a model (for example, are sales statistics mea-
sured by quarter or by month?).
Determine if any attributes are used to represent more than one
notion (attribute overloading).
Determine if any data domains are limited in their ability to
express the represented notion.
9.7.7 Precision of Domains
While the number of attributes in a set characterizes granularity, preci-

sion refers to the number of values that a single attribute may take. The
precision is partially determined by the data type of the attribute, but it
is also predicated on subtyping that is applied to the domain.
Domain precision can be measure these ways:
Count the number of possible values that an attribute can take (if
the attribute data type is afiniteset).
Determine the degree to which the data type has been subtyped.
This can be seen as a characterization of the specialization
assigned to specific domains (for example, two-letter strings vs
U.S. state abbreviations vs Northeast U.S. state abbreviations).
Determine the precision of attributes that can take on a represen-
tation of a continuous domain (such asfloatingpoint numbers or
decimals).
9.7.8 Homogeneity
Homogeneity refers to the overloading of value classes within a used

attribute. While a data model is intended to maintain a single class of
entities, the use of the model may evolve to overload the data set to
maintain multiple classes of entities within the same database.
Homogeneity can be measured in terms of the degree to which an
entity is being used to exclusively represent the real-world object it is
meant to model. This can be done these ways:
Identify all attributes that have been overloaded to represent more
than one object.
Count the number of attributes that take values from more than
one defined data domain.
Examine the data domains used by all attributes and make sure
that each describes a single aspect of the data model.
9.7.9 Naturalness
Naturalness is a qualification of the relationship between what is being

modeled and the model's representation, meaning that each represented
attribute should pair with a natural object in the world being modeled.
Also, each attribute should represent a single fact about the object and
pull its values from a meaningfully natural (not artificially restricted)
domain.
Naturalness is a qualitative dimension, but it can be measured by
grading the degree to which the data in the collection match the objects
that are being modeled. This can be measured these ways:
Does each data value pair with a real-world object.^
Does each attribute reflect a single aspect or fact about the object
being modeled?
9.7.10 Identifiability
There is a presumption that when each entity in a particular set of enti-

ties (used in the generic sense, not the relational database sense) is
expected to be unique, there must be some absolute means for distin-
guishing between any pair of entities. If not, there is a possibility that
there are multiple records representing the same individual entity.
Therefore, each entity type must allow for unique identification.
Identifiability is measured by determining which attributes (or sets
of attributes) should represent unique entities, as well as a means for
distinction, and then measuring how many duplicate records exists for
the same entity. In Chapter 13 we explore ways to determine the identi-
fiability attributes (commonly known as a primary key).
9.7.11 Obtainabiiity
Obtainability qualifies whether the information to be modeled can be

collected and stored. This obtainability may be characterized by a mea-
sure of the ease and feasibility of collection and storage, as well as more
obtuse issues such as legality of information collection or storage. Build-
ing a model to represent values that are difficult (or impossible) to obtain
will probably result in having attributes that will remain mostly empty.
There are two kinds of obtainability that can be measured. The

first is whether data attributes represent values that are obtainable at
all. The second measures the difficulty of obtaining data. These are
measured these ways:
Assign a grade that scores the difficulty in obtaining information,
and score each table and attribute with an obtainability grade.
Count the number of people required to participate in obtaining
the data.
Measure the amount of time required to obtain data.
9.7.12 Relevance
An irrelevant attribute is one that, if it were removed, would have no

effect on either the application as it is being run today or on any of the
planned features of the application. This is similar to essentialness, but
it is more focused on data that is being collected or was already col-
lected and no longer has any business relevance.
Relevance is measured by counting the number of attributes and
tables, which if removed would have no business relevance. This implies
some degree of reverse engineering to find out what the intended use is
for each table and its attributes, but this exercise is particularly useful in
that the reverse engineering is a good source of knowledge/content to be
collected and managed.
9.7.13 Simplicity/Complexity
The simplicity dimension refers to the degree of complexity in the data

model, whether it is due to extra attributes or a complicated relation-
ship between attributes. Complexity in the data model manifests itself
over time in terms of complications in the associated application code.
Simplicity (or more accurately, complexity) is very difficult to mea-
sure, and consequently, we must choose some aspect of the model to
measure. Since it is commonly accepted that the more interconnections
there are within a system, the greater the complexity, one place to start
is by examining the number of tables in the system and the number of
relationships that exists within each table and across tables.
A large number of anything in a data model may reflect a lack of
simplicity. We can count (1) the number of attributes in each table, (2)
the number of tables in a database, and (3) the number of foreign key
relations in a database.
9.7.14 Semantic Consistency
Semantic consistency refers to consistency of definitions among attrib-

utes within a data model, as well as similarly named attributes in differ-
ent data sets. Semantic consistency is a dimension that characterizes
how the meanings and the names of objects within the data set are con-
sistent.
Here are some ways to measure semantic consistency:
Determine the degree to which similarly named attributes make
use of the same data domain.
Determine the degree to which data attributes using the same
domain are similarly (or exactly) named.
Grade data attribute definitions in terms of consistency across the
enterprise data resource.
9.7.15 Structural Consistency
Structural consistency is a characterization of the care taken by the

database modelers, administrators, and stewards in guaranteeing
"strong-typing" of similar attributes. This dimension can be measured
across two axes.
Thefirstregards how different attributes of the same virtual type
are represented. For example, are customer names represented in
all database tables using threefieldssuch asfirstname, last name,
middle initial? Or are some represented using just last name and
first name without the middle initial? The measurement here
counts the number of times the same kind of item is represented in
more than one way.
The second regards the degree to which there are different repre-
sentations for the same value. In other words, this measures the
different ways that the same data item appears in the data envi-
ronment. For example, numeric data can be represented using
integers, character strings, decimals, and so on. Another example
is to ask if state names are spelled out or abbreviated.
There is a subtle difference between these two measures. The first is

looking at how the same idea is spread out across the columnar struc-
ture of a database table; the second looks at whether different data
domains are used to represent the same value.
9.8 MEASURING DATA QUALITY OF DATA VALUES
Next, we look at how to measure the data quality of data values. This is
the kind of measurements most frequently associated with data quality
and, consequently, is also the easiest for which to secure management
support.
9.8.1 Accuracy
Data accuracy refers to the degree with which data values agree with an
identified source of correct information. There are different sources of
correct information: a database of record, a similar, corroborative set of
data values from another table, dynamically computed values, the
result of a manual workflow, or irate customers.
Accuracy is measured by comparing the given values with the iden-
tified correct source. The simplest metric is a ratio of correct values and
incorrect values. A more interesting metric is the correctness ratio along
with a qualification of how incorrect the values are using some kind of
distance measurement.
It is possible that some of this measuring can be done automati-
cally, as long as the database of record is also available in electronic for-
mat. Unfortunately, many values of record are maintained in a far less
accessible format and require a large investment in human resources to
get the actual measurements.
9.8.2 Null Values
A null value is a missing value. Yet, a value that is missing may provide
more information than one might think because there may be different
reasons for the missing value, such as:
Unavailable values
Not applicable for this entity
There is no value in the attribute's domain that correctly classifies

this entity
The value may actually be missing.
The simplest way to measure null values is to count the number of
empty or null attributes that are expected to have a value. More precise
measurements can be collected after specifying the exact rules regarding
the types of null values and measuring conformance to those rules. This
involves specifying rules such as described in Chapter 8, and automati-
cally checking conformance.
9.8.3 Completeness
Completeness refers to the expectation that certain attributes are

expected to have assigned values in a data set. Completeness rules can
be assigned to a data set in three levels of constraints and can therefore
be measure using these constraints as well.
1. Mandatory attributes require a value the measure counts the
number of records conforming to some assertion but still missing
the value. As an example, stock option records must have an
expiration date field; any records without that attribute filled in
are incomplete.
2. Optional attributes, which may have a value again, this must
be measured based on conformance to the specified data quality
rule described in Chapter 8.
3. Inapplicable attributes (such as a maiden name for a single male),
which may not have a value. The measurement, again, is based
on conformance to a specified data quality rule.
9.8.4 Consistency
Consistency can be curiously simple or dangerously complex. In its

most basic form, consistency refers to data values in one data set being
consistent with values in another data set. Note that even though two
data values drawn from separate data sets may be consistent with each
other, both can be incorrect, so measuring consistency is not the same as
measuring correctness.
How do we measure consistency? The simplest measure is to per-

form record linkage among the data sets under investigation and verify
that the shared attributes in each set have the same values. The reported
measure is the number of entities that do not have consistent represen-
tations across the enterprise.
9.8.5 Currency/Timeliness
Currency refers to the degree to which information is current with the

world that it models. Currency can measure how "up-to-date" informa-
tion is and whether it is correct despite possible time-related changes.
The measure for currency can be similar to that of consistency check-
ing the data against a known up-to-date source. The measure counts the
number of records that are not current.
The measurement for timeliness involves defining the time criteria
and constraints that are expected for the arrival of the data and then
measuring in the period how frequently the data are available when it is
expected. A more interesting measurement details how late the data
actually are. In other words, if the data acquirer expects that delivery
will take place by a certain time each day, we would measure not only
how often the data is delivered on time but exactly what time the data is
delivered, how close to the expected time it was delivered, and the range
of lateness over a period of time.
9.9 MEASURING DATA QUALITY OF DATA DOMAINS
Having made use of the concept of data domains and mappings

between those domains, the next area of measurement is how well the
different players in the enterprise are cooperating with the use of enter-
prise-wide data subtypes.
9.9.1 Enterprise Agreement of Usage
The notion of abstracting information into a data domain implies that

there are enough users of the same set of data that it makes sense to
manage that data set separately as a resource instead of having separate
groups manage their own versions. The dimension of enterprise agree-

ment of usage measures the degree to which different organizations
conform to the usage of the enterprise data domain of record instead of
relying on their own data set.
The first measurement involves counting the number of shared
data domains that have been agreed to throughout the enterprise. We
can further measure this dimension by the number of organizations
within the enterprise that have agreed to use the shared data resource
versus the number of holdouts.
9.9.2 Stewardship
A dimension of data quality is the degree to which responsibility has

been assigned for the stewardship of information domains. Since a data
domain represents a collection of data values that are recognized as a
data source of record, it is advisable to appoint a steward to be respon-
sible for the upkeep and maintenance of the data domains. Stewardship
is measured by counting the number of groups that have appointed a
steward to be responsible for a set of data.
9.9.3 Ubiquity
As a data quality-oriented organization matures, the agreement of

usage will move from a small set of "early adopters" to gradually
encompass more and more of the enterprise. Ubiquity measures the
degree to which different departments in an organization use shared
reference data and can be measured two ways. The first way looks at
the number of distributed versions of the same reference data sets that
exist in the enterprise. As more groups agree to the usage agreement or
subscribe to a data ownership policy, they yield their control over their
versions of the data sets and, therefore, the number of replicated sets
will be reduced. The second way is counting the number of data users
subscribed to the shared data resource. Even if a suborganization
refuses to give up its copy of the reference data, there is nothing that
prevents individual users from subscribing to the shared resource on
their own. Therefore, we can measure individual usage as well.
9.10 MEASURING DATA QUALITY OF DATA PRESENTATION
Moving on to our next set of data quality dimensions, let's look at ways
to measure the quality of data presentation. Many of these dimensions
can only be measured through dialog with the users, since the presenta-
tion is an issue that affects the way information is absorbed by the user.
In each of these cases, though, the best way to get a feel for a score
in each dimension is to ask the user directly! In the question, formulate
each dimension definition as a positive statement for example, "The
format and presentation of data meets my needs" and see how
strongly each user agrees or disagrees with the statement.
9.10.1 Appropriateness
Appropriateness is the dimension that we use to categorize how well the

format and presentation of the data matches the users' needs. To mea-
sure this dimension, we must explore the history of the interaction
between the user group and the designers and implementers. If there are
many occurrences of user requests that result in changes to the data
model or to the data presentation layer, this may indicate a low level of
appropriateness.
9.10.2 Correct Interpretation
A good presentation provides the user with everything required for the
correct interpretation of information. Applications with online help
facilities are fertile territory for measuring this dimension. The help
facility can be augmented to count the number of times a user invokes
help and log the questions the user asks.
Applications without online help can still be evaluated. This is
done by assessing the amount of time the application developer spends
explaining, interpreting, or fixing the application front end to enhance
the user's ability to correctly interpret the data.
9.10.3 Flexibility
Flexibility in presentation describes the ability of the system to adapt to

changes in both the represented information and in user requirements
for presentation of information. Problems associated with flexibility are
reflected in requests to change a user interface or to add a new compo-
nent to a user interface. We can measure flexibility across two axes: (1)
counting the number of times users make requests for changes in the
presentation and (2) measuring the difficulty in implementing a change.
That is measured either in the amount of time required, the number of
parties involved, or the number of files, tables, or programs that need to
be modified.
9.10.4 Format Precision
The presentation of an attribute's value should reflect the precision of

the value based on both the internal representation and the needs of the
users. Typical issues include the ability to display all possible values and
to distinguish different values.
For example, floating point numbers may have a greater precision
in storage than in presentation, or decimal data types may maintain six
places to the left of the decimal point. In some cases, only two decimal
places are displayed despite the greater precision, with the value pre-
sented having been rounded. In some cases, this may be tolerable (for
example, when computing sales tax), but in others it may not be accept-
able (displaying the concentration levels of arsenic in drinking water).
To measure this dimension, it is necessary to prioritize displayed
values based on the importance placed on the users' ability to differenti-
ate by degree of precision. In other words, we need to isolate those vari-
ables to which users focus and measure how well the precision
conforms to the users' expectations.
9.10.5 Portability
In today's enterprise, it is likely that there will be a multitude of systems

and software platforms. In this kind of environment, being able to take
advantage of portability becomes critical in maintaining a high level of
application availability. Aside from the ability to migrate an application
from one platform to another, portability with respect to presentation

will incorporate the use of standards and recognized symbolism and the
ability to perform context-sensitive customization.
In an environment that uses different kinds of systems and applica-
tions, a portable interface is important so that as applications are
migrated from one platform to another, the presentation of data is
familiar to the users.
Alternatively, when we are "migrating users" (from one office to
another office, for example) it is important to continue to provide a
familiar interface (or not) dependent on user preferences. For example,
consider an employee's transfer from a United States office to a London
office. If the user has specified certain display preferences for daily
reports, should the report's presentation be updated based on location
when the user has moved? Would distance now be measured in kilome-
ters instead of miles?
We measure portability based on the following:
Subscription to data standards do the systems make use of
published recognized standards for representation of data?
Ability to internationalize how well do the systems reconfigure
the presentation of information depending on reported location?
Ability to allow personalized customization how well do the
systems reconfigure presentation of information based on user
preferences?
Use of known symbols and icons are they well known and con-
sistent throughout the enterprise?
Platform transparency does the presentation remain the same
when seen from different hardware (or software) platforms?
We can grade the answers to these questions to provide a measurement
of portability.
9.10.6 Representation Consistency
This dimension refers to whether instances of data are represented in a

format that is consistent with the domain of values as well as consistent
with other similar attribute values. In Chapter 5, we looked at the
example of the display of time in a nonmilitary, 12-hour format as
opposed to the 24-hour military format. If both formats are used in dif-
ferent parts of the same system, this could create confusion.
One of thefirstways to measure this dimension is to see whether

there is a style guide for data representation throughout the enterprise.
This can range from printed guideUnes that are to be followed by imple-
menters to electronic style sheets and templates that are integrated into
the development process.
A more granular investigation should be done to determine if there
is a standard representation format associated with every base data type
and domain. The next step would examine all presentations of values
associated with every data type and domain and see if the representa-
tion is consistent with the standard representation.
9.10.7 Representation of Null Values
In Chapter 5, wefirstdiscussed null values, and in Chapter 8 we looked

at the storage representation of nulls more carefully. Given that we
know that there are what we might call "standard nulls," as well as our
defined null types, when the null value (or absence of a value) is
required for an attribute, there should be a recognizable form for pre-
senting that null value that does not conflict with any valid values.
For example, a numeric null should not be displayed as a 0, which
represents a specific numeric value. For character strings, the "missing"
method of displaying the data should only be used if there are no distin-
guishing null type values.
If there are ways to distinguish the different kinds of null values,
there should be different ways to present them to the user. These factors
should be considered:
Are there special ways of representing null values?
If user-defined null types are used, can the user distinguish
between null types in the presentation.^
Can the user distinguish between a null value and valid default or
0/blank values?
9.10.8 Use of Storage
We have established that despite the increasing availability of low-cost

large data storage platforms, it is still important to measure how effec-
tively the storage is being used. This includes the following considerations:
If Storage performance is meeting user requirements

How well the use of storage will scale
Whether there are embedded performance glitches inherent in the
architecture
Whether data replication is being used to good advantage
Whether constraints on data normalization should be relaxed for
performance reasons
9.11 MEASURING DATA QUALITY OF INFORMATION POLICY
In Chapter 3, we looked at the issue of defining a data ownership policy,

and the topics of data policies (with respect to data sharing, usage, and
management) are all themes that flow through the technical compo-
nents of this book. An assessment of data quality in an enterprise can-
not dismiss the importance of the existence of and the concurrence with
general information policies.
In this section, we review those dimensions introduced in Chapter
5 associated with information policy, and we look at ways of character-
izing the degree of conformance with information policy.
9.11.1 Accessibility
Accessibility refers to the degree of ease of access to information, as well

as the breadth of access (whether all the information can be accessed).
We can measure this dimension by answering these questions:
1. For each data set, how easy is it to automate access? In other
words, we can measure the degree of ease of access by how easy
it is to implement a system to access the data. Some systems may
be set up such that special programs must be written to provide a
hook into the data, while in others, the means for access for any
data set is template and is easily configured.
2. Does the presentation allow for the display of all data? Some-
times the method of presentation does not adequately display all
pertinent data items, either by fiat or through the constraints of
"real estate." If an attempt is made to integrate a full display lim-
ited to a single screen, something is bound to be short-changed.
3. Is the presentation in a form that allows the user to absorb what
is being presented? This question differs from the previous one in
that even if all the information has been presented, can the user
get as much as possible out of the data.
4. How easy is it to get authorized to access information? This is
not a question of automation but instead measures the steps that
must be taken to authorize a user's access.
5. Are therefiltersin place to block unauthorized access.^ This ques-
tions whether there is a way to automate the access limits.
This dimension refers to that information that is allowed to be pre-
sented to any selected subset of users. Even though the last two questions
border on the issue of security and authorization, there is a subtle differ-
ence. The dimension of accessibility characterizes the means for both
providing and controlling access, but the dimension of security charac-
terizes the policies that are defined and implemented for access control.
9.11.2 Metadata
As we will explore in greater detail in Chapter 11, metadata deals with

the data in the system. When we attempt to measure the metadata pol-
icy, we investigate whether there is an enterprise-wide metadata frame-
work and the support policies that go along with it.
To measure the metadata policy, we must score the following ques-
tions:
1. Is there a metadata policy defined? In some systems, there is no
defined metadata policy, only small islands of managed meta-
data. This question focuses on whether the managers of the
enterprise have deemed it important to have an organizational
policy for managing metadata.
2. Is there a metadata repository? This next question focuses on the
infrastructure provided for metadata management. Even with a
metadata policy in place, there is a difference in usage if there is a
well-defined infrastructure available.
3. Where is metadata stored and under whose authority? Is meta-
data stored in a location accessible to all users? Can users browse
the metadata, especially if they are integrating a new information
system component? The greaterflexibilitythere is in reading
metadata, the better chance there will be of reusing information
that already exists instead of creating new tables and so forth.
MEASUREMENT AND CURRENT STATE ASSESSMENT S27
9.11.3 Privacy and Security
Privacy is an issue of selective display of information based on inter-

nally managed permissions. Security is similar to privacy, except that
privacy deals with protecting the entities being modeled by the system,
whereas security protects the data itself.
We speak of privacy when asking about the kind of information
that is being stored, the kind of control imposed over the storage and
retrieval of that information, and who controls access to that informa-
tion. In this chapter, we couple the measurement of privacy and secu-
rity, since they are very closely intertwined. When we want to measure
privacy policy, we need to ask these kinds of questions:
1. Is there a privacy policy? The lack of a privacy policy should be

noted, but existing privacy policies can be confusing and should
be examined to make sure that there is clarity in which levels of
privacy are protected and which are not.
2. If there is a privacy policy in place, how well is privacy pro-
tected?
3. Are there safeguards in place to maintain privacy and confiden-
tiality? How is data secured from unauthorized viewing?
4. Is private information encoded or encrypted in a way that pre-
vents unauthorized reading, even if the data itself can be
accessed?
5. Are there different storage procedures for confidential data ver-
sus nonconfidential data?
6. How does the policy enforce security constraints (such as loading
secure information onto a portable data device like a laptop, per-
sonal digital assistant, or even pagers and mobile telephones)?
Privacy and security policies should cover more than just the trans-
ference and storage of information. The privacy policy should also
cover the boundaries of any kind of information dissemination, includ-
ing inadvertent as well as inferred disclosure. An example of inadver-
tent disclosure is the lack of care taken when using a mobile telephone
in a public location. An example of inferred disclosure is when the accu-
mulation and merging of information posted to, then subsequently
gleaned from multiple Web sites can be used to draw conclusions that
could not have been inferred from the individual Web sites.
9.11.4 Redundancy
Redundancy refers to the management of multiple copies of the same

data sets. In some environments, redundancy is planned and desirable;
in others, it is a by-product of poor management. Planned redundancy
is desirable in that it provides fault-tolerance and may improve the
accessibility of information. On the other hand, redundancy becomes
unwieldy when there are many copies of the same data. In order to mea-
sure redundancy, we must look to these issues:
1. Is redundancy planned in the system or not?
2. If redundancy is planned, what is the hardware infrastructure for
storage.^
3. What is the policy for copy updates? Is updating performed in
real time, or is synchronization performed across the enterprise
at a specified time?
4. Who manages the source copy? Or are the multiple copies viewed
as equals with a synchronization process coordinating all copies?
5. If redundancy is unplanned, is it undesired?
6. With unplanned redundancy, how well are the copies synchro-
nized? Does the lack of synchronization create an environment
for the creation of information errors?
7. If redundancy is unplanned, how do multiple copies affect the
efficiency of operations?
8. How do multiple copies affect data synchronization across the
enterprise?
9.11.5 Unit Cost
The cost of maintaining information contributes greatly to a company's

ability to provide information-based services, information processing,
as well as general overall efficiency. When the unit cost of information is
low, greater volumes of information can be handled, which can open
opportunities for increased margins.
To measure unit cost, we must measure (1) the cost to obtain values,
(2) the cost of storage, (3) the cost of processing per unit, (4) the cost to
maintain levels of data quality, and (5) the cost of building data quality
into an information product, which must be weighed against the cost of
not having data quality. These are issues that we dealt with in Chapter 4.
MEASUREMENT AND CURRENT STATE ASSESSMENT S29
9.12 STATIC VS DYNAMIC MEASUREMENT
The different dimensions of data quality can be measured in two ways.

The first way, static measurement, explores the aspects of data quality in
a static set of data. The second method, dynamic measurement, assesses
the levels of data quality as information flows through a working system.
9.12.1 Static Measurement
When we talk about static measurement, we refer to measuring what

already exists within the system. Static measurement is a useful process,
since it can expose many opportunities for measuring against our differ-
ent data quality dimensions. A static assessment will essentially take a
snapshot of the data environment that can then be copied to another
location for separate analysis.
Static measurement is a tool that was used frequently in addressing
the Y2K problem. Many organizations completely duplicated their pro-
cessing systems in a separate, enclosed environment, grabbing copies of
all data at a particular point in time. The assessments performed looked
for instances of chronic problems that would be apparent in the system
itself, as opposed to day-to-day data.
9.12.2 Dynamic Measurement
Alternatively, dynamic measurement inserts probes into the system in

operation in order to assess and measure data quality on an ongoing basis.
Dynamic measurement has many advantages over static measurement.
Ability to see how well in-process data quality checks and trans-
formations work
Access to real-time measurements
Ability to track poor data quality through a working system
Unfortunately, because dynamic measurement must be integrated
into a working system, the coordination involved is much greater than
the static measurement approach.
9.13 COMPILING RESULTS
In Chapter 4, we discussed the construction of a data quaUty scorecard

that measures the cost associated with poor data quaUty. In this chapter,
we actually drilled down into this process in greater detail, trying to
characterize the specific issues regarding data quality and assessing
exactly how well or poorly the actual data quality meets our expecta-
tions.
In Section 9.4, we specified that in order to perform a current state
assessment, we would select a subset of the data quality dimensions for
measurement. Having made that selection, we can use the guidelines
laid out in this chapter to "insert probes" into the system and collect
measurements.
We collect as much information as possible about the measure of
data quality for each dimensions selected at each selected location. We
hope to find the location of those sentinel rules that can "make or
break" the data quality conformance in the system.
The last step in performing the current state assessment is to corre-
late the results with the data quality scorecard. For each location in the
information chain and for each area of data quality, we will now be able
to attach a set of dimensions of data quality and measurements of those
dimensions. The result is a more detailed report that discusses the fol-
lowing:
The cost of poor data quality throughout the system
The locations associated with information errors
The kinds of errors and/or nonconformance that exists in the
system
The measurement of the errors or nonconformance
The distribution of cost of poor data quality
If we can attribute the costs associated with low data quality with
specific dimensions of data quality, it presents the opportunity to use
those results to define the data quality expectations inherent in a set of
requirements. We discuss this process in Chapter 10.
9.14 SUMMARY
In this chapter, we discussed the process of actually measuring the levels

of data quality with respect to the dimensions of data quality intro-
duced in Chapter 5. Our goal is to perform current state assessment

that we then use as input to the data quality requirements definition
process.
We began the chapter by reviewing the process of mapping the
information processing environment into an information chain. The
information chain is the base map used to target the insertion of mea-
surements of the different data quality dimensions. We then examined
each of the dimensions of data quality from Chapter 5, deriving mea-
surement procedures.
Finally, we looked at the consolidation of the measurements into a
single report, correlating the occurrence of low data quality to the
actual costs associated with it, giving us the input to the next step, the
definition of requirements.
10
DATA QUALITY REQUIREMENTS
In Chapters 4, 6, and 9, we discussed creating a map of the information

processing environment, determining the costs associated with low data
quahty, and measuring the levels of data quality. We looked at the hier-
archy of associated roles within the organization, as well as the roles
that are played by different system operators and components. Through
the process of defining a data quality policy, hopefully we will have
assigned the responsibility associated with those roles to individuals
with the authority to manage data quality.
Now that we understand how low data quality can affect the eco-
nomics of the information faaory and we have explored ways to
expressly define rules regarding assertions about our expectations about
information, we have all the tools we need to express specific data quality
requirements.
We have mapped the flow of information throughout the system
and clearly identified each location within that chain. We have evalu-
ated the actual costs and impacts associated with low data quality, and
we have performed a current state assessment to determine the sentinel
rules that gauge data quality. We are now at a point where we can eval-
uate the results of the assessment by identifying the critical locations in
the information chain and attributing the costs of low data quality to
those locations, as well as define our minimum standard expectations
for data quality.
Our goal is to narrow the scope of each specific problem into a
manageable unit of work. This is done by completely restricting the
problem domain to specifics: the current measurement, the lowest mea-
sure of acceptance, and the optimal level of data quality. Additionally,
233
we want to be able to rank each project in terms of its importance to the

enterprise, and so we must attach some characterization of cost at each
work granule. In order to achieve this, these are the areas of further
analysis that we will perform:
1. Distribution of cost that combines the result of the data quality
scorecard, which attributed the information chain with the costs
associated with low data quality, and the current state assess-
ment, which attributed the information chain with the measured
levels of data quality.
2. Assignment of responsibility, which combines the result of the
current state assessment with the enumeration of data quality
actors within the environment, and assigns the responsibility for
maintaining data quality with the appropriate actor.
3. Determination of internal requirements, which involves setting
the required levels for data quality for the dimensions measured
in the current state assessment, for places in the information
chain where an individual internal to the organization is held
responsible.
4. Determination of external requirements, which involves setting
the required levels for data quality for the dimensions measured
in the current state assessment, for places in the information
chain where an individual external to the organization is held
responsible.
In this chapter, we are going to look at the results of our assessment
methods and see how to turn those assessments into actionable require-
ments. This process will incorporate the merging of collected data, along
with the use of a requirements analysis tool called use-case analysis. This
tool explores the way that the system is used in order to define how the
system should interact with the users. Even though the operations of the
system have already (most likely) been defined, we use this method to
impose the requirements for the quality of content, which we can then
use as input to the data quality and business rule definition process.
We also need to distinguish between "internal" requirements, which
are requirements that are imposed on actors and system components
within the control of the organization, and "external" requirements,
which are imposed on data suppliers external to the organization. We
deal with those requirements in Chapter 15.
In this chapter, we review some of the components that are used as
input to the requirements definition process and ways to bound each
DATA QUALITY REQUIREMENTS 235
requirement as a small task that can easily be implemented. Additionally,

we look at how to define requirements and store them as metadata within
a database table, yielding a flexible platform for modifying, searching,
and reporting on the compliance with the requirements.
10.1 THE ASSESSMENT PROCESS, REVIEWED
It is worthwhile to review the components of the assessment that we

have collected so far.
A set of actors within the system
A map of the information chain
The impacts of low data quality
The costs associated with those impacts
A selection of dimensions of data quality that are measured
Basically, we have archived a record of the impact that bad data
quality has with respect to the organization. When we are finished with
the requirements phase, we will essentially have "finished the job" by
translating the causes of those impacts into requirements.
10.1.1 ACTORS
Each stage in an information chain represents an action that can be per-

formed by one of the actors in the system. Our use of the term actor will
be apparent in Section 10.4, when we discuss use case analysis. Recall
that we defined the following actors within the system.
1. Suppliers Data suppliers provide information into the system.
2. Acquirers Acquirers accept data from external suppliers for
provision into the factory.
3. Creators Internal to the factory, data may be generated and
then forwarded to another processing stage.
4. Processors A processor is any agent that accepts input and gen-
erates output, possibly generating some side effects.
5. Packagers A packager collates, aggregates, and summarizes
information for reporting purposes.
6. Delivery agents A delivery agent delivers packaged information
to a known data consumer.
7. Consumer The data consumer is the ultimate user of

processed information.
8. Middle manager These people are responsible for making sure
the actors are correctly performing their jobs.
9. Senior manager The senior manager is responsible for the
overall operation of the factory.
10. Deciders These are senior-level managers associated with
strategic and tactical decision making.
10.1.2 Information Chain
The information chain is a graphical representation of the way the data

moves through the system. A node in the graph represents each process-
ing stage, and a directed link from one node in the graph to another indi-
cates a communication channel. Every node and every channel in the
graph is allocated a unique name. This allows us to associate measure-
ments, costs, and impacts with any specific point, or "location," in the
graph.
10.1.3 Impacts
In Chapter 4, we discussed how the realm of impacts was divided into

the soft impacts, which are clearly evident but hard to measure, and the
hard impacts, whose effects can be estimated and measured. Hard
impacts include the following:
Customer attrition
Costs attributed to error detection
Costs attributed to error rework
Costs attributed to prevention of errors
Costs associated with customer service
Costs associated with fixing customer problems
Costs associated with enterprisewide-data inconsistency
Costs attributable to delays in processing
Soft impacts include the following:
Difficulty in decision making
Time delays in operation
Lowered ability to effectively compete
Data ownership conflicts
Lowered employee satisfaction
In Chapter 4, we described how to build the data quality scorecard
by determining the locations in the information chain where these kinds
of impacts are felt. In Chapter 9, we discussed how to measure the
kinds of data quality problems that occur at each location by perform-
ing a current state assessment (CSA).
10.1.4 CSA
The process of performing a current state assessment includes the selec-

tion of some dimensions of data quality that apparently are related to
the identified impacts, which are pinned down to selected locations
within the information chain. During the CSA, we insert probes to mea-
sure the data quality levels of the selected dimensions, using defined
metrics for each dimension. Ultimately, we measured the conformance
of each dimension to either some level expected by users or some histor-
ical average as posited by the statistical control process.
10.2 REVIEWING THE ASSESSMENT
In reviewing the assessment, we are trying to achieve three goals. The

first goal, distribution of impact, gives us the opportunity to more fully
associate particular impacts we identified as part of the scorecard with
the specific data quality dimension we measured. The second goal is to
distribute the cost associated with each impact with the levels of data
quality measured as part of the assessment. The third goal, which we
will treat in its own section, is determining the expectations of data
quality to be incorporated as requirements.
10.2.1 Distribution of Impact
At each location in the information chain, we indicated the impacts asso-

ciated with low data quality and the measurements of specific dimensions
of data quality. Now, for each dimension of data quality measured, we

will assign some percentage of the overall impact at that location. For
example, if we had projected that associated with this location was a hard
impact due to error detection, at this point we should be able to specifi-
cally attribute the amount of error detection being performed for any of
the data quality dimensions measured. Thus, if there are errors due to a
problem with the data model, and there are errors due to timeliness, some
percentage of the full impact will be allocated to the data model problem
and the balance attributed to the timeliness problem.
Recall that our original impact decisions were estimates, and this
can produce two possible outcomes. The first is that we have correctly
estimated the impacts, and when we are finished, we will have fully
assigned all assumed impacts to specific data quality problems.
The second, and more likely, possible outcome is that our original
estimates were wrong, and we have discovered greater (or lesser) levels
of impact at a specific information chain location. Another possibility is
that there are impacts that we have not seen that suddenly become
apparent. In this case, it is worthwhile to adjust our original scorecard
to reflect this new insight, and then repeat the impact assignment step
until we are satisfied with the assignment.
10.2.2 Distribution of Cost
The second goal, distribution of cost, can be achieved after the assign-
ment of impact, as each impact is associated with a cost. This step is
actually a refinement of the overall scorecard process, but it gives us a
means for attributing a specific cost to a specific data quality problem,
which then becomes input to the final step, determination of expecta-
tions. Figure 10.2 shows distribution of impact and cost.
10.3 DETERMINING EXPECTATIONS
This last goal of assessment review poses two generic questions for each
data quality dimension measured with each associated set of impacts
and costs.
1. What quantitative level of measured quality will remove the neg-
ative impact? This represents a base requirement of expected
data quality.
ATA QUALmr REQUIREMENTS 239
Data Information
Reference Quality Chain Overall
ID Problem Location Activity Impact Cost Percentage Cost
1 Malformed
credit card Credit Card
Numbers Node 5 Processing Detection $60,000,00 40.00% $24,000.00
Contact
customer Correction 30.00% $18,000.00
Rework 30.00% $18,000.00
2 Invalid Direct
addresses Node 0 marketing Detection $150,000.00 60.00% $90,000.00
Correction 40.00% $60,000.00
Reduced Acquisition
reach overhead $600,000.00 33.00% $198,000.00
Lost
opportunity 67.00% $402,000.00
3 Incorrect Shipping
pick lists Node 7 processing Detection $24,000.00 20.00% $4,800.00
Correction 80.00% $19,200.00
Customer
service Warranty $209,000.00 20.00% $41,800.00
Spin 10.00% $20,900.00
Attrition 70.00% $146,300.00
FIGURE 10.1 Distribution of impact and distribution of cost
2. What quantitative level of measured quality will most efficiently

improve the overall cost-benefit structure? This represents a tar-
get optimal level requirement for expected data quality.
This process assigns two threshold levels for our measurements
and defines the parameters for any improvement project to be under-
taken. The first threshold is the level at which the negative impact can
be erased; the second is the threshold at which we have most efficiently
achieved the most improved data quality. Note that we can always
assign the second threshold to be 100 percent, but it may actually be
more cost effective to achieve a lower threshold and take into account
the Pareto Principle that 80 percent of the work can be achieved
with 20 percent of the effort.
Note that we must integrate the users' needs into the data quality
requirement thresholds, since it is the users that can decree whether their
expectations are being met. We therefore make use of a mechanism
called use case analysis, which, although it is usually used in the system
design process, is easily adapted to the data quality process.
10.4 USE CASE ANALYSIS
Use case analysis is a process that developed over a number of years and
is described by Ivar Jacobson in his book Object-Oriented Software
Engineering as a way to understand the nature of interaction between
users of a system and the internal requirements of that system. Accord-
ing to Jacobson, a use case model specifies the functionality of a system
as well as a description of what the system should offer from the user's
perspective. A use case model is specified with three components.
Actors, representing the roles that users play
Use cases, representing what the users do with the system
Triggers, representing events that initiate use cases
We will use of this model but altered slightly into what we could
call the "data quality view." The data quaHty view focuses on the con-
text and content of the communication interactions, so we can incorpo-
rate user expectations for the quality of information as a basis for
defining the minimum threshold of acceptance.
10.4.1 Actors
We have already seen the term actors used in this book, and that choice
of terminology was not arbitrary. In a use case analysis, an actor is a
representation of one of the prospective users of the system, describing
the different roles that a user can play.
Note that a single user may take on different roles, and each role is
represented as a different actor. Actors model anything that needs to
interact or exchange information with a system. Therefore, an actor can
represent anything external to the system (either a human or another
system) that interacts with our system. With respect to data quaUty, we
want to look at those actors that can have some effect on the levels of
data quality, so our list of actors, as reviewed in Section 10.1.1, is the
appropriate set.
In Jacobson's model, actors correspond to object classes, and the
appearance of a user taking on the role of an actor represents an instance
of one of those objects. The collection of actor descriptions forms a model
of what is external to the system. In our data quality view, we are less
concerned with object instances and more concerned with expectations
an actor will associate with information quality.
It is worthwhile to show a simple example that demonstrates a use
case analysis. Let's consider a simple candy vending machine that makes
use of a debit card. For the design of this candy machine, we expect
there to be three different actors: the purchaser, the maintenance per-
son, and the accounting system that keeps track of transactions that
have taken place on the machine.
10.4.2 Use Cases
A use case is a description of a specific way of using the system by per-

forming part of the functionaUty. It represents a course of events that
takes place when an actor interacts with the system. The collection of
use cases constitutes a specification of the system.
In our previous example, there are three actors. The purchaser
inserts the debit card into the debit card slot, selects a candy bar by
pushing a sequence of buttons, and then waits for the product to be dis-
pensed. The cost of the candy bar is deducted from the debit card, and
then it is released. The purchaser takes the candy bar, and the debit card
is ejected from the card slot. We will call this the purchase use case.
The maintenance person checks that all the advertised products are
available in the machine, that the passageways are clear, and that all
trash is removed. He or she initiates this use case, called service machine,
by opening the machine with a key. All material levels are checked and, if
necessary, changed. A third use case, generate report, is initiated by the
maintenance person to determine the activity that has taken place on the
machine.
There may be many use cases in a system or only a few. Often, many
of the use cases start out the same, and it may be difficult to identify the
difference between two use cases until it the procedure is completed. For
example, we might add another use case to the vending machine

system change parts. Here, the maintenance person initiates the use
case by opening the vending machine with a key before replacing an
inner part of the machine. Note that this new use case begins the same
way as service machine.
10.4.3 Identifying Use Cases
In standard use case analysis, we identify use cases by first examining

the set of actors and asking questions about the main tasks of each
actor. These are the kinds of questions that are asked.
Will the actor need to read or write to the system?
Will the actor need to modify information in the system?
Will the actor need to inform the system of changes in the external
world?
Does the system need to inform the actor about unexpected
changes or events in the system?
In the data quality view, the systemic use cases are basically prede-
fined. Instead, we want to identify the affiliated data quality compo-
nents of the use cases, which address both the nature of communication
within the use case and the content of those communications. We will
want to ask these kinds of questions.
What kind of information is communicated during this use case?
What constraints are there (or should there be) for data entry?
What data quality dimensions are involved in this communica-
tion?
Is any data being communicated dependent on any other known
data set?
How is data quality guaranteed?
What is the error resolution process?
What are the minimum expected thresholds for data quality?
Figure 10.2 shows a use case analysis process.
10.4.4 Triggers
In the use case model, a trigger is an event that initiates a use case. A
trigger event may occur as a result of an input data structure, an actor
Use Case Book Hotel Room:

1. The vacationer provides dates and credit
card data to the hotel clerk.
2. The hotel clerk puts in a request to the
reservation system.
3. The reservation system returns a list of
available rooms for those dates.
4. The data are sent back to the vacationer
by the hotel clerk.
5. If the dates and available rooms are
satisfactory, the vacationer agrees.
6. The hotel clerk then updates the reservation]
in the reservation system.
7. Erroneous data are propagated to the
reconciliation system, which notifies the
vacationer to resend the data.
8. A correctly inserted reservation is reported
back to the vacationer by the hotel clerk.
Data Layout 1
Modifications
to the system?
FIGURE 1 0 . 2 Uses case analysis characterizes a component of the information chain
requesting an action (for example, a report) but providing no input

data, time, or some internal database or system event. In the vending
machine example, time may be used as a trigger to initiate the mainte-
nance person's machine service routine.
The specification of the actors, the collection of use cases, and the
collection of triggers will completely model the user requirements of a
system. This model, which defines the user expectations, can be used to
derive the user's quality requirements, the user's performance require-
ments, and perhaps the user interfaces. Because the model focuses on
what the actors need from the system, there are no direct impositions
on the system implementation. The use cases enumerate the design con-
straints of the system needed to provide the required service.
10.4.5 Identifying Triggers
We begin identifying triggers by examining the operations of the system

that are not initiated by an actor. Any modification to the data set or
generated report that is not initiated by an actor is probably associated
with a trigger. For example, time-based events (for example, hard real-
time constraints, end of trading day) or capacity issues (for example,
limiting the number of transactions allowed per user) are likely related
to triggers.
In the data quality view, executing a rules engine to test a data
quality condition will also act as a trigger. The implication is that there
are embedded data quality use cases that describe the process for testing
data quality and business rules and performing some action based on
the result of the test.
10.4.6 Variants
Often it isn't clear if certain functionality is to be placed in a new use

case or if it is a variant of a previously described use case. Small differ-
ences would indicate that the new use case is a variant of the original,
whereas large differences would lead to creating a new use case. When
describing variants, the most important course of events is called the
basic course, whereas the variants (such as different options of the same
use case or errors that can occur) would be alternative courses. A use
case will normally have one basic course, and zero or more alternative
courses. In the data quality view, variants typically represent different
courses dependent on different kinds of validation errors.
10.4.7 Extensions
Extensions describe how one use case can be inserted into (extend)
another use case. In the vending machine example, the product may get
stuck during product release. The result would be that the purchaser's
debit card is debited, but the product is not properly dispensed. This
implies a new use case, product stuck, to extend get product. In this use
case, the maintenance person is alerted to fix the machine and provide a
refund for the purchaser's debit card.
Here are some other reasons for extensions.
Optional parts of use cases

Complex alternatives that seldom occur
Subcourses that are executed only in special cases
When several different use cases can be inserted into a special (or
general purpose) use case
In the data quality view, we might assign to each validation test
and error reconciliation a specific use case. These use cases can then
extend other system use cases that will make use of the validation and
reconciliation components.
10.5 ASSIGNMENT OF RESPONSIBILITY
In Chapters 2 and 3, we talked about data ownership and the different

roles that are played in the information factory. At this point in our analy-
sis, we have determined the locations where low data quality affects the
operation of the system, and we have allocated both the impacts and costs
among the different sources of low data quality. We now have a dissection
of the source of problems at level granular enough to be able to assign
responsibility for managing the information up to the standard.
Our next step is to refer back to the data ownership policy and
decide who has ultimate responsibility for each individual problem.
That person will then be tasked with defining the project that brings the
level of quality in line with the expectations. Because the tasks are well
defined in terms of the metric against which success can be achieved and
each task is a small, manageable unit, it is more Ukely that any political
ownership issues can be overcome.
In other words, we have broken the amorphous "bad data" prob-
lem into individual regions of operation, management, and ownership.
We can now make use of the data ownership policies we defined earlier
to invoke the responsibilities of maintaining a high level of quality, but
hopefully the fine granularity will ease the implementation process.
10.6 CREATING REQUIREMENTS
When defining data quality requirements, the system designer can

derive requirements from the actors, the use cases, and the triggers.
These requirements can be derived from thesefiveareas.
1. Model generation/relational inputs

2. Invariants, boundary conditions, constraints
3. Explicit quality questions
4. Report generation and queries
5. Performance
In the general sense, all of these components are important, but in
the data quality view, the most important are the middle three: invari-
ants and constraints, explicit quality questions, and (to a lesser extent)
report generation and queries.
10.6.1 Model Generation and Relational Inputs
In general system design, the systems designer begins with the use case
model to determine the information model that will encompass the
users' requirements. Typically, the process of figuring out the domain
objects begins with isolating all the nouns used in the use cases. These
nouns, along with their meanings, become a glossary for the system.
For the requirements stage, the degree of detail in describing each
object's attributes and its interactions with other objects in the system is
probably sufficient. Often, a set of objects with similar attributes can be
isolated and abstracted using a higher description. For example, the
vending machine may vend different sizes of candy bars. Each of these
objects are different kinds of products, which might imply that product
is a base class with the different sizes or flavors of candy being attrib-
utes of the product class.
10.6.2 Invariants, Boundary Conditions, Constraints
Invariants are assertions about a system that must always be true and are
used to identify error conditions. In the case of a vending machine, an
invariant may be that dispensing a product may only be performed if the
inserted debit card has sufficient credit. Boundary conditions describe the
extents of the usability of the system. In the vending machine, one bound-
ary condition is that each row can only hold 15 candy bars, so there is a
maximum number of products that may be sold between maintenance
periods. Constraints deal with issues that impede the usability of the sys-
tem. A constraint of the vending machine is that only one actor may be
served at a time.
We have already seen these issues in Chapters 7 and 8. In the gen-

eral sense, invariants and constraints focus on what we expect to be
true about the system, and so we have encompassed this as our data
quality rules. Invariants, boundary conditions, and constraints are used
to identify usability, correctness, and perhaps performance require-
ments of a system. These force the designer to augment the use cases
with error checking and handling and performance metrics.
10.6.3 Quality Issues
Quality issues also drive usability requirements. In our case, much of

the quality requirements will focus on data quality, and so we drive this
aspect the same way as discussed in Section 10.3.
Ultimately, for each important aspect of data quality, we will have
defined the minimum expected level based on a specific measurement of
a well-defined metric associated with some aspect of one of the dimen-
sions of data quality. Since we will have measurement methods in place,
we will be able to provide a running update using control charts, graph-
ing each day's progress toward meeting the specified goals.
10.6.4 Report Generation and Queries
Both of these issues focus on the retrieval and presentation of accessed

information, and, therefore, this also drives the data quality require-
ments. This aspect concentrates on both what information is to be
reported and what information is to be requested from the system.
Based on the information that is to be published out, we must incorpo-
rate measurement of the aspects of presentation, accessibility, as well as
the dimensions of data quality related to data models.
10.6.5 Performance
Despite the importance of system performance, we have not addressed this

issue to any great extent with respect to the data quality requirements. In
fact, performance is a critical concern because we do not want to put
strain on the system by inserting data quality measurement and validation.
Therefore, we must consider the means for implementing the
data quality projects. In all cases, we must impose some performance
constraints on both measurement and validation, such that there is no

significant decrease in system performance. This may imply that addi-
tional resources will be needed, and those resources will need to be inte-
grated with a production system.
10.7 THE DATA QUALITY REQUIREMENTS
We employ a use case analysis to impose our data quality requirements on

top of each interaction within the system where a decrease in data quality
can cause a relevant impact. Our data quality requirements become those
threshold measurement levels associated with our selected data quality
dimensions.
The details of each of the requirements are then compiled into a
requirements catalog that can be stored, reviewed, and queried. Each
individual requirement can be retrieved and measured to ascertain how
well the requirement is being met, which provides a single measure of
data quality within the system. The overall level of data quality of the
system can be defined as a function of the levels of each individual
requirement. Some scaling factors may be introduced based on the costs
and impacts. For example, if we have defined 10 requirements, but one
of them is tied to a problem that accounts for 60 percent of the costs of
poor data quality, we might weight it by a larger factor in accounting
for the overall system data quality score.
In terms of requirements management, it is useful to have a plat-
form from which we can both accumulate and query the details of the
definition of the requirements as well as the implementation thereof. We
can actually implement this catalog as a database table and then incor-
porate our data quality requirements as part of our system metadata.
10.7.1 A Data Quality Requirement
The format for the statement of a requirement specifies that at a specific

location in the information chain, a specific measurement must be equal
to or greater than a specified threshold. The measurement may be based
on any of the criteria for measurement described in Chapter 9, or it may
be defined by the users of the system. All measurements should be trans-
lated into a value between 0 and 1.0 to provide a percentage score for
conformance.
10.7.2 Data Quality Requirements Database
Each specification of a data quality requirement must include the fol-

lowing.
A unique identifier for the requirement
A name for the requirement
The name of the responsible party
The location in the information chain where the requirement
is applied
A reference to the dimension of data quality that is being
measured
A description of the measurement method
The measurement rule, if possible
The minimum threshold for acceptance
The optimal high threshold
The scaling factor as a percentage of the overall system data quality
We can create a database table to hold this metadata.
c r e a t e t a b l e requirements (
reqID integer,
reqName varchar(lOO),
responsible varchar(lOO),
locID integer,
dimensionID integer,
description varchar(lOOO),
reqRule varchar(1000),
minThresh float,
optThresh float.
scaling float
);
As we will see in Chapters 12 and 17, if the data quality rule sys-
tem is implemented using a rules engine, we can automate the measure-
ment and thresholding process by retrieving the rule for each
requirement, executing the measurement in the location in the informa-
tion chain, reporting whether the threshold(s) have been met, and inte-
grating the overall score based on the scaling factor.
10.7.3 The Data Quality Requirements Document
By managing our requirements in a metadata database, we provide the

means to automatically generate a requirements document that can be
reviewed by the end users for correctness and compliance. The require-
ments document will display all the pertinent information described in
the previous section, but the benefits of maintaining the requirements as
content include the following.
The statement of the requirements is easy to augment or modify,
since it can be managed via a simple user interface.
Since the requirements are in a table form, subselection of the
requirements based on specific attributes of the requirements table
(such as data quality dimension, or responsible party) yield spe-
cific data about the requirements.
Integrating the requirements table with a history table indicating
daily measurements provides a platform to measure improvement
over time.
As opposed to a static definition in a print document, the require-
ments specification can be kept up to date in a more manageable
way, and a new print version can be regenerated any time the
requirements change.
10.8 SUMMARY
In this chapter, we looked at how to use the analytical information col-

lected via the map of the information chain, the construction of the data
quality scorecard, and the current state assessment to provide input to the
data quality requirements specification. Our goals in defining require-
ments include the distribution of the weight of the impacts and costs asso-
ciated with low data quality and the definition of a pair of threshold
values for each measurement we take.
The first threshold is the minimum acceptance threshold, below
which we cannot decree the information to meet our criteria for fitness
for use. The second threshold represents an optimal level of data qual-
ity, defined as the point at which the most efficient expenditure of
resources achieves the greatest improvement. Together we can use these
thresholds to determine data quality scores for individual requirements,
as well as customize an overall system data quality score.
We looked at how a system design technique called use case analy-

sis can be used for the assignation of requirements based on the applica-
tion services provided to the user. We can build on top of the use case
analysis to see at which points the more significant issues with respect
to data quality emerge.
We examined the definition of the requirements and the manage-
ment of those requirements in a well-defined database. By using a data-
base table, we can manage the requirements themselves as content, as
well as provide a platform for automatically generating requirements
documents and reports of requirements conformance, indexed by data
quality dimension, by responsible party, by location in the information
chain, or any other stored attribute.
11
METADATA, GUIDELINES, AND POLICY
In Chapters 7, 8, and 10, we discussed maintaining a repository of infor-

mation about the data within the system. We used the term metadata to
refer to this repository. In this chapter, we explore metadata in more
detail, since a major component of a knowledge management program is
the maintenance of as much information as possible about an organiza-
tion's data resource.
The simplest definition of metadata is "data about data." In prac-
tice, metadata is a repository of information that enables a user to inde-
pendently learn and understand what information is available and
extractable from a given data resource. Because of the many different
ways that information is used, this reflects the many different ways
metadata can be represented and used. In general, metadata represents
all structural and definition aspects of any set of data, whether from a
technical or a business perspective.
On the technical side, metadata is used for the design, implementa-
tion, development, maintenance, and ongoing management of data. On
the business side, metadata encompasses all contextual meaning associ-
ated with a data set, covering definition of all objects within the system,
the relations between the objects, the derivation of the data, the change
history, data quality rules, and business rules associated with the use
and exploitation of the data.
In today's world, where the number and magnitude of data sets
grow exponentially with the ability to manage them, without a frame-
work for encompassing information about the data, we would get lost
in a sea of fragments. Different kinds of information platforms depend
on different kinds of metadata. Clearly, the metadata needs of an online
253
transaction processing system are different from those of analytical sys-

tems or a Web environment. Yet, industry players have recognized the
need for a common ground for defining and communicating metadata.
The Meta Data Coalition (MDC) is a consortium of vendors and end-
users that has agreed to cooperate on the definition of a standard form
for defining, sharing, and managing metadata.
In this chapter, we summarize some of the best practices from the
data industry to provide the baseline version of a metadata framework.
We will rely on some of the ideas present in the MDC's Open Informa-
tion Model (OIM), an evolving standard for the definition, communica-
tion, and management of metadata, and we will incorporate other ideas
found in the literature. Our goal is to provide a baseline guide for build-
ing a metadata repository, and this list should by no means be considered
complete.
We will explore how metadata relates to data quality and how meta-
data can be used as a management tool for driving enterprise information
policy. Since the kinds of data domains, mappings, and rules discussed in
Chapters 7 and 8 can be viewed as enterprise reference data that doubles
as metadata, we will discuss how to manage domains, mappings, and
rules. We will also look at how a well-defined metadata repository can be
come a useful enterprise information publication and subscription tool
when used as the basis for information browsing.
11.1 GENERIC ELEMENTS
In any metadata framework, there are some generic elements that can
tag most data or metadata elements in the system. This section enumer-
ates some of the generic metadata elements that would be incorporated
into the repository.
11.1.1 Contact Information
This is all the information needed to contact a particular party, includ-

ing the following.
Name
Title
METADATA. GUIDELINES, AND POLICY 255
Location (see Section 11.1.5)

E-mail address
Telephone numbers, including the type of telephone number
(office, mobile, home, pager, FAX, home office, etc.)
11.1.2 Description
This is a text description of any entity. This might also incorporate ref-
erences to more advanced descriptive objects, such as text documents,
graphical pictures, spreadsheets, or URLs.
11.1.3 Iconic Representation
Is there any kind of iconic representation for an entity.'^ If so, is there

more than one? This kind of metadata describes any types of icons, as
well as the visual representation of the icon based on the state of the
represented object. This incorporates the graphical representation and
specific picturefiles,colors, conditions, and rules for display.
For example, an icon can be associated with the access attribute of
a data field. If the specified user has read-only access, the color of the
icon might be red to indicate that the user may not write to the field. If
the specified user has write access, the same icon might be green.
11.1.4 Keywords
Keywords are used for classification of objects. Associated with any

entity is a set of keywords that can be used for indexing purposes when
building a browsable metadata catalog.
11.1.5 Location
A location represents a physical location, such as a street address, an

office building, a room number, as well as any attribution such as
whether the location is home or office. A location attributes contact
information as previously described.
11.1.6 Author/Owner/Responsible
These are ail contact information references to parties. The author is

the original major creator of the entity being described. The owner is
the person that manages the described entity. Sometimes, there is an
additional party who is responsible for the described entity but is not
the owner.
11.1.7 Help Source
This represents a resource that provides help for the described entity, such
as a URL, a README file, or the phone number of a support person.
11.1.8 Version
This is a way to keep track of user-defined version information, such as

major version, minor version, and revision numbers associated with a
particular entity.
11.1.9 Handler
This describes some kind of object that serves as a handler for the entity,
such as an application component that handles the user interface for
directing operations on the entity.
11.1.10 Menus
These are the menus associated with the user interface for a described
entity.
11.2 DATA TYPES AND DOMAINS
In all types of computer systems, data objects are defined as belonging

to a preassigned data type. When information is passed from one sys-
tem to another, it is good to use some base standard representation of
METADATA. GUIDELINES. AND POLICY 257
the types that are available among all platforms. In our metadata repos-
itory, we keep track of all the types that are available within the envi-
ronment, as well as the ways that types are defined and used.
11.2.1 Alias
An alias is a name used to describe a type that is known by another

name. In C/C++, the typedef denotes a type alias. We associate as meta-
data all type aliases in the enterprise essentially as a mapping (as
described in Chapter 8).
11.2.2 Enumeration
Any enumerated list of data values that is used as a type is kept as meta-
data. Enumerated types, such as our domains from Chapter 8, are rep-
resented using a base type and a list of values.
11.2.3 Intrinsic
Any defined type that is intrinsic to the data model should be denoted as
such, and the information about both the size of objects and the physi-
cal layout of values of that type should be maintained as metadata. For
example, decimal values may be maintained in a number of different
ways on different platforms, although the presentation of values of type
decimal may be identical across all those platforms. In this case, the
exact representation should be denoted in case the data need to be
migrated to other platforms.
11.2.4 Namespace
This is the namespace associated with the object types or origination

context in which an entity exists, such as an application name and ver-
sion number in which the entity was created. This is used to guard
against incompatibilities that crop up between objects created in differ-
ent contexts. We have already talked about namespaces in reference to
rule sets. A single collection of rules accompanying a data set must be
accumulated under a single "rule set," the name of which defines a rules
namespace.
11.2.5 Object Type Mapping
This represents a mapping from object types in one namespace to object

types in another. This gives the user the opportunity to see the best
matches between object types that belong to different namespaces, so
that any kind of comparison or migration between objects of the differ-
ent types can be planned accordingly.
11.2.6 Scalers
Atomic data types used in a system as base types are scalar types. Exam-
ples include integers, floating point numbers, and strings. We enumerate
a set of base intrinsic scalar types in Section 11.2.11.
11.2.7 Structure
A structure is the definition of a structured data type used within the

system.
11.2.8 Union
Similarly, a union is the representation of the union data type definition,

such as in C or C++.
11.2.9 Array
An array is a linear collection of objects of the same type.
11.2.10 Collections
With objects that can be bundled using collections, the collection meta-
data maintains all information about the maximum size of a collection.
the way that collection members are accessed, whether there is any
inherent ordering to members in the collection, and the operations that
are permissible on the collection, such as insert, delete, sort, and so forth.
If there is some inherent ordering, there must also be descriptions of
equality of objects and an ordering rule (that is, a "less-than" operation).
11.2.11 Base Types
Here is a list of base types that are typically used. Any metadata reposi-
tory should keep track of which of these types are valid within each
enterprise system.
Binary A large unbounded binary object, such as a graphic
image, or a large text memo
Boolean TRUE or FALSE
Character The metadata repository should log which represen-
tation is used for characters, such as ASCII or UNICODE
Date A date that does not include a time stamp
Datetime A datestamp that does include a timestamp
Decimal The exact decimal value representation, as opposed to
float, which is approximate
Double Double precision floating point
Float Single precision floating point
Integer Standard integer type
Small integer Small integer type
Long integer Extended size integer type
Long long integer Larger-sized extended integer type
Numeric Fixed-size decimal representation. This should be
attributed by the number of numeric digits that can be stored and
the number of digits to the left of the decimal point.
Pointer A reference to another data object
String A sequence of characters that must be attributed with the
length as well as the character type
Time A timestamp data type
TimePrecision The precision of the counting period associated
with a timestamp type
Void The standard intrinsic void data type (as in C/C++)
11.3 SCHEMA METADATA
In our data environment, the actual layout of information is critical meta-

data as well as any relation, access, and usage data. We characterize all of
this metadata as schema metadata metadata about the data schema. In
each of these sections, we use the metadata elements described in the pre-
vious sections.
11.3.1 Catalog
A catalog is a collection of schemas, and combined, a catalog of

schemas can be used for browsing through the enterprise information
metadata resource.
11.3.2 Connections
In an operational system, there must be a means to describe the ways

that users can interact with the database system. Client references to
databases are called connections, and for each connection we must
maintain the following as metadata.
Names used to establish the connection
Connection string (may include user name and password,
although typically this may include other information)
The name of the database used in the connection
The name of the data source used by the connection (DSN)
Indication of read-only status
Whether the connection can be shared
Connection timeout (the time in which initialization takes place; if
this time is exceeded, a timeout may have occurred)
11.3.3 Tables
For each table in the system, we want to maintain the following.

Table name
Description of what the table models
Physical location of the table (which server does it live on?)
Size of the table and growth statistics (and upper size limit, if
necessary)
Source of the data that is input into the table
Table update history, including date of last refresh and results of
last updates
Primary key
Foreign keys
Referential integrity constraints
Cross-columnar data quality assertions
Functional dependences
Other intratable and cross-tabular data quality rules
Data quality requirements for the table
11.3.4 Attributes/Columns
For each table, we will also maintain a dictionary of the attributes or

columns of the table. For each column, we want to maintain the following.
Column name
Description of what is in the columns
Business definitions associated with the column
Data type
Column size
Whether the column is searchable
Domain membership
Source of the data values that populate the column
Is this an auto-generated unique key?
Null value rules
Value restriction rules
Data transformation rules
Consistency rules
Data quality requirements for the attribute
11.3.5 Load Programs
This is a list of the programs used to load data into the tables or that
feed data into the tables. We maintain this information as metadata.
The name of the program
The version of the program
Description of what the program does

The source code
Revision history
Dependency information, such as what other programs must pre-
cede this program in execution
The author of the program
The current owner of the program
11.3.6 Views
Views are table-Hke representations of data joins that are not stored as
tables. For each view, we want to maintain the following.
The name of the view
The owner of the view
The source tables used in the view
The attributes used in the view
Whether updates may be made through the view
11.3.7 Queries
Queries represent sets of records extracted from database tables. We

can maintain this metadata for queries.
A name for the query
The columns involved in the query
The text of the SQL for the query
11.3.8 Joins
We can represent joins as metadata as well. We must maintain the fol-

lowing.
The name for the join
The join key used for comparison from each table
The tables that participate in the join
The columns that participate in the join from each table
11.3.9 Transformations
In any data extraction, transformation, and loading scheme, we can

maintain metadata about the transformations that are performed. We
already have a format for defining transformations using our rules from
Chapter 8, but for clarification, each transformation can be represented
Uke this.
The collection of source data items (including table/column
references)
The destination value
The transformation that is applied
The person responsible for this transformation
The description of the transformation
The business aspect of the transformation
11.3.10 Data Sources
In all data sets, data come from different original sources. While some
originates from user applications, other originates from alternate
sources. For supplied data, we want to maintain the following.
The name of the data package
The name of the data supplier
The name of the person responsible for supplying the data (from
the supplier side)
The name of the person responsible for accepting the data (on the
organization side)
The expected size of the data
The time the data is supposed to be delivered
Which tables and attributes are populated with this data
The name of any data transformation programs for this data
The name of the load programs for this data
External data quality requirements associated with provision
agreement
11.3.11 Triggers
Triggers are rules that are fired when a particular event occurs. These
events may be table update, insert, or deletes, and in the metadata we
want to maintain the following.
Name of the trigger
The author of the trigger
The owner of the trigger
Whether the trigger fires on an update, insert, or delete
Whether the trigger fires before or after the event
The trigger frequency (once per row or once per table)
The statements associated with the trigger
Which columns are associated with the trigger
11.3.12 Stored Procedures
Stored procedures are collections of SQL statements that may be exe-

cuted (as a program) by database users. We want to maintain metadata
for the following.
The name of the stored procedure
The author of the stored procedure
The version history
The description of what the stored procedure does
The SQL statement sequence (or at least a pointer to the
sequence)
11.3.13 Indexes
We can manage the collection of indexes as metadata. We will keep

track of the following.
The table or view on which the index is made
The columns associated with the index
Whether nulls are allowed
Whether the index needs to be rebuilt or is updated automatically
Whether the index is sorted
11.4 USE AND SUMMARIZATION
Usage statistics for a data set are useful for system resource manage-
ment, system optimization, and for business reasons (answering ques-
tions such as "Which data set generates the most interest among our
customers?").
11.4.1 Time Constraints
If there are any time constraints, such as in processing or delivery of

data, these should be documented as metadata.
11.4.2 Users
We want to maintain this metadata associated with users in the com-

munity.
Name of user
Location
Telephone numbers
E-mail address
List of data objects to which this user has access rights
11.4.3 Access Rights
For each touchable entity in the system, we want to keep track of which
users in the community have which levels of access.
User name
Data object name
Access rights
This metadata is used both for enumerating the access rights a user
has for a particular data set, as well as enumerating which users have
rights for a given data set.
11.4.4 Aggregations
In a data warehouse, aggregations are used for precalculating informa-

tion for performance reasons. Aggregation metadata keeps track of
what dimensions make up the aggregation.
11.4.5 Reports
End users that expect summarization of information stored in a data-

base, as well as usage or behavior characteristics, may request reports.
A report is a formatted view of summarized data. For each report, we
will want to maintain the following.
The name of the report
The creator of the report
The consumers of the report
The time the report is available
The tables and columns that are source data for the report
The data transformations or calculations used in making the
report
The format of the report
The means by which the report is distributed
11.5 HISTORICAL
11.5.1 History
We will want to keep track of different kinds of history. Thefirstkind is

versioning associated with the data schema itself, and the second kind is
usage history.
11.5.2 Enumeration of Readers
For all data objects, we want to keep track of who is reading the data.
The user name
The program name
The data object being read
The frequency of object reads
11.5.3 Enumeration of Writers
For all data objects, we want to keep track of who is writing the data.
The user name
The program name
The data object being written
The frequency of object writing
11.5.4 Last Update
It is also useful to keep track of which user or program last updated a

data object.
11.5.5 Change in Representation, Precision, Derivation
Any time a change is made to the data environment that involves a

change to the data model (representation), a change to the precision of
the information (such as changing from a float type to a double type),
or a change is derivation (such as modifying the inputs that populate a
datafield),that change should be logged in a change history. The infor-
mation that is logged may include:
Modification name
Date of change
The object that was modified
The reason for the modification
Was this a change in representation, precision, or derivation.^
If this was a change in representation, what was the old represen-
tation and the new representation?
If this is a change in precision, what is the old precision and what
is the new precision?
If this is a change in derivation, what is the replaced source and
derivation process and what is the new source and derivation
process?
The person responsible for effecting the change
Transformations associated with the change
This last item is quite important. It provides a way to bridge histor-
ical aggregation and summarization across time from the point before
the change to after the change.
11.6 MANAGING DATA DOMAINS
In Chapter 7 we discussed the definition of data domains. Since it is

probable that the same domains are being used in different parts of the
enterprise, maintaining a single source directory of those domains is
beneficial when it comes to being able to browse a domain catalog. If
we want to manage our data domains as reference metadata, we can do
this using three data tables.
The initial component of our domain metadata catalogs the names
of all defined domains used within the system as well as the domain
types. Therefore, the first table is a domain reference table, which con-
tains the name of the data domain, the class of the domain (base data
type, enumerated or rule-based), a reference to the base domain or data
type on which it is based, a text description of the domain, and the
source of the data that populates this domain, along with an assigned
identifier for the domain. We will include in this table a reference to the
base data types, and the dType component refers to another domain ref-
erence within the table. Here is one way to define the table.
create table domainref (
name varchar(30),
dClass char(l),
dType integer,
description varchar(1024),
source varcharC512),
domainid integer
);
The values can all be stored in a single table, referenced by a
domain identifier. The component domainiD refers to one of the named
domains in the domainref table. In this case, we arbitrarily limit the
size of the values to 128 characters or fewer.
create table domainvals (
domainlD integer,
value varchar(128)
);
Finally, we represent our rules-based domains using records that
consist of rule statements.
create table domainrules (
domainid integer,
rule varchar(1024)
);
One other table is useful in the metadata repository, although it is

not used for representing the domain. Rather, just as we maintain usage
data about our data tables, we will also maintain usage data about our
domains, basically keeping track of which data table attributes use
defined domains (see Figure 11.1). We will look at the representation of
rules in Section 11.8.
11.7 MANAGING DOMAIN MAPPINGS
Similar to representing data domains as metadata, we will also like to

accumulate domain mappings as enterprise reference metadata. Even
more so for mappings, since they can represent embedded functionality
that may have ubiquitous use within the enterprise, it is desirable to make
use of database tables to represent mappings between data domains.
Just as with domains, we can represent mappings using three data
tables, where the first table is used as a catalog. The mapping reference
table contains the name of the domain mapping, the domain id of the
source domain, the domain id of the target domain, a text description
of the mapping, and the source of the data that populates this mapping,
along with an assigned identifier for the mapping.
Domai nRef
domainID
name
dClass
dType domai nVals
description 1
domainID (FK) 1
source ^
w value 1
i1
domai nRi les
domainID (FK)
rule
FIGURE 1 1 . 1 Domains database

create table mappingref (

name varchar(30),
sourcedomain integer,
targetdomain integer,
description varcharC1024),
source varchar(512),
mappingid integer
);
The value pairs can all be stored in a single table, referenced by
mapping identifier. In this case, we arbitrarily limit the size of the values
to 128 characters or fewer.
create table mappingpairs (
mappingid integer,
sourcevalue varchar(128),
targetvalue varchar(128)
);
Finally, we represent our rules-based mappings using records that
create table mappingrules (
mappingid integer,
rule varchar(1024)
);
By absorbing all the mappings into a centralized repository, we can
provide users with a source for browsing the mappings that are used
within the system, as well as allowing users the chance to explore the
meanings of these mappings based on their use. Therefore, we will also
maintain usage metadata on which attribute pairs conform to the rela-
tionship defined by known mappings (see Figure 11.2).
11.8 MANAGING RULES
Managing rules is slightly different from managing the domains and

mappings. With the latter, most frequently we are managing enumer-
ated data sets, and these are straightforward to represent using data
tables. The rules we describe, though, are flexible enough that they defy
standard data representation. The reason for this is in the nested nature
of rule expressions. Many of our rules contain conditions or actions,
which may be composed of smaller pieces, each of which is a valid con-
METADATA. GUIDEUNES. AND POLICY 271
Mappings
1 1
mappingID 1
Name 1 mappingPairs
r A
+ .
sourcedomain 1 mappingID (FK) I
targetdomain 1 "^"^H
1 1. . ^
targetvalue I
description 1
sourcevalue 1
source 1
mappingRules
FIGURE 1 1 . 2 Mappings database
dition or action. Nested structures are not completely suited to storage

in a database, mostly because the complete structure cannot be
extracted with a single query. For example, consider this rule.
I f (employees.salary < 20000) AND (employee.status <>
"fulltime") then exempt (healthcarrier);
The condition in this rule is a conjunction between two simpler
conditions, joined with the AND operator. In this case, the condition
has two parts, each of which has its own operands. Naively, to repre-
sent this condition, we would need to keep track of four operands and
three operators. The problem with representing this in a database table
is that once a table is defined, we do not add additional columns to the

table to account for each rule's essentially unlimited number of compo-
nents. If we were to add more columns, most of them would be empty,
which is a tremendous waste of space. We simply cannot represent all
the pieces of this rule's condition in a single row of a database.
11.8.1 Approaches to Managing Rules
In fact, all operations can be represented using a tree structure, where the
root of the tree represents the operator (in this case, "AND") and the
leaves of the tree represent the operands (in this case (employees.salary <
20000) and (employee.status <> "fuUtime")). We can apply this recur-
sively, so that each of those two operands is also represented using a tree.
This is shown in Figure 11.3.
There are two alternatives for maintaining the rules as metadata,
both of which rely on programmatic means for implementation. The
first alternative is to use a data table for embedding nested structures. We
can create a single table to represent expressions, where each expression
contains an operator (like "AND" or "<>") and afinitemaximum num-
ber of operands. For argument's sake, let's constrain it to two operands.
FIGURE 1 1 . 3 Metadata operations

create table expressions (

expressionID integer,
isRule logical,
operator integer,
operandi integer,
operand2 integer,
attrName varchar(lOO)
);
The isRule attribute is set to true if this record represents the root
of a rule tree. We will enumerate operators in a separate table, and the
operator attribute refers to that table. Operands may take different
forms; they may be attribute names, or references to other expressions.
Therefore, we will have our expressions refer back into the expression
table and allow for one attribute to be a character string representing
attribute names. Here is how we could represent our previous example
condition, although for clarity we will make the operator names explicit.
Expression ID Operator Operand 1 Operand 2 attr Name
0 NAME Employecs.salary
1 CONSTANT 20000
2 LESSTHAN 0 1
3 NAME Employees.status
4 CONSTANT Fulltime
5 NOTEQUAL 3 4
6 AND 2 5
By careful inspection, we will see that all of our rule components

can be represented in this format. Even lists of attribute names can be
implemented in this binary format by using a LIST operator and mak-
ing the second operand of a LIST operator be either an attribute name
or another LIST node.
There are three issues with this representation, however. Thefirstis
that a program is needed to transform rules from a linear format (such
as the syntax we used in Chapter 8) into the nested tree format for the
database. The second is the difficulty in extraction; a stored procedure
or a dedicated application is needed to cursor through the table, recre-
ating the rule. The third issue centers on modifying the rules. When a
user wants to change a rule, this requires a lot of maneuvering within
the table.
The rule must be recreated from the representation in the table, but
if the rule is truly modified, then those tuples that compose the rule
within the table must be removed from the table, and the new version of
the rule must be inserted into the table as if it were a new rule. This
means that there is little opportunity for reuse, and rule editing is a very
read/write-intensive operation.
11.8.2 Alternate Rule Management
A different approach to rule management is to maintain the text repre-

sentation of the original rule within the database table. In this
approach, the responsibility for managing the form of each rule is
removed from the database and placed in application code.
A formal definition of the syntax of all rules must be defined, and a
rule parser can be written to validate that each rule conforms to the cor-
rect syntax. The application would also incorporate semantic analysis
to make sure that rules make sense contextually. This way, we can
maintain each rule as text within a single field of a database table.
create table rules (
rulelD integer,
ruleText varchar(1000)
);
Using this approach, we have lost the ability to directly query the
rule table to search for expression associated with a specific table
attribute (except through substring searches). On the other hand, this
approach greatly simplifies rule insertion, rule editing, and rule extrac-
tion. In addition, the benefit of using a rule parser for validation of the
format is that the same parser can be used for transforming the rule
specification back into a format suitable for in-process execution or for
incorporation into a rules engine.
11.8.3 Rule Sets
One more table is needed to keep track of which rules are associated
with specific rule sets, since different users may have different sets of
rules, even for the same data. Managing rule sets requires defining a
rule set name, associating it with a unique identifier, and having a sepa-
METADATA, GUIDELINES. AND POLICY 275
rate table linking the rule set id with the identifiers associated with each
defined rule.
Create table rulesets (
RuleSetID integer,
RuleSetName varcharClOO)
);
create table rulesetCollections (
RuleSetID integer,
RulelD integer
);
If we are using the first approach, the rule ID is the identifier asso-
ciated with the root of the expression tree for that rule. If we are using
the second approach, we use rule ID.
11.9 METADATA BROWSING
If we have competently integrated all metadata into a centralized repos-

itory, we now have a platform from which information about the enter-
prise can be communicated with all appropriate users within the
organization. Users can query the domain tables and be provided with a
list of domains, the description of each domain, and the other tables
that use that domain. The same is true for mappings.
Data quality and business rules can be queried and reviewed, and
domains, mappings, and rules can be reused. All of this can be easily
accomplished with a metadata browser a front-end tool that enables
the dissemination of reference data and metadata to systems designers.
Through the noncontrolled sharing of this information, the organiza-
tion will organically become more intelligent and consistent, since de
facto standards of usage will emerge through "natural selection." Valu-
able data domains, mappings, and rules will be accepted by growing
groups of users, while ineffective metadata objects will wither and die
from lack of use.
This also means that information usage policies can be effectively
established through the definition of the accepted reference data compo-
nents and controlled browsing. By selectively allowing users to subscribe
to a data domain or a mapping, or by limiting the kinds of schema meta-
data available for review, senior information managers can influence
many aspects of information system design, including the way that data-
base tables are defined and the ways that information is extracted and
presented to users. By controlling the flow of metadata, the managers of
a data resource can control the way that knowledge is integrated
throughout the enterprise. Figure 11.4 shows metadata browsing.
11.10 METADATA AS A DRIVER OF POLICY
The use of a metadata repository does not guarantee that the systems in
use in the enterprise will subscribe to the philosophy of centralized
knowledge management and control. Unfortunately, the existence of
useful tools does not necessarily induce people to use them.
As a more practical matter, the instituting of metadata use can
enhance an organization's overall ability to collect knowledge and reuse
data. But without an enforceable policy behind the decision to incorpo-
rate a metadata framework, the benefits will not be achieved. In tan-
dem, information policy drives the use of metadata, but the use of
metadata will also drive information policy.
Creating a central core of reference data and metadata requires an
organizational commitment to cooperation. As with the data owner-
ship policy, the information management policy must specify the ways
that data and metadata are shared within the company.
One more aspect of information policy embedded in the metadata
question is that of enterprise-wide accessibility. Between issues of pri-
vacy and security and knowledge sharing, there is some middle ground
that must indicate how metadata is shared and when accessibility con-
straints are invoked. This is a significant area of policy that must be
effectively legislated within the organization.
11.11 SUMMARY
In this chapter, we explored the notion of metadata, starting with the

basics about generic elements and moving through data types and
domains, schema, and use and summarization. We then looked at the
management of the sort of reference data and metadata that we have
introduced in earlier chapters, namely data domains, mappings
between domains, and data quality rules. Last, we looked at how the
publication of metadata throughout the organization can improve
knowledge management and data quality overall and that information
policies are integrated closely with metadata.
O
i-i
2 A
UJ
\i D
12
RULE-BASED DATA QUALITY
In Chapter 8, we discussed the definition of a formal system for describ-

ing data quality and business rules. For an organization, there is a sig-
nificant advantage to consolidating its business and data quality rules
and using an automated rules system to implement those rules.
Consolidating business and data quality rules is a way to capture and
control strategic knowledge. In executable systems, this knowledge is
most frequently incorporated into program logic, which is both hard to
access and to control. Capturing and controlling the embedded knowl-
edge requires that it be moved from the opaque representation of the
computer program to a form that business users can manage and control.
Over the past 20 years, rule-based system technology has evolved
from an assortment of niche products and languages into a set of inte-
grated approaches and tools. Many of today's products are capable of
supporting the knowledge consolidation process. Because a collection
of rules is easier to browse and understand than a complex piece of C++
code, rule-based systems provide a reasonable approach to software
flexibility and componentization.
In this chapter, we investigate the ideas behind rule-based systems,
providing an introduction to business and data quality rules and the
ways that rule-based systems make use of these rules. We also look at
some of the "insides" of a rules engine, as well as some criteria for eval-
uating rule-based systems.
279
12.1 RULE BASICS
What is a rule? Back in Chapter 8, we defined a way to describe data

quaUty rules, but we did not focus on the specifics of rule semantics. In
reaUty, the formaUsm described in Chapter 8 carries to a large extent
what is called "syntactic sugar" an increase in the kinds of words
used to describe rules to make them easier to read and understand.
In general, though, if rules are going to be used as a descriptive
means for encapsulating operational business flows, it is important to
be familiar with the lingo used in the definitions. In business systems,
rules form the programmed representation of business policies and
practices.
Here are some examples of rules.
If the request is for more than $5,000, then a senior manager must
sign the request form.
If the customer's daily withdrawal limit is not exceeded, dispense
the requested cash amount.
If a reservation cannot be found for this ticket, escort the ticket
holder to Customer Service.
No customer may request a purchase that is greater than an
assigned credit limit.
For all intents and purposes, a rule contains a condition and an
action. More precisely, a rule is a statement that asserts some truth
about the system, along with an optional action to be performed if the
assertion is not true. Alternately, a rule is a condition followed by an
action to be performed if the condition is true. Note that the previous
four examples meet these definitions, and all the rules described in
Chapter 8 can be restated in these simple forms. Rules stated in either
way can be transformed into the other form. For convenience we will
use the condition followed by action form.
Actions may consist of modifying the environment (which in turn
may turn on other conditions or assertion violations) or restricting some
modification to the environment, such as disallowing a transaction. Con-
ditions are evaluated when some trigger event occurs. When a condition
is evaluated to be true, the action is taken, and that rule is said to have
been fired.
RULE-BASED DATA QUALITY 281
12.2 WHAT IS A BUSINESS RULE?
A business rule is a statement that describes some structural aspect of a

business, defines some relationship between entities in a business, or con-
trols or influences the behavior of the business. As Barbara von Halle
explains in an article by Ellen Gottesdiener in the March 1997 Applica-
tion Development Trends, business rules take the following forms.
1. Declarative and not procedural rules stated as assertions, not
as program logic.
2. Atomic when each rule refers to one and only one issue.
3. Expressed in a well-formed language there is a formalism for
expressing rules.
4. Distinct independent constructs each rule refers to a specific
business notion.
5. Business oriented the rule refers to the way business is done
and is not bogged down as the technical implementation of busi-
ness logic.
6. Business owned ownership of the rules lies with the business
stakeholders, not with the implementers of that business logic.
Even though these attributes are desirable in a business rules sys-
tem, separating the management of business rules from business rule
implementation does not necessarily remove the element of program-
ming from the instantiation of a business rule system. Rather, the state-
ment of and the execution of business policies as just described need not
be tightly coupled with the implementation of those policies.
As an added benefit, the separation of business rules from both the
data on which the rules operate allows for easier implementation of
multitiered client/server applications. In other words, business opera-
tions can be divorced from the client side by implementing the execu-
tion of business rules at an application service level. Modifications to
the environment, typically represented as an enterprise database, are
abstracted from the business operation as well, since the rules transcend
the data values themselves.
Business rules are a manifestation of a rule-based system that is
bounded specifically by descriptions of business policies and relation-
ships. The GUIDE Business Rules Project attempted to identify and
articulate the rules that define the operation of a business. As described
in the final report of November 1995, a business rule must be one of the
following.
Structural assertion A fact that describes some aspect of the

enterprise
Action assertion A constraint or condition that Umits or con-
trols the actions of the enterprise
Derivation A statement of knowledge derived from other
knowledge in the enterprise
The GUIDE project drives the expression of rules in a formal way
that can be superimposed on both a rules programming system and an
entity-relationship form while maintaining a business focus. By doing
this, the project members abstract the essence of business-oriented rules
but allow for flexibility in the implementation of those rules while sug-
gesting ways to archive those rules as content.
Since the specification of business rules can be transformed into a
specification for any reasonably established rule-based system, we will
concentrate on the ideas of a rule-based system and return to business
rules later.
12.3 DATA QUALITY RULES ARE BUSINESS RULES

(AND VICE VERSA)
We already saw data quality rules in Chapter 8. We can see that the
rules described there fall into the definitions described in Section 12.2.
Our data quality rules are structural assertions (such as domain and
mapping definitions), action assertions (such as our transformation
rules and domain assignment rules), or knowledge derivation rules
(such as our domain mapping and assignment rules or our derivation
rules). In any of these cases, our data quality rules match the specifica-
tions listed in Section 12.2.
1. Our data quality rules are declarative. The fact that we attempt
to move data quality rules from the executable program to the
world of content proves that our data quality rules are declara-
tive and not procedural.
2. Each specific data quality rule applies to one specific operational
or declarative assertion, demonstrating atomicity.
3. We have defined a well-formed semantic for specifying data qual-
ity rules, yielding a well-formed specification language.
4. Each rule in a system exists in its own context and can be viewed,
modified, or deleted without affecting any other rule in the set.
5. Since we can view data quality as the "oil" in the machinery of

operations, the data quality rule is business oriented.
6. In a perfect world, the ownership of the data quality rule lies
with the data owner, who typically will be the major business
stakeholder in the data.
Thus, in a not too roundabout way, data quality rules are business
rules. An even more daring comment would be that a majority of busi-
ness rules could actually be characterized as data quality rules. Data
quality rules are specifications on the activity associated with allowing
information to pass from one platform to another, whether that means
from data suppliers to the enterprise, between clients inside the enter-
prise, or between the enterprise to their clients or customers.
12.4 WHAT IS A RULE-BASED SYSTEM?
A rule-based system is a set of rules that describes business policies to be

executed in process. A rule-based system provides a mechanism for cre-
ating and editing a rule base, along with an engine that tests conditions
or triggers and fires off actions.
12.4.1 Rules Languages
A rules language is a framework for describing sets of rules. Because

our description of a rule includes both a condition and an action, any
standard programming language that contains IF-THEN-ELSE con-
structs like C and C++ can be used as a rules language. However,
because the standard procedural programming languages impose a
sequential order of execution, their use as rules languages is limited.
These languages are limited because there is nondeterminism asso-
ciated with events and assertions in the modeled world. A procedural
language will impose an artificial dependence on the sequence of events
and can lead to unexpected results if those events take place in different
orders. Instead, rules languages, by virtue of their declarative syntax,
provide a way to express a set of rules without imparting any order or
execution.
12.4.2 Rules Engines
A rules engine is an application that takes as input a set of rules, creates

a framework for executing those rules, and acts as a monitor to a sys-
tem that must behave in conjunction with those rules. Rules engines
work together with a rules language. A rules engine always acts
abstractly. In other words, the operation of a rules engine is the same no
matter what the rule base specifies. It is through this mechanism that the
rules are separated from the implementation.
Complex rules engines allow for multiple concurrent evaluations
of rules as well as parallel triggered execution. Other kinds of rules
engines perform complex mathematical reductions to verify the contin-
ued truth values of the system. Simpler rules engines will (perhaps in a
nondeterministic way) select a rule for evaluation and possibly execute
an action if the condition evaluates to true (see Figure 12.1).
12.5 ADVANTAGES OF THE RULE-BASED APPROACH
12.5.1 Augmented Capabilities
A major difference between procedural programming languages (such

as C and C++) and rules languages is the constraints on order of execu-
tion. Rules engines operate in stages: evaluate the environment, check
for triggered conditions, and execute corresponding actions. Only those
rules that are affected by inputs or triggers at each execution stage are
going to be executed. In procedural languages, execution of a sequence
of rules is predetermined by the programmer. This not only forces many
"rules" (as embedded in "if-then" statements) to be unnecessarily exe-
cuted, it opens up the possibility for incorrect execution if the nesting of
rules is not appropriate.
12.5.2 Efficiency in Policy Automation
Rule-based programming provides a more efficient manner for

automating business policies. One reason is that the formality of de-
scription encourages a more precise thought process that in turn is more
thorough. This provides better "up front" definition of the business
process.
Here are some more reasons.
The rules execution

environment provides for
- rule evaluation
- triggered execution
- rule base validation
The application hooks into

the rules engine, which
incorporates the application's
rule set during execution.
FIGURE 1 2 . 1 Rules engine
Separating the implementation of complex business logic from

nonexperts will help narrow the time needed to complete an
application.
The time needed to test will be reduced.
The rules engine manages the mechanics of the application, so the
amount of coding is decreased.
12.5.3 Opportunities for Reuse
Data quality and business rules reflect the ongoing operations of a busi-
ness. In any large environment, there are many situations where the
same business rules may affect more than one area of operations. This
permits us to collapse enterprise-wide usage of predefined domains and
mappings into a coordinated centralized repository. We can see that not
only data-driven defining rules represent enterprise knowledge, but exe-
cutable declarations can also represent operational knowledge that can
be centralized. Once the repository of rules is centralized, the actual
processing and execution of these rules can be replicated and distrib-
uted across multiple servers located across an enterprise network.
12.5.4 Rapid Response to Changes in the Business Environment
Because the rules are not embedded in source code in unmanageable (and
indeterminate) locations, when the business operation changes, it is more
efficient to update the rule base to speed up the implementation of modi-
fied policies. Changes to the rule base, as long as they do not cause incon-
sistencies within the rule base, can be integrated quickly into execution.
12.5.5 Ease of Reengineering
Managing business rules as content enables easier business process

reengineering. Having all the policies situated in one location enables
analysts to understand what the application under investigation was
meant to do.
12.6 INTEGRATING A RULE-BASED SYSTEM
12.6.1 The Rule Base and the Knowledge Base
There is a subtle difference between a rule base and a knowledge base

(though for our purposes we can use the terms interchangeably). A rule
base is the collection of rules that have been predefined in the system. It
is a statically defined set and is not modified during the execution of an
application. A knowledge base contains all true statements in a system,
whether those statements are predefined rules or truths derived during
the execution of the system.
A pure rule base has no state, and any external event triggers one
(or more) of certain rules that may apply. A knowledge base incorpo-
rates both state ("memory") and inference ("knowledge") into the sys-
tem. As the execution progresses, the choices of rules change as infor-
mation about the external events is integrated into the knowledge base.
As an example, consider these two rules.
1. If the oil consumption exceeds 21 gallons, then fax an order to
the oil delivery company.
2. If the oil consumption exceeds 21 gallons for 10 days in a row,
then fax an order to the oil delivery company.
Rule 1 implies no state, and if the external event occurs, the rule should
be fired. Rule 2 requires that some knowledge be maintained days in
a row so these facts are integrated into a knowledge base.
12.6.2 Scalability
Naturally, resource requirements increase as the number of rules increases.

A knowledge base may require even more storage because all facts that
may be relevant to the triggering of rules will need to be maintained.
12.6.3 Integration Granularity
How tightly coupled is a rules engine to other components of an applica-

tion? The integration granularity measures how closely the rules engine
must be affiliated with any other application code. Depending on the
application, either a tightly coupled or loosely coupled integration may be
desired.
When the rules engine is completely integrated with the application
at compile and link time, the granularity of integration is very tight.
When a separate engine is executing as an intact server at one or more
distributed locations, the granularity of integration is very loose.
12.7 RULE EXECUTION
12.7.1 Firing a Rule
Rules can be restated in action form, if not already worded that way.
All assertional form rules are changed so the assertion becomes the
condition, and if the assertion is not true, then an alert action takes
place. A rule is then said to be fired when its condition is evaluated to
true. This is also referred to as triggering a rule.
For interactive systems, a set of inputs signals the evaluation stage
of the rules engine. An input value that allows one or more conditions
to evaluate to true is called a triggering event.
12.7.2 The Rules Engine
A rules engine is an application that manages the rule base, interacts

with the outside world for the triggering events, and fires associated
rules when conditions are satisfied. The rules engine acts as the con-
troller of the rule execution; it operates in a loop.
1. Evaluate the current state.
2. Identify the rules whose conditions are fulfilled.
3. Select one of those rules for firing.
4. Execute that rule's action.
5. Start again at Step 1.
While some rules engines allow for multiple rule firings (based on
firing all the rules whose conditions are fulfilled), because the actions
may actually create a contradictory state, most rules engines will arbi-
trarily select one of those rules to fire. Another option is to put all fired
rules on a worklist, and as each action is taken, the items on the work-
list are reexamined to make sure there are no inconsistencies.
A rules engine monitors interaction, and as rules are fired and new
facts are established, the rules engine will integrate the new facts into
the knowledge base. New facts (and new rules) are established as rules
are fired and actions are taken. Frequently, an action may be to assert a
new rule. For example, consider these rules.
1. If the customer's credit limit has been reached, then increment the
occurrence count of invalid credit charging attempts.
2. If three invalid charging attempts have occurred, then invalidate
the purchase.
Each time rule 1 is fired, there is a change in the knowledge base. At
some point, the effect of having fired rule 1 some number of times will
trigger the firing of rule 2.
As a more complex example, consider these two policies.
RULEÂSED DATA QUALITY S89
1. All successful customer purchase requests must be reported back

to the customer within 90 seconds of execution.
2. If the product is not in stock, the failure to accommodate the cus-
tomer request must be reported back to the customer within 60
seconds.
The first policy can be encapsulated as a rule.
A If a customer's request is satisfied at time tl, then the customer
must be notified no later than time tl + 90 seconds.
The second policy can be encapsulated as these rules:
B. If the product is not in stock, then replace rule A with rule C.
C If a customer's request cannot be fulfilled, then the customer
must be notified no later than time tl + 60 seconds.
Therefore, the second policy not only has an effect on the order of
events, it actually changes the set of rules. This shows the difference
between a static rule set and a dynamic knowledge base.
12.8 DEDUCTION VS GOAL-ORIENTATION
In dynamic rules systems, the knowledge base is increased as new facts

are established. There are two ways that facts are derived in a knowl-
edge-based system The first way, forward chaining, is a generative
mechanism that uses the knowledge base incorporated with all trigger
events to deduce new information. The second way, backward chain-
ing, starts with a specific goal and attempts to establish the truth of the
goal by establishing the truth of the conditions that lead to that goal.
The suitability of these two methods depends on the actual appli-
cation. In applications where the intention is to identify as much infor-
mation as possible to be presented to the application owners, forward
chaining is appropriate. If the intention of the application is to alert the
users when a certain fact can be determined, backward chaining may be
the ticket. In some cases, a combination strategy may be in order
when some facts need to be established before there are enough facts to
establish the truth of the goal.
12.8.1 Forward Chaining
In a forward chaining system, the knowledge base starts with a default

collection of facts and rules. From the set of facts and any inputs to the
system, the rule conditions are iteratively checked, and new facts are
generated. A forward chaining system derives conclusions from the data.
An example of forward chaining uses this rule:
1. If a customer is purchasing automobile insurance, and that cus-
tomer's car is red, then the customer's insurance premium is
$1,000 a year.
If the following facts are introduced
Mary has a red corvette, and
Mary is purchasing automobile insurance, then the conditions on
rule 1 are satisfied, leading us to the conclusion
Mary's insurance premium is $1,000 a year.
Forward chaining works through a process of pattern matching.
All the rules initially are represented by holding spots for the compo-
nents of each condition. Then, as facts are introduced, the rules engine
attempts to fill in the holding spots with each fact. If all the holding
spots can be filled, then the conditions associated with the rule have
been estabUshed, and the associated conclusion is now added to the
knowledge base.
12.8.2 Backward Chaining
In a backward chaining system, the system begins with a conclusion

and iteratively works backward to find justification for having made
that conclusion. For each rule conclusion, the system looks in the
knowledge base to satisfy that rule's conditions.
An example of backward chaining uses these rules.
1. If a customer has any energy accounts, that customer's monthly
expenditure is the sum of the expenditures on all of that cus-
tomer's energy accounts.
2. If a customer spends more than $100 a month on natural gas, she
is considered to be a high-volume customer.
3. If a customer is a high-volume customer, she is eligible for a 10
percent discount on all charges over $100 a month.
For example, look at these account records.
Customer Meter ID May Natural Gas Charges
Smith, John A. 71804940951 $25.00

Johnson, Steven 71808751965 $76.64
Johnson, Steven 71808751966 $25.98
Corhno, Ralph 71803871094 $140.02
Radzov, Vanessa 71803875823 $18.88
If we ask if Steven Johnson is eligible for the discount, a backward

chaining system would attempt to establish the truth of the conclusion
"Steven Johnson is eligible for a discount of 10 percent on all charges
over $100." Since there is a rule with that conclusion (3), the system then
tries to establish the condition to that rule: "Steven Johnson is a high-
volume customer." As rule (2) addresses the conclusion of high-volume
customers, the next backward step is to establish that "Steven Johnson
spends more than $100 a month on natural gas." The next step is apply-
ing rule (1) backward, implying that the total charges are composed of
the sum of all accounts. Since Steven Johnson has two accounts, the total
monthly charge is $102.62. Rolling back, this establishes the truth of the
conditions of all of the rules, establishing that the answer to the query is
"yes" Steven Johnson is eligible for the discount.
12.8.3 The Execution Sequence
In a rule-based system, execution is not sequentially specified but sched-

uled. Execution is based on stepped iterations. This means that at each
step all the rules with conditions set to be true by any inputs are col-
lected and placed on a queue (the "agenda") for evaluation (that is, the
rules are "activated"). The rule engine determines the sequence of eval-
uation, and the actions are executed.
12.9 EVALUATION OF A RULES SYSTEM
In the article "Business Rules Automating Business Policies and

Practices," Mitchell Kramer describes four basic requirements of a rule-
based approach: isolation, abstraction, integration, and responsiveness
to change. We will use these requirements as the dimensions across
which vendor products will be evaluated.
12.9.1 Isolation
Throughout the book, we have seen that a major driver for defining and
using a rules system is the disengagement of the statement of the busi-
ness operations and poUcies from the technical implementation of those
rules and policies. Therefore, the ability to isolate the rules from the
application that uses those rules is a requirement.
A strategic benefit of isolation is the encapsulation of the rules as
content, which can then be managed separately from any application that
uses those rules. The rule definitions in Chapter 8 are designed to provide
this isolation, especially when managed through a separate interface.
12.9.2 Abstraction
Abstraction refers to the way that rules are defined in the system. This
can encompass a GUI that queries the user for rules or a rules language
for describing rules. This requirement dimension covers the question of
how rules are defined and not necessarily how the rules actually interact
within an executing system. The rules in Chapter 8 are meant to be
edited, modified, and tested from within a separate GUI.
12.9.3 Integration
Integration refers to the way that a rule-based system is integrated with

existing applications. Refer back to Section 12.6.3 for more informa-
tion on integration.
12.9.4 Responsiveness to Change
As a business changes, so do its policies and rules. A robust rules system

will allow significant responsiveness to changes in policies, rules, opera-
tions, and so on, allowing an easy means to browse, edit, modify, vali-
date, and test changes to rules as well as controlling the integration of
new rules into a running system.
12.9.5 Evaluating a Rules Engine
When looking at potential rules engine platforms, these questions

should be asked.
1. Can the rules engine execute as a server? If so, then the rules
engine is likely to be loosely coupled and can be implemented as
a distributed component or even replicated across the enterprise.
2. Are the rules read at runtime or compile time? The rules being
read at compile time would indicate that the rules are read once
when the application is built, as opposed to the application hav-
ing access to a rule base during execution. The implication is that
at execution/runtime, there is a more dynamic system that allows
for rules to evolve as time moves forward.
3. Does the rules engine require access to a database? On the one
hand, requiring a database forces an additional cost constraint,
but on the other hand, the engine may store rules in a proprietary
format.
4. How is a rule system created? How are rules defined, and how
are they stored? How are rules moved from the definition stage
to the execution stage?
5. Is the rule system integrated as a library in an application, or is it
a standalone application? This question asks whether you can
integrate a rules engine and a set of rules as part of another appli-
cation.
6. When updating a rule base, does the entire application need to be
rebuilt, or are builds limited to a subset of the application? To
what degree is the system dependent on the definition of rules?
7. How is the knowledge base maintained during execution? Is
there a database that holds the newly defined assertions?
8. How is the knowledge base maintained during application
update? This question asks how a knowledge base is stored and
restored when the application is halted. If there is a state that
exists in the system, then when the system is brought down, there
should be some format for state persistence.
9. If the rules are integrated directly into a compiled application, is
there a facility for management of the distribution of the applica-
tion? This question asks how rules are distributed to server or
broker applications that execute in a distributed fashion.
10. How are rules validated when updated? It is possible that a new
rule can invalidate or be inconsistent with other rules that are
already in the rule base. Is there a way to check the validity of a
rule set.^
11. Is there a debugging facility? Is there a separate means for testing
out a rule set before putting it into production? If there is a prob-
lem with a production rule base, and is there a tracing facility
that can be turned on so that the execution can be monitored?
12.10 LIMITATIONS OF THE RULE-BASED APPROACH
12.10.1 Detail Management
Associated with any set of rules is the specter of a rule base gone wild,
filled with meaningless trivialities and stale rules that only clog up the
system. When using a rules system, one must be detail oriented to the
extent that the rules engineer is willing to commit to understanding the
rule definition system and the rule management system.
The rules approach requires a dedication to detail, since all objects
operating in the business process as well since all attributes of each
object must be specified. This requires a dedication to detail as well as
an understanding of business process analysis.
In Chapter 10, we looked at use-case analysis and how it affects
requirements. It is at this level of granularity that these skills come in
handy. Ultimately the rule specification operates on a set of objects rep-
resenting entities in the real world, and whether the implementation of
rules is data-oriented or object-oriented, the application must be aware
of all potential actors.
12.10.2 Inflated Expectations
Enthusiastic adopters of rule-based technology may have expectations

that a converted set of complex business policies into a rule-based sys-
tem will always be faster to implement, easier to maintain, and more
efficient to execute. While this may be true for some applications, the
fact is that frequently business policies themselves are poorly specified,
and the business process of converting a policy statement into a rule
base can be very complicated.
The reality is that implementing an application based on rules

requires a senior management-level commitment to seeing the project
all the way through the analysis and definition process. Also, as dis-
cussed in Section 12.10, there must be a dedication to details. The rea-
son for this is that the rule definition process bridges a gap between
consumer and technical producer. The analysis required involves both
domain expertise and technical savvy and forces a close partnership
between the technology provider and the application consumer. The
reward is that the implementation of the technology is not "hoarded"
by a technician and managing the application after it has been com-
pleted can be done in a more cost-effective manner.
12.10.3 Programmers Are Not Eliminated
Because of the "natural language" qualities of rule descriptions, there is a

general impression that a business process application can be built by non-
programmers and that the IT resources and association can be eliminated.
This is a naive assumption because rule specification itself is dominated by
strict syntactical structure, and even if it is not called "programming," the
skill and experience required match that of a progranmier.
With the right tools, though, the learning curve for a nonprogram-
mer should not be a steep one. A visual tool that encourages the user to
define more effective rules but lets the user retain the capability for test-
ing and debugging rules without affecting a production system would
help bring a nonprogrammer up to speed.
12.11 RULE-BASED DATA QUALITY
So now that we know more about rules and rules engines, how do we
use the data quality rules described in Chapter 8? This requires a few
steps.
12.11.1 Rules System Selection
The first step involves the selection of a rules system that will execute our
data quality rules. We can use the guidelines and questions discussed in
Section 12.9. In the case of data quality rules, depending on whether the
rules system is being used for real-time purposes or for offline purposes,
there may be different answers to each of the questions in Section 12.9.5.
For example, if we are using the data quality rules for validating a
data mart, we might want our rules engine to be slightly decoupled but
still integrated with the data loading process. If we are using a rules
engine to generate GUIs, the system can be completely decoupled. If we
want to use rules to ensure the correct operation of our systems, we
might want a tightly coupled rule engine (see Figure 12.2).
In this diagram, the rules engine

and the rule set are both tightly
integrated with the application
acquiring the information.
Rule Set
Data Input
W Application
Process
Rules Engine
In this diagram, the rules engine

and the rule set are slightly
decoupled from the application
acquiring the information.
Rule Set
Rules Engine
FIGURE 12.2 Rules systems

12.11.2 Rule Translation
To use the rules in Chapter 8, we must translate them from their syntac-
tic definitions into a form suitable for an off-the-shelf rules engine. To
demonstrate, let's turn some of our rules from Chapter 8 into a form
that either is a direct assertion or an if-then statement.
Non-null value rule
Attribute B nulls not allowed
is changed into
Assert !isNull(B);
Attribute value restriction
Restrict GRADE: value >= *A' AND value <= T AND value != T '
is changed into
Assert (GRADE >= 'N) AND (GRADE <= T') AND (GRADE != *E')
For the most part, this translation is relatively straightforward. In
order to represent domains and mappings, any chosen rules engine
must support set definitions with a syntax that allows for inserting
strings into sets and checking for set membership. Domains are then
represented as sets of string values, and mappings can be represented as
sets of strings composed of the source domain value, a separator string,
and the target domain value.
One more example would be the completeness rule.
IF (Orders.Total > 0.0), Complete With
{Orders.Billing_Street,
Orders.Billing_City,
Orders.Billing_State,
Orders.Billing_ZIP}
this would change into the following.
IF (Orders.Total > 0.0) THEN !isNull(Orders.Billing^street) AND
!isNull(Orders.Billing_City) AND !isNull(Orders.Billing_State) AND
!isNull(Orders.Billing_ZIP);
A thorough walkthrough of the rules will clarify a mapping into
the assertion/if-then format. Then the rules are ready for the next step
(see Figure 12.3).
12.11.3 Rule Validation
We briefly discussed rule validation in Chapter 8, but rule validation is

outside the scope of this book. No matter what, any rule set must be
Rule Type Generated Test
Code to extract records

Null Value Rule with nulls where they
are not allowed

Value Restrictions Negate restriction where negated assertion
is true

Domain Membership ' \ f Check for ^ ^ INJ
II >A nonmembership J . A where negated assertion
is true

where assertion is true and
any of the specified
attributes are null

where negated assertion is
true and any of the specified
attributes are null

Consistency where condition is true and
the negated assertion is true
Code to update record

Derivation A Check for condition J A when condition is true
FIGURE 1 2 . 3 Rule translation

validated, and that entails guaranteeing that no set of two or more rules
are contradictory.
Actually, the brevity of this section beUes many issues, the most
important one being that validation of rule-based systems is extremely
difficult and is a current topic of research in the world of databases, arti-
ficial intelligence, and knowledge-based systems. In essence, it is easy to
say, "VaUdate the rules" but much harder to actually do it. Hopefully, if
we constrain our system somewhat and don't allow dynamic inclusion of
new rules, the process may be easier (see Figure 12.4).
An example of a An example is if one rule An example is if An example is when

contradiction occurs indicates that if a condition one rule indicates a the action of one rule
when one rule holds, an attribute may be condition that is creates a state where
asserts the negation exempt, while a second rule disallowed by another another rule's condition
of another rule*s indicates that if the same set of rules. may be fired.
assertion. condition is true, the
attribute may not be null
Figure 1 2 . 4 Rule validation

12.11.4 Rule Importing
The last step is importing the rules into the rules system. Depending on
the system, this process may involve something as simple as inserting rules
into a database or as difficult as converting rules from a natural-language-
based format to instances of embedded C++ or Java class objects.
An interesting goal is to use automatic means to translate rules
defined with the natural language format into a format that can be
loaded into a rules system. The process of translation from one format
to another is called "compilation."
12.12 SUMMARY
In this chapter, we looked more closely "under the hood" of a rules sys-
tem. First we further refined our understanding of what a rule is, and
then we discussed the specifics of business rules. In reference to our
rules formalism defined in Chapter 8, we again posited that data quality
rules and business rules are really the same thing.
We then looked a little more closely at rules systems. A rules engine
drives the execution of rules by identifying which rule conditions have
become true and selecting one of those rules to execute its action. This
process can be repeated on all rules whose conditions have been fulfilled.
We discussed the advantages and disadvantages of the rules
approach, including the augmented capabilities, the ability to implement
business policies relatively quickly, and opportunities for reuse. A discus-
sion of how rules engines worked followed. We also looked at how to eval-
uate a rules system and different philosophies embedded in rule systems.
We focused a bit on some of the drawbacks of using a rule-based
system, including the work required to focus on details, as well as the
fact that many users have inflated expectations when moving from an
exclusively programmed system to one that is rule based. We also saw
that using a rules system does not eliminate programming from the
environment, although rules "programming" may be simpler than
using complex programming languages.
Finally, we looked at the steps in transforming the data quality rules
that were discussed in Chapter 8 into a format that can be imported into a
real rules system. This involves selecting a rules system, translating the
rules into the appropriate form, validating the rules to make sure that there
are no contradictions, and importing those rules into the rules system.
13
METADATA AND RULE DISCOVERY
Up until now, our discussion has centered on using data quality and
business rules as tools for leveraging value from enterprise knowledge.
We have shown that we can use data domains and the mappings
between those domains to consolidate distributed information as a sin-
gle metadata resource that can be shared by the entire organization.
Thus far, however, we have concentrated on the a priori definition of
data domains, mappings, data quality and business rules, and general
metadata.
In this chapter, we focus on analyzing existing data to distinguish
the different kinds of metadata. The processes and algorithms presented
help us find metadata that can be absorbed into a centrally managed
core resource and show us how to manage the metadata and its uses.
Domain discovery is the process of recognizing the existence and
use of either enumerated or descriptive domains. The existence of the
domain is interesting metadata; the domain itself can be absorbed as
enterprise reference data. Having identified a domain, it is also interest-
ing to explore whether there are any further derived subdomains based
on the recognized domain. Along with domain discovery and analysis is
the analysis of domains to detect any multiple domains embedded in the
discovered domain. We can also analyze columns that use data domains
to see whether attributes are being overloaded (used for more than a
single purpose) and attempt to perform mapping discovery.
Another significant area of investigation is the discovery of keys in
collected data, merged data, and legacy databases. Older, nonrelational
database management systems did not have primary key requirements,
so when migrating a legacy database into a modern RDBMS, it may be
301
necessary to find primary and foreign keys and to validate the referen-
tial integrity constraints.
A third area of investigation involves the discovery of data qual-
ity/business rules that already exist in the data. There are two steps to
formulating these rules: identifying the relationships between attributes
and assigning meaning to those relationships. We will look at a tech-
nique for discovering association rules and the methodology of figuring
out what those associations mean.
13.1 DOMAIN DISCOVERY
You will recall that data domains, based on a base data type, consist of
either enumerated lists of values or a set of rules that specify restrictions
on the values within that data type, using operations that are allowed
on values within that data type. In Chapter 7, we covered the definition
of data domains through enumeration and description. In many data
sets, common data domains are already used, whether by design or not.
We can presume that if data domains are used by design, they will
already have been documented. But if not, we can make use of heuristic
algorithms to find used data domains.
Domain discovery is the recognition that a set of values is classified
as a set and that one or more attributes draw their values from that set.
Once a domain has been discovered, there is a manual validation phase
to verify the discovered domain. Subsequent analysis can be performed
to understand the business meaning of the discovered domain. This stage
applies a semantic meaning to a domain and can then be used as the
basis for the validation of data domain membership (see Figure 13.1).
13.1.1 Benefits of Domain Discovery
These are some of the benefits of the domain discovery process.

Determining that an attribute belongs to a domain makes it possi-
ble to document the domain and create a means for validating
domain membership rules.
Determining that attributes from different tables use the same
domain provides a means for consoUdating shared reference data
from across the enterprise, thereby lowering risk of inconsistency
METADATA AND RULE DISCOVERY 303
Data Domain 1
Data Domain 2
Data Domain 3
Metadata and
Reference Data
Repository
Data Domain 4
Data Domain 5
Data Domain N
FIGURE 1 3 . 1 Domain Discovery
in enterprise data. It also establishes the existence of a relationship

between those attributes that is inadvertent, or simply had been
forgotten, but may prove to be valuable.
Managing domains under explicit stewardship is a key component
of the management and maintenance of information as an enter-
prise resource.
If the data stewards choose, the domain values may be stored
once and referenced via a numeric encoding, thereby saving space
when referring to domain values.
Extracting the rules into a formal definition transforms an execu-
tion-oriented object into one that can be treated as content,
increasing the value of the enterprise information resource.
Relational metadata is embedded in domain information. For
example, if all customer account numbers referred to by
ACCOUNT in one table are the same accounts referred to by
ACCT_NUM in another table, we infer that the two attributes
refer to the same set of objects.
Information about descriptive domains may be embedded in

application code; extracting these domain descriptions from pro-
gram code helps uncover hidden historical business rules.
13.1.2 Representing Data Domains
As we discussed in Chapter 11, we will have a means for storing, cata-

loging, browsing, and managing data domains. As a review, we will
need three metadata database tables to represent our data domains.
The first table is the domain reference table, which contains the
name of the data domain, the class of the domain (base data type, enu-
merated or rule based), a reference to the base domain or data type on
which it is based, a text description of the domain, and the source of the
data that populates this domain, along with an assigned identifier for
the domain.
create table domainref (
ncime varchar(30),
dClass char(l),
dType integer,
domainid integer
);
The values can all be stored in a single table, referenced by domain

identifier. In this case, we arbitrarily limit the size of the values to 128
characters or fewer.
domainid integer,
value varchar(128)
);
Last, we represent our rules based domains using records that con-
sist of rule statements.
create table domainrules (
domainid int e ger,
rule varchar(1024)
);
13.1.3 Domain Identification Through Expertise
While our predilection may be to rely on automatic knowledge discov-

ery methods to find our domains, before we focus on automated ways
to identify potential data domains, there is another often overlooked,
yet valuable method. This is the process of interviewing the experts
familiar with the data to benefit from their domain expertise.
Our first step should be to avoid using automatic means to extract
knowledge that can be gleaned from conversations with those who
know the most about the data. Assuming that we have already taken
some of the steps described earlier in this book, we will have mapped
out the information chain as well as determined the actors and their
roles within the information factory. We can use the results of this exer-
cise to pinpoint the best individuals to explain known data domains. An
afternoon of conversations may prove to be more productive than
weeks of computer processing time.
13.1.4 Domain Identification Through Exhaustive Set Analysis
Our first focus is in identifying enumerated domains. Despite the fact

that these domains are not explicitly enumerated in any place, we can
propose sets of data values that might possibly be enumerated domains
by finding sets of values that exhibit the behavior of enumerated
domains. This includes but is not limited to the following:
The number of values is relatively small when compared to the
context in which it is used (that is, the number of possible values
that an attribute might take is limited to a small set).
The values are "intuitively distributed." This means that the dis-
tribution, while not always even, will take on characteristics spe-
cific to the context. In some cases, there is a relatively even
distribution, and in other cases, there may be more weight given
to a small subset of those values.
Other domains exist that may be derived from this domain.
The domain is used in more than one table.
The attribute that uses the value from the domain is rarely null.
Unfortunately, these are more guidelines than rules, since exceptions
can be found for each characteristic. The brute force method for identifying
enumerated domains is to look at all possible value sets. We begin by pre-

suming that each column in each table potentially draws its values from a
defined domain. For each table, we go through each column and extract the
distinct values. This set is now a candidate domain, and we apply heuristics
to make a decision about whether to call this set a domain. It turns out that
sometimes we can make some kind of determination early in the analysis,
and sometimes we have to wait until we know more.
Here are some of the heuristics applied:
1. Are there columns in different tables with the same column
name? If so, this may be a good indication that the base data set
is shared among multiple tables, and if some other of these
heuristics apply, this is likely to house a data domain.
2. If the number of distinct values is similar to the number of
records from which the values were extracted, there is a smaller
likelihood of that set being a domain than if the number of dis-
tinct values is much smaller than the record count.
3. Conversely, if the distribution of the values over the column is
relatively equal, then there is a greater likelihood of that set being
a domain.
4. The number of null occurrences is counted. If it is relatively
small, then the values are likely to belong to a domain. If it is not,
it does not eliminate the data set from domainhood, since there
may be other rules that apply within the data set that indicate
that the value is null under some circumstances and takes its
value from a domain under others. In the case of many nulls,
hold the data set for further analysis.
5. Is this data set similar or the same as another data set already
identified as a domain.^ If so, then it is likely to be the same
domain or perhaps a super- or subset of the other domain.
The automated process can only propose candidate domains. The
set of data values only becomes a domain once a user assigns a context
and meaning to those values. When a user decides that a candidate data
set is a true domain, the following steps should be taken:
1. A name is assigned to the domain.
2. A semantic meaning is assigned to that domain.
3. The domain values, the domain name, and the semantic meaning
are inserted into the reference data tables.
4. For each attribute that takes its value from that domain, a
domain membership rule is documented.
Repeating this process builds the domain inventory. Because some

domains are easier to recognize than others, the method of document-
ing domains, as well as their intended meanings, eases the way for fur-
ther domain analysis. Being able to build on a known base will provide
insight into some of the less identifiable domains (see Figure 13.2).
Total number of records: 12834723
, AccountNumber Date XCODE TransTypc Amount Cur

1 AccountNumber
1 Name distina
Street value 12834723 23765
1 TelephoneNumber count
1 SocialSecuritylD
1 AccountNumber
1 CustomerName
1 HomePhone
1 BusinessPhone
1 MobilePhone
1 CurrencyType
1 MaximumTransaction
1 AccountNumber
1 TransactionType Total number of records: 12834723
1 TransactionDate
1 TransactionAmt I AccountNumber I Date I XCODE I TransType I Amount I Cur
1 CurrencyType
I I QiE]/t] / 0 / L Z Z I S
1 In this example, both
1 the column name
1 ''AccountNumber*'and
1 the column name
1 "CurrencyType* are
1 used in more than one
1 table. This may incticate
1 that the values assigned
1 to diese attributes may Potential domains,
1 belong to a specific doe to the low
1 domain. occurrence of nulls
FIGURE 1 3 . 2 Heuristics for Domain discovery

13.1.5 Domain Membership Analysis Through Value Matching
Presuming that we have already started to build the domain inventory,

the next step is to see whether other data attributes make use of the
same domain. We do this by a process called value matching, in which
we analyze how well the set of values used to populate one attribute
match the values of a known domain. This process is useful in deter-
mining if any table attributes use a subdomain of a known defined data
value set.
In this process, we can be a little more exacting when it comes to
deciding domain membership. The value matching process for a specific
attribute can be described using these steps.
1. The attribute's distinct values are collected and counted.
2. The set of unique values is matched against each domain. Fast
matching techniques are used for scalability.
3. For each domain, we compute three ratio values. The agreement
is calculated as the ratio of distinct attribute values that are pre-
sent in a domain to the total number of distinct values in the
attribute. The overlap is calculated as the number of domain
member values that do not appear in the attribute divided by the
number of domain values. Finally, we compute the disagreement
as the number of values that appear in the attribute but are not
members of the domain.
4. The domains are sorted by their agreement percentages. The
highest agreement percentages are presented as likely identified
domains.
When we compare an attributes value set to a known domain,
there are four cases.
1. All of the values used in the attribute are members of the known
domain, and the all of the values in the domain are used in the
attribute (agreement = 100%, overlap - 0%, disagreement -
0%). In this case, it is safe to say that the attribute takes its values
from the known data domain.
2. All of the values in the attribute are members of the known
domain, but there are domain members that are not used in the
attribute {agreement =100%, overlap > 0%, disagreement =
0%). In this case, it is also likely that the attribute takes its values
from the domain, but this may also indicate the attribute's use of
a subdomain, which should be explored.
Some of the attribute values are members of the known domain,

but some of the values used in the attribute are not members of the
known domain (agreement < 100%, disagreement > 0%). In this
case, there are two possibilities. The first is that the there is no real
agreement between the attribute's values and the domain, in which
case the search for a match should continue. The second case is an
indication that the known domain may actually be a subdomain of
a much larger set of values, which should be explored. The deci-
sion will probably depend on the percentages computed.
4. None of the values used in the attribute are taken from the
known domain (agreement = 0%, overlap = 100%, disagreement
= 100%). In this case, it is probably safe to say that the attribute
does not take its values from the domain. Figure 13.3 shows the
domain membership analysis process.
Database
Table
The unique value set is matched

against each domain using
fast set matching techniques.
i ^ i k i i I i i k i k
\r 1 r ^f ^f 1 r ^r
D D D D D D
o o o o 0 o
m m m m m m
a a a a a a 1 Agreement Report |
n n n n n n Domain Agreement
0 1 2 3 4 N Domain ig 97% 1
Domain ij 65% 1
Domain {2 23% 1
Domain '\^ 12% 1
Domain i^ 0% 1
FIGURE 1 3 . 3 Domain membership analysis

13.1.6 Domain Identification Through Pattern Analysis
Aside from looking at enumerated value sets, which are easily analyzed
when presented with already extant data value collections, it is also use-
ful to look for patterns that might signal membership in a rule-based
data domain. For string-based attributes, our goal is to derive a rule-
oriented domain by analyzing attribute value patterns.
For example, if we determine that each value has 10 characters, the
first character is always A, and the rest of the characters are digits, we
have a syntax rule that can be posited as a domain definition. We can use
discovered definition as a validation rule, which we would then add to a
metadata database of domain patterns. Simple examples of rule-based
data domains include telephone numbers, ZIP codes, and Social Security
numbers. What is interesting is that frequently the pattern rules that
define domains have deeper business significance, such as the geographical
aspect of Social Security numbers, as we discussed in Chapter 7, or the
hierarchical location focus associated with ZIP codes.
As a more detailed example, consider a customer accounts data-
base containing a data field called ACCOUNT_NUMBER, which
turned out to always be composed of a two-character prefix followed
by a nine-digit number. There was existing code that automatically gen-
erated a new account number when a new customer was added. It
turned out that embedded in the data as well as the code were rules
indicating how an account number was generated. Evidently, the two-
character code represented a sales region, determined by the customer's
address, while the numeric value was assigned as an increasing number
per customer in each sales region. Because this attribute's value carried
multiple pieces of information, it was a classical example of an over-
loaded attribute. The discovery of a pattern pointed to a more compli-
cated business rule, which also paved the way for the cleaving of the
overloaded information into two separate data attributes.
Our first method for pattern analysis is through the superimposi-
tion of small, discrete "meaning" properties to each symbol in a string,
slowly building up more interesting patterns as meanings are assigned
to more symbol components. Initially, we make use of these basic sym-
bol classifications:
Letter
Digit
Punctuation
White space
Symbol pattern assignment is the first pass at pattern analysis. In

each string, we assign one of these classifications to each character
appearing in a data value. When all the characters in a string have been
classified, the string will have an associated pattern string as well. For
each value string, we prepare and record its pattern string. When all
value strings have been analyzed, there is a column of associated pat-
tern strings ready to be collated and counted.
At this point, there are two tasks to be done. The first is to look for
recurring patterns within the set of generated pattern strings; the second
is checking the generated pattern strings against the known sets of pat-
terns. Either way, the goal is to present candidate patterns representing
rule-based domains to the user.
If no candidates reveal themselves through simple symbol pattern
assignment, then there may be additional embedded information in the
patterns themselves that should be investigated. Our next method for
pattern analysis takes a more macro view of the data by categorizing
strings instead of symbols. At this point, all strings can be classified in
one of the following ways.
Alphabetic
Alphanumeric
Numeric
First name
Last name
Business word
Address words
One of any other categorized word class
In each attribute value, we now assign to each white space-sepa-

rated string one of the word categories, forming a new pattern string.
After all the strings have had patterns assigned, these patterns can be
collated and counted and examined for recurring patterns and matches
to previous patterns (see Figure 13.4).
If during the symbol or whole word analysis we discover common
patterns that have not yet been registered, they may be a new pattern,
which should be documented. At this point, it is a good idea to do a lit-
tle detective work and see if there is some implicit business rule embed-
ded in this pattern. It may turn out that there is some historical reason
for the pattern, which may reflect some business condition that cur-
rently exists and must be maintained or that existed at some point but is
no longer valid, thereby allowing the rule to be changed.
AccountNumber SocialSecuritylD Address
f f y
contains a Street
DD-AAAAA-DD DDD-DD-DDDD
street designator Designator
Domain
Business
f
contains a business word
1
1
_ _
Business
Word
Domain
FIGURE 1 3 . 4 Pattern analysis
The existence of a reason for a rule is one of the subtle differences

between a data quaUty rule and a business rule. Declaring that all values
assigned to afieldin a database record must conform to a pattern sup-
plies a means to validate data within a record, but it gives no insight
into why the value must be in that form. When we associate meaning
with the data quality rule, we suddenly have a context that allows us to
understand why values must conform to that format as well as a con-
text to explain what deviations from that format mean.
13.1.7 Superdomains and Subdomains
The domain discovery process is adaptable to parallelization, since

many attributes may be analyzed simultaneously. In addition, we may
also find that as our domain inventory grows, it is possible that similar,
but not exact, data value sets may appear in our set of domains. When
this occurs, there are two possibilities. The first is that one domain may
be completely contained within another domain, and the other is that
two (or more) similar domains may actually be subsets of a larger data
domain.
We refer to the first case as a subdomain, which is a set of values
that is a subset of another domain. The second case is a superdomain, in
which the true domain that the attributes rely on is the composition (or
union) of the smaller domains.
Note that unless an occasional "self-test" of each domain against
the rest of the set of domains is performed, the only way to recognize
these kinds of overlapping domain value issues is when comparing an
attribute's values against many different known domains. If an analyzed
attribute appears to belong to more than one domain, it may signal the
existence of a subdomain lurking among the known domains. This sub-
domain may represent another business rule, or we might infer that two
similar domains represent the same set of values, in which case the
domains might be merged.
13.1.8 Keeping Domains Current
Recalling the currency/timeliness dimension of data quality, it is impor-

tant that we keep our discovered domains current as well. Since our
sources for data domains are either through expert input, discovered
enumerated domains, and discovered rule-based domains, we must
address the upkeep associated with each of these sources.
For expert-oriented data domains, typically if there is a domain
expert consulting on the analysis of used domains, that expert is likely to
know when there are any changes to the value set. Whether the domain
values are sourced from an external party or the values are based on busi-
ness rules, the expert user will be aware of any modifications, although a
proactive knowledge manager will contact the expert on a periodic basis
to make sure any changes can be documented and updated.
For example, consider a data domain consisting of International
Standard Organization (ISO) country codes. This data may be sourced
directly from the ISO, but since the enumeration of world nations
changes relatively infrequently, this is not the kind of data set that is
refreshed on a consistent basis. Yet, any geopolitical event that causes
an addition or subtraction to the set of nations should trigger a refresh
of the domain. This is the kind of event where an expert's opinion

would properly keep the information current.
For discovered domains, there is a more serious issue, which has to do
with the degree of uncertainty that exists when we rely on discovered
knowledge that cannot be vetted through a corroborative source. If we
claim that a set of data values extracted from a column is a domain, we
basically state that any values that are not in that data set are incorrect for
that attribute. Yet, it is possible that the data values that were extracted are
a subset of the data values that, over time, could be used for that attribute.
This means that we may have documented an incomplete domain, and if
we use a domain membership rule to validate the data values in that col-
umn, we will have spurious violations due to the incomplete domain.
This implies that in order to maintain currency with discovered
domains, some amount of continuing investment must be made by the
organization. If we see that certain attribute domain membership rules
are being violated with an increasing frequency, this may be a sign that
the associated domain(s) are not current. We can incorporate the duties
of domain maintenance into the role of data steward.
13.1.9 Overloaded and Split Attributes
When checking an attribute's data values against known domains, what

is indicated by medium-to-high agreement with relatively low overlap
between an attribute and more than one data domain? This can happen
when there is two attributes' worth of information in one column! This
could also indicate an overloaded attribute, where the same column is
used to represent more than one actual attribute (see Figure 13.5).
Alternatively, when a single attribute uses more than one domain,
it might be an indication that more complex business rules are in effect,
such as the existence of a split attribute, which is characterized by the
use of different domains based on other data quality or business rules.
As a simple example, consider a banking application that registers cus-
tomer account activity. When a customer makes a deposit, a credit is
made to the customer's account in the form of a positive amount of
money. When the customer withdraws money, a debit is made to the
account in the form of a negative amount of money. In this case, the
same column is used, but in fact there are two domains used for this
attribute positive decimal values (when the action is a credit) and
negative decimal values (when the action is a debit).
HIGH
LOW
HIGH
GREEN
RED
HIGH
HIGH
YELLOW Domain 1
YELLOW LOW
LOW
LOW
HIGH
RED
GREEN
LOW
GREEN
HIGH
GREEN
HIGH
HIGH
LOW
RED
YELLOW
HIGH
LOW
YELLOW
RED
HIGH
LOW
LOW
GREEN
GREEN
LOW
YELLOW Domain 2 RED
HIGH
YELLOW
HIGH
LOW
YELLOW
LOW
RED
FIGURE 1 3 . 5 Example of overloaded attributes

The affinity of the values in a single data field to multiple domains

may also mean that the domain as recorded is incomplete and an
adjustment must be made in the reference data set. One plan of action
would be to merge the domains, as long as the constraints associated
with those domains are consistent (both are character strings).
This would result in a superdomain. In this case, any attributes that
use either of the original subdomains may need to be "reregistered."
Either the previously analyzed attributes use the newly created super-
domain with a set of constraints (making it a true subdomain) or, in fact,
the attribute may use any of the values in the newly formed superdomain.
One other possibility is that there are real inaccuracies in the col-
umn. The records with nonconforming data values should be flagged
and made available for correction.
13.2 MAPPING DISCOVERY
Recall that a mapping between domain A and domain B is a set of pairs

of values [Uy b] such that aisz. member of domain A and & is a member
of domain 6, and that there is an intuitive understanding of the rela-
tionship between the values a and b. In Chapters 7 and 8, we looked at
the definition of mappings between domains that we could later inte-
grate as data validation rules. In this section, we look at the discovery of
inherent data mappings embedded within preexisting data sets.
13.2.1 Mapping Discovery
A mapping membership rule applies between two attributes if each

attribute's values are taken from a known domain and each pair of val-
ues in each record exists in a domain mapping. Mapping memberships
represent consistency relationships between the values within each
record and can be used for validation of data records on entry to the
data warehouse. Unfortunately, identifying mappings between domains
is not as simple as identifying domains because domain mappings repre-
sent relationships between sets of data values. Since in each data set
there are many possible mapping relationships (between any pair of
attributes and, in fact, between any pair of sets of composed attribute
values), the algorithm must explore all sets of pairs, which results in a
quadratic number of comparisons of attribute pairs.
The simplest method for identifying a mapping is to iterate over all

pairs of attributes and test to see if the value pairs within each attribute
pair constitute a mapping. For each attribute pair, extract all value pairs
as they appear in the data set. This set can be subjected to some heuris-
tics to see if the value pairs within the set form a mapping. Since there are
different kinds of mappings, we will want to examine this data set to see
if the potential mapping is one-to-one, one-to-many, many-to-one, and
so on. In any mapping, we refer to the first attribute as belonging to the
source domain and the second attribute belonging to the target domain.
13.2.2 Benefits of Mapping Discovery
Mapping discovery is the recognition that two distinct sets of attributes

X and Y each belong to a specific domain and that the relation between
those two sets is embodied in a mapping between those domains.
These are the benefits of mapping discovery.
Managing mappings between domains under explicit stewardship
provides a framework for the maintenance of that enterprise
resource.
Discovering that two sets of attributes are related via a domain
mapping establishes a business rule going forward with respect to
those sets of attributes.
Normalization metadata, such as identifying relations between
attributes within the same table when there is pairwise correlation
(for example, there is a one-to-one relationship between street
address and telephone number), can be used to transform an
unnormalized table into normal form.
13.2.3 Representing Data Domain Mappings
As we discussed in Chapter 11, we have a means for storing, cataloging,

browsing, and managing mappings between data domains. As a review,
we need three metadata database tables to represent our data domain
mappings.
The first table is the mapping reference table, which contains the
name of the domain mapping, the domain id of the source domain, the
domain id of the target domain, a text description of the mapping, and
the source of the data that populates this mapping, along with an
assigned identifier for the mapping.
create table mappingref (
name varchar(30),
targetdomai integer,
mappingid integer
);
The value pairs can all be stored in a single table, referenced by
mapping identifier. In this case, we arbitrarily limit the size of the values
to 128 characters or fewer.
create table mappingpairs (
mappingid integer,
);
Finally, we represent our rules-based mappings using records that
create table mappingrules (
mappingid integer,
rule varchar(1024)
);
13.2.4 Mapping Discovery Through Expertise
Just as with data domains, the best way to identify mappings between
domains is through conversations with experts in the area. Since map-
ping discovery is more complex than domain discovery, any way to
avoid relying on an automated process can only be beneficial.
13.2.5 One-to-One Mappings
A one-to-one mapping represents an interesting value determination

rule within the data set that probably represents an embedded business
rule. It implies that for each unique occurrence of a value of the first
attribute, the second attribute must match what is described in the map-
ping. Therefore, given the value of the first attribute, the second
attribute's value is predetermined, so this rule can be used for both vali-
dation and automated completion. These kinds of mapping member-
ships may exist between composed sets of attributes as well and are also
referred to as functional dependencies.
A one-to-one mapping has certain characteristics that are useful
for integration into the discovery process:
1. All values from the first attribute must belong to one domain.
2. All values from the second attribute must belong to one domain.
3. When all distinct pairs have been extracted, there may not be any
duplicate entries of the source attribute value.
Note that because different source values can map to the same tar-
get value, the third characteristic does not hold for the target attribute
value.
These are some of the heuristics used to determine a one-to-one
mapping:
The number of distinct values in the source attribute is greater
than or equal to the number of distinct values in the target
attribute. This reflects the one-to-one aspect. Since each source
must map to one target and multiple sources may map to the same
target, we cannot have more values in the target than in the
source.
The attribute names appear together in more than one table or
relation. The existence of a tightly bound relationship between
data sets will be evident if it appears more than once, such as a
mapping of ZIP codes to cities.
The attribute values appear in more than one table. In this case,
we are not looking for the names of the attributes occurring fre-
quently but the actual usage of the same domains.
In reality, one-to-one mappings are essentially equivalent to func-
tional dependencies, and we can use the association rule discovery algo-
rithms (in Section 13.5) to find mappings.
13.3 CLUSTERING FOR RULE DISCOVERY
In a number of the rule classes discussed in Chapter 8, there is a significant

difference between the definition of the rules from the automated dis-
covery of embedded rules. When users describe their expectations of the
data, they already have some ideas as to the set of vaUdation rules that
should be applied. But in the discovery process, we don't always know
what to look for, which makes that process particularly hard. In addi-
tion, we need to be able to distinguish between rules that make sense or
have some business value and spurious or tautological rules, which do
not have any business value. Automated rule discovery is not a simple
task, but we do have some tools that can be uses for rule discovery. The
first tool we will look at is clustering.
13.3.1 What Is Clustering?
In general, clustering refers to a set of algorithms used to segment items

in a set into separate partitions, where each value is assigned to a spe-
cific partition. One of the major purposes of clustering is for classifica-
tion, in which we group items by some set of properties and assign a
meaning to each cluster. Clustering hinges on the notion of "distance"
between items in a set, using a distance function to determine how close
any two items are to each other.
There are two kinds of clustering: hierarchical and nonhierarchical.
Hierarchical clustering is used when the set is subject to nested classifi-
cation, such as the taxonomies applied to biological species. Nonhierar-
chical clustering is used to group items together in situations where there
is no subclassing.
Clustering can be performed either as a divisive or an agglomera-
tive process. In the divisive method, the process begins with all data
items in a single cluster, and iterative steps break down clusters into
smaller pieces. In the agglomerative approach, each data item starts out
in its own cluster, and each iterative step attempts to join clusters (see
Figure 13.6).
While the many uses of clustering are applied to very high-tech
applications, such as focused direct marketing or advanced data mining
techniques, remember that the concept behind clustering is identifying
subsets of items within a larger set where all the values within the subset
share some properties. Chapter 8 briefly discussed that given a data set
and a rule base, for each rule the data set can be divided into two
groups: those that conform to the rule and those that do not.
In general, this means that we are classifying data records a priori
by defining a rule (which is equivalent to a set of properties) that can be
imposed on all those records. If we are trying to discover rules instead
Clustering segments the

elements of a set into a
collection of partitions,
where all the elements within
a partition share some set of
properties.
This diagram, shows a
partitioning of elements
based on proximity within
the two-dimensional space.
FIGURE 1 3 . B Clustering
of defining them from the start, our goal is the same: classifying records
based on a rule. We can use the clustering process to help perform this
classification. In the next sections, we will look at ways of using cluster-
ing to discover certain kinds of rules.
13.3.2 Distance and Similarity
Whenever clustering is performed, there must be some measure of distance

between two data items being compared. In order to cluster data items, we
need to create some kind of quantization of the value and then project
each data item (for example, record) into a point in an n-dimensional

space (where n is the number of attributes being observed).
Distance is used in calculating nearness for data items during the
clustering process. There are a number of measures used in clustering.
Euclidean distance This uses the distance measure derived from
the Pythagorean theorem, namely that the distance between any
two points is the sum of the squares of the difference between the
corresponding data points.
City block distance Similar to the way that a pedestrian must
traverse a street grid, this measures the distance in terms of walk-
ing along the line segments in a grid. The measure is the sum of
the differences between each pair of points.
Exact-match distance For variables whose values are not allo-
cated along continuous dimensions, the values assigned are either
1 for an exact match or 0 for not a match.
Figure 13.7 shows these more clustering methods. We will look more at
similarity and distance in Chapter 16.
13.3.3 Divisive Clustering
This is the basic approach for divisive clustering.

1. Assign all values into a single cluster.
2. Determine a division of each cluster that will maximize the simi-
larity between all the elements within one side of the division and
maximizes the difference between any pair of items across the
division.
3. Repeat Step 2 until either some predefined limit of clusters has
been found or some predefined limit of cluster size (minimum, of
course, is one) has been reached.
What is glossed over in the simple statement of Step 2 is that the
determination of this division requires examining a huge amount of
information and processing. For example, for each proposed division,
each pair of clusters must be examined to measure the difference
between them, and all the values within each proposed new cluster
must be examined for similarities between values. This is particularly
computationally intensive, and divisive methods are typically not as
popular as agglomerative methods.
Value A == Value B !
Value A != Value B O
Euclidean Distance Exact Match Distance
City Block Distance
FIGURE 1 3 . 7 Distance measures
13.3.4 Agglomerative Clustering
The approach for agglomerative clustering is as follows.

1. Assign each data element to its own cluster.
2. Find the two closest clusters and combine them into a single cluster.
3. Repeat Step 2 until there is only one cluster left.
Again, the essence of the algorithm is embedded in the second
step, where the two closest clusters are combined. It turns out that the
determination of the two nearest neighbors is the tricky part, and there
are a number of methods used for this determination.
Single link method The distance between any two clusters is
determined by the distance between the nearest neighbors in the
two clusters.
Complete link The distance between any two clusters is deter-
mined by the greatest distance between any two data items in the
two different clusters.
Unweighted pair-group average The distance between any two
clusters is calculated as the average distance between all pairs of
data items in the two different clusters.
Weighted pair-group average This method is similar to the
unweighted pair-group average, except that the size of the cluster
is used as a weight.
Unweighted pair-group centroid A centroid is a point calculated
as the average point in a space formed by a cluster. For any multi-
dimensional space, the centroid effectively represents the center of
gravity for all the points in a cluster. In this method, the distance
between any two clusters is measured as the distance between cen-
troids of the two clusters.
Weighted pair-group centroid This is similar to the unweighted
pair-group centroid method, except that the size of the cluster is
used as a weight.
Ward's method This method measures the distance between two
clusters based on the total distance between all members of the
two clusters.
13.3.5 Value Ranges Using Clustering
In Chapter 8, we discussed value restrictions, which can be applied to

both numeric and character string typed values. In the analysis process,
it would be useful to be able to identify occurrences of the use of value
restriction rules, which we could then incorporate into the rule base.
For any numeric value set, we will initially assume that there are no
restrictions on the values. Then, the simplest value restriction rule for
numeric values bounds the set by a maximum and a minimum value,
which we respectively call the upper and lower bounds. Numeric re-
strictions can be composed of either a single contiguous range or a
collection of noncontiguous regions. A contiguous region is any

numeric range that incorporates at least an upper or a lower bound, or
both. Two regions are noncontiguous if the upper bound of one is less
than the lower bound of the other.
The best way to remedy this problem is to reverse it from being a
divisive clustering problem to being an agglomerative problem. We
begin by assuming that each value is its own cluster, and then we begin
to aggregate clusters based on the distance between the different points.
Since these points are all along a single dimension, this clustering is rel-
atively simple. Each cluster represents a value range.
If we believe that there are noncontiguous regions incorporated into
the set of value ranges, we can propose a minimum distance between
clusters that we expect will exist within the data set. This becomes a
parameter to the clustering algorithm that signals to stop trying to merge
clusters when the distance between clusters is greater than the minimum.
For example, if we think there might be more than one value range with
a distance of at least 15 points between them, we will stop clustering
unless there are two clusters whose distance is less than 15.
13.3.6 Clustering and Generalized Rule Discovery
Clustering is useful for more than just value range detection. When we
can characterize each data record as a combination of measurable prop-
erties, we can use clustering as a classification method that depends on
those properties. Since all members of a cluster are related because they
are near to each other based on some set of metrics, we can aim toward
translating the clustering characteristics into a description that explains
why those records belong together. If the distinction is clear, it is a
strong argument to represent those characteristics as one of our data
quality rules.
For example, let's assume that we have a database of customers
that includes a significant amount of demographic information as well
as sales histories. If we wanted to determine whether there were differ-
ent classes of customers based on both their sales histories and their
demographics, we can apply a clustering algorithm to break down the
set of customers into different classes. If our goal is to determine the
characteristics of different classes of customers, we can direct the clus-
tering based on fixing the "type of customer" property by assigning one
of a set of values based on sales volume. Once the clustering has been
done, those characteristics that are the drivers in determining similarity

can be used as the basis for defining new business rules.
13.3.7 K Means Clustering
The example from the previous section is a good opportunity to use a K

means clustering algorithm. In the K means algorithm, the K stands for
the number of clusters we would like to find. These are the steps we
take next.
1. Arbitrarily select K seeds that are to be the first guess at deter-
mining the centroids of the clusters.
2. For each record, assign the record to a cluster based on the clus-
ter centroid to which it is nearest.
3. After all records have been assigned to a cluster, recompute the
centroids of the clusters.
4. If the centroids changed from the previous iteration, then repeat
the process starting at Step 2.
5. If the centroids did not change from the previous step, then the
process is finished.
We use our distance measurement in Step 2, where we assign each
record to a cluster based on nearness to the presumed centroid.
13.4 KEY DISCOVERY
In legacy data from systems that do not have key requirements, we

might like to be able to discover candidate keys within the data. Recall
that a table key is a set of attributes such that for all records in a table
no two records have the same set of values for all attributes in that
key set.
In the key discovery process, our goal is to find sets of attributes
within a table that we expect to have a key such that the set of attributes
uniquely identifies a record in the data set. Initially, we can presume
that the combination of all the attributes must uniquely define an entity
in the table; otherwise, there would be duplicates, which violates our
assumption.
The process for determining keys is an exhaustive one. We can use
a recursive process.
1. Initialize key list K to null.

2. Call the tester function with the complete set of attributes.
The tester function looks like this, using pseudocode and the use of
the SQL select statement. The input is a set of attributes, the original
table, and the count of records in the original table. The second state-
ment selects out the count of records based on the distinct appearances
of the attributes in the attribute set S. If the count of records is the same
as the complete table count, the attribute set forms a key. For each
attribute in the set, remove it and call Tester recursively until the set is
of size 0.
Tester(S: attribute s e t , Tab: database t a b l e , count:
integer) {
I f s i z e ( S ) == 0, return;
Integer i = s e l e c t count ( d i s t i n c t S) from Tab;
If (count == i ) then add S to key l i s t K;
For each attribute a in S {
Set S' = S - attribute a;
Call T e s t e r ( S ' , Tab, count);
}
}
When this process completes, the key list K will contain all
attribute sets that are unique across the table. We can sort this list in
ascending order by attribute set size, which will tell us which attribute
sets are the smallest candidate keys.
13.5 DECISION AND CLASSIFICATION TREES
A different way to look at the classification of data is based on a tool

called a decision tree. In a completed decision tree, each node in the tree
represents a question, and the decision as to which path to take from
that node depends on the answer to the question. For example, we can
have a binary decision tree where one internal node asks whether the
employee's salary is greater than $50,000. If the answer is "yes," the left
path is taken, but if the answer is "no," the right path is taken.
At each step along the path from the root of the tree to the leaves,
the set of records that conforms to the answers along the way continues
to grow smaller. At each node in the tree, we have a representative set of
records that conforms to the answers to the questions along that path.
Each node in the tree represents a segmenting question, which subdi-

vides the current representative set into two smaller segments. Every
path from the root node to any other node is unique. Each node in the
tree also represents the expression of a rule, and at each point in the tree
we can evaluate the set of records that conform to that rule as well as
the size of that record set (see Figure 13.8).
13.5.1 Building Decision Trees: The CART Algorithm
Obviously, decision trees are useful for harvesting business rules from a
data set. But how do we build a decision tree? Here we discuss the
CART algorithm. CART is an acronym for Classification and Regres-
sion Tree, and a CART tree is built by iteratively splitting the record set
at each step based on a function of some selected attribute.
The first step in building a CART tree is to select a set of data
records on which the process is to be performed. We might want to sub-
mammal?
yes
domesticated? reptile?
I yes I no \
lives in a cage? long snout amphibian?
yes : I yes yes yes :
I rodent? "| | barks? 11 feline? "| [ long neckT^ | Everglades?] | Hawaii? j [ bumpy skin?) | flies? |
I yes I n o ] | yes | n o ] | yes | n o | [ yes | n o ] | yes | n o ] | yes yes no I

I yes I yes I no ]
r T rTTTIXT: TJ. I o
8
O u
FIGURE 1 3 . 8 Decision trees

METADATA AND RULE DISCDVERY 329
select some records that we know already have properties in common

such that we can use the result of the decision tree to perhaps enhance
new records based on these dependent variables. All other data attri-
butes become independent variables. This selected data set is used as a
training set, and the results of the decision tree process are to be applied
to the entire data set at a later date.
The next step in the process is deciding which of the independent
variables is the best for splitting the records. The choice for the next split
is based on choosing the criteria that divide the records into sets where,
in each set, a single characteristic predominates. The standard method is
to evaluate the possible ways to split based on each independent vari-
able, measuring how good that split will be. The measure for evaluating
a decision to split is based on a heuristic. Here are some examples:
Gini This splitting rule looks at ways to maximize the set differ-
entiated by a split, with the goal of isolating records with that
class from other records.
Twoing This is a method that tries to evenly distribute the
records at each split opportunity.
When using the Gini method, we basically want to create a split
such that the difference in the amount of diversity from the original
data set to that within each set (after a split) is increased. Diversity is
measured based on the probabilities of records falling into the classes as
defined by the split.
The complete tree is built by recursively splitting the data at each
decision point in the tree. At each step, if we find that for a certain
attribute all values are the same, we eliminate that attribute from future
consideration. When we reach a point where no appropriate split can
be found, we determine that node to be a leaf node.
When the tree is complete, the splitting properties at each internal
node can be evaluated and assigned some meaning. There are some
advanced issues regarding error rates and probability (especially at the
leaves, where the lack of a decision may not necessarily define a class),
as well as pruning algorithms to reduce the size of the tree.
13.5.2 Turning Decision Trees into Rules
Each path of the tree represents a set of selection criteria for classifica-
tion. If we decide that the leaf represents a vaUd set of records (that is.
there is some valuable business relevance), we walk the tree from the
root to the leaf, collecting conditional terms, to accumulate search crite-
ria in the data set. By the end of the traversal, we basically have the con-
dition under which the records in that set are classified. If that
classification should enforce some assertion, we can use the condition
as the condition in a data quality consistency rule. Other rules can be
constructed similarly.
13.6 ASSOCIATION RULES AND DATA QUALITY RULES
Association rules, as applied to databases, are rules that specify a rela-

tion between attributes that appears more frequently than expected if
the attributes were independent. Basically, an association rule states
that the values of a set of attributes determine the values of another set
of attributes, with some degree of confidence and some measure of
support.
Note that this definition strongly resembles our functional depen-
dence rule from Chapter 8, except that in the association rule sense, we
are looking for distinct occurrences of values within the attributes as
opposed to all values within the attributes. Actually, if there are associa-
tion rules that cover all possible values of the attributes, we can declare
the existence of a functional dependence. Otherwise, association rules
define potential consistency rules.
An association rule is specified as {source attribute value set}fi{tar-
get attribute value set}. The source attribute set is referred to as the
"left-hand side" (for obvious reasons), and the target attribute set is the
"right-hand side." The confidence of the association rule is the percent-
age of the time that the rule applies. For example, if 85 percent of the
time that a customer buys a network hub he also buys a network inter-
face card (NIC), the confidence of the rule {iteml: Buys network hub} fi
{item2: Buys NIC} is 85 percent. The support of a rule is the percentage
of all the records where the left-hand side and right-hand side attributes
have the assigned values. In this case, if 6 percent of all the records have
the values set {iteml: Buys network hub} and {item2: Buys NIC}, then
the support for the rule is 6 percent.
An association rule with a high confidence may represent a data
quality rule, where the occurrences that do not conform to the rule are
actually errors that could be corrected. What is nice about the associa-
tion rule is that it can be used for both vaUdation as well as correction!
Algorithms for association rule mining are relatively complex and

are outside the scope of this book. On the other hand, many vendors
provide tools that can discover association rules, as well as tools for
building decision trees, along with other data mining applications that
can be used for data quality rule discovery.
13.7 SUMMARY
Finally, what do we do with the results of our discovery? By accumulat-

ing the discovered domains, mappings, and rules as managed content,
integrated with an access control system and a metadata browsing sys-
tem, we can build an environment for user subscription to the central-
ized metadata resource. This is consistent with our discussion in
Chapter 11, where we saw how to manage and present enterprise meta-
data.
The bulk of this chapter has been spent looking at ways to discover
data quality and business rules. Part of any data quality program
should be a knowledge discovery component, where the goal is to dis-
cover the domains, mappings, and rules that have not been disclosed by
the experts or have been lost due to employee turnover. This enterprise
reference and metadata, embedded in the data sets themselves, can be
set free through the use of relatively straightforward heuristic algo-
rithms and data mining techniques.
14
DATA CLEANSING
It may be surprising that a book on data quality doesn't talk about

"cleansing" until nearly the end of the text. Yet, while data cleansing may
be the first thing on one's mind when thinking about data quality, cleans-
ing actually encompasses only a small part of the data quality universe.
The reason for this is that if we have treated data quality properly,
planned from the beginning, then we should not really have to worry
about ever cleaning data, since it should already be fit for use by the time
the user sees it.
In actuality, though, since we have been largely operating for 40
years without considering data quality a priori, we live in a world
where data have been subject to significant entropy and there is a signif-
icant data cleansing problem. But first, what is data cleansing?
For our purposes, we define data cleansing as transforming data
perceived to be incorrect into a form that is perceived to be correct. In
some environments, we rely on a data standard to define what is and is
not correct. In other realms, the notion of correct or incorrect is almost
inherent in the application.
Most data quality tool vendors concentrate on name and address
cleansing. This typically consists of the following areas.
Customer record parsing, which takes semistructured
customer/entity data and breaks it up into component pieces
such as title, first name, last name, suffix, and so forth. This also
looks for connectives (DBA, AKA, &, "IN TRUST FOR") that
indicate multiple parties in the data field.
Address parsing, which is a similar activity for addresses.
333
Address standardization, which makes sure that addresses con-

form to a pubUshed postal standard, such as the USPS Postal Stan-
dard. This includes changing street designations to the standard
form (ST for street, AVE for avenue, W for west).
Address cleansing, which fills in missing fields in addresses (such
as ZIP or ZIP + 4 codes) and corrects mistakes in addresses, such
as fixing street names or reassigning post office locality data,
such as changing the city field in an address from a vanity address
("ROLLING HILLS HEIGHTS" to "SMALL VALLEY").
A primary example is in the world of direct mail marketing. Mar-

keting material is addressed and mailed, but if the mailing address is
incorrect, the mail cannot be delivered and is returned to the sender. But
there is a well-designed standard for names and addresses defined by
the U.S. Postal Service, and a process called address standardization
relies on the standard plus data published by the Postal Service to auto-
matically correct addresses that are thought to be undeliverable.
While there is a well-defined standard for addressing, the address
does not have to be perfect to reach its destination. It only has to be
addressed "well enough." This is a fine distinction, since in the case of
U.S. addresses, there is a standard, and it is easy enough (for the most
part) to correctly modify an incorrect address. But in those cases where
there is no defined standard, there may still be issues of data cleansing
that will bring data into a form that is, as per our definition of data
quality, fit for use.
In this chapter, we explore data cleansing. We start with a discus-
sion of data standards and the process of bringing data into confor-
mance with a standard (standardization). We also look at common
error paradigms that create the need for cleansing. We then discuss
what we call metadata cleansing making sure that the enterprise
metadata reflects the defined metadata guidelines, as well as making
sure that the data resource conforms to the metadata.
We then look at some of the more routine forms of data cleansing.
The first is merge/purge and duplicate elimination, which looks at the
combination of multiple data sets and the determination (and subse-
quent elimination) of duplicate entries. This is followed by a section on
updating missing fields how to determine what a missing field means
and how it can be automatically filled. A specialized treatment of
updating missing fields is the addition of salutations, titles, and other
kinds of expansions.
DATA CLEANSING 335
We also focus a significant section on U.S. Postal address standard-

ization, since it is a good working example of a well-defined standard
for which there is a well-developed application solution for data cleans-
ing. Finally, we talk about data cleansing in terms of migration of data
from legacy systems to new application systems.
14.1 STANDARDIZATION
According to the Webster's New Collegiate Dictionary, a standard is

"something set up and established by authority, custom, or general con-
sent as a model or example." In the data world, we can say that a stan-
dard is a model to which all objects of the same class must conform.
When we talk about standardization, it means confirming that a
data record or a message conforms to a predefined expected format.
That format may be defined by an organization with some official
authority (such as the government), through some recognized authori-
tative board (such as a standards committee), through a negotiated
agreement (such as electronic data interchange (EDI) agreements), or
just through de facto convention (such as the use of hyphens to separate
the pieces of a U.S. telephone number, as opposed to other national for-
mats that might use commas or periods).
We examine the definition of a standard and the benefits of using
one. More important, we talk about two critical ideas with respect to a
good standard. The first is the ability to automatically test for standard
conformance, and the second is the provision of a means for automated
transformation into standard form. A well-designed standard form will
allow for both of these concepts to be realized.
14.1.1 What Is a Standard?
A data standard is a format representation for data values that can be

described using a series of rules. In addition, in order for a format repre-
sentation to be a standard, there must be an agreement among all inter-
acting parties that their data values will conform to that standard.
A standard may enforce some cosmetic string formatting. An
example is the format of a U.S. telephone number. As per our definition,
the format of a U.S. phone number is discussed by a committee called
the Industry Numbering Committee, and the standards are agreed to by
the members of that committee, comprising all major telecommunica-

tions companies.
The simplest standard governs the grouping of symbols
1-XXX-XXX-XXXX
where each X is a character symbol. We can further refine this standard
by imposing the rule that each X represents a digit between 0 and 9.
A standard may also enforce the use of a specific set of data
domains and domain mappings. To continue our U.S. telephone num-
ber example, we can insist that the standard format is
1-AAA-XXX-XXXX
where, as before, each X represents a digit between 0 and 9 but where
the AAA component must be drawn from the list of valid Numbering
Plan Area (NPA) codes for the United States. To take our standard even
further, we can also impose a rule that says that the standard format is
l-AAA-BBB-XXXX
which adds the rule that the second triplet of numbers, EBB, must refer
to a valid NXX sequence (NXX represents the central office code) asso-
ciated within an NPA area (see Figure 14.1).
There are many further constraints that can be added, depending
on the application. If you are working for a phone company and you
need to validate billing data, you may have more constraints than if you
are collecting telephone numbers for a customer database.
One quick note: Telephone numbers in North America use a differ-
ent format standard than telephone numbers in other parts of the
world! So, even if you have all the rules down for the United States,
there is still a world of rules (literally) that might need examination.
14.1.2 Benefits of Standardization
If we try to compare two items that are not of the same class, we are
told that we cannot compare apples with oranges. When we talk about
standardization, we are really trying to make sure that we compare
apples with apples and oranges with oranges. Standardization is a
process by which all elements in a data field (or a set of related data
fields) are forced to conform to a standard.
There are many benefits to this process, the first of which we have
already mentioned: conformity for comparison. When aggregating data
down a column (or set of columns), we will be in much better shape
when we know that all the values in those columns are in standard
DATA CLEANSING 337
Values defined by the North American standard

for telephone numbers, as defined by the Industry
Numbering Committee (INC), are actually good
examples of an overloaded attribute. The telephone
number itself represents a unique means for
connectivity but also represents a routing mechanism
for making the conneaion.
FIGURE 1 4 . 1 Telephone number standards
form. That way, our sums will not be skewed by erroneous data values.
The same goes for using standard values as foreign keys into other
tables we can feel comfortable that referential integrity is more likely
to be enforced if we stay within the standard.
Another interesting benefit of standardization is the ability to
insert an audit trail for data error accountability. The process of stan-
dardization will point out records that do not conform to the standard,
and these records can be tagged as incorrect or forwarded to a reconcil-
iation process. Either way, by augmenting a table with audit trail fields
and recording at which point in the data processing chain the tables are
standardized, we can trace back any significant source of errors. This
gives us a head start in analyzing the root cause of nonstandard data.
Probably the most important benefit of standardization is that

through the standards definition process, organizations create a stream-
Uned means for the transference and sharing of information. As compa-
nies agree on how to exchange information, they also create a means
for increasing the volume of exchanged information. With formal stan-
dards, more organizations can subscribe to the standard and partici-
pate. Examples include the use of data standards for financial
transaction exchange (such as SWIFT or FIX), standards for the health
industry, standards for geospatial mapping, and so on.
14.1.3 Defining a Standard
The first step in defining a standard should be obvious: Invite all data
consumers together to participate in defining the standard because a
standard is not a standard until it is recognized as such through the con-
currence of the users. In practice, a representative body is the best vehi-
cle for defining a data standard.
The second step is to identify a simple set of rules that completely
specify the valid structure and meaning of a correct data value. The
rules in this set may include syntactic rules that define the symbols and
format that a data value may take as well as data domain and inclusion
rules that specify the base data domains from which valid value compo-
nent may be taken. The most important part of this step is making sure
that there is a clear process for determining if a value is or is not in stan-
dard form.
The third step in standard definition is presenting the standard to
the committee (or even the community as a whole) for comments.
Sometimes small items may be overlooked, which might be caught by a
more general reading. And, remember, a standard only becomes a stan-
dard when it is accepted by all the data consumers. After a brief time
period for review and comments, an agreement is reached, and the stan-
dard is put in place.
14.1.4 Testing for Standard Form
One critical idea with respect to data value standards is that if a value is
dictated to conform to a standard, there must be a way to test to see if a
DATA CLEANSING 339
data value conforms to the standard form. This means that the defini-
tion of a standard must by association imply the test.
This test can usually be embodied using application code and refer-
ence data sets. Going back to our U.S. telephone number example, it is
easy to write a program that will check if the telephone number string
itself matches the defined format. In addition, two reference tables are
needed. The first is a domain table listing all valid NPA prefix codes that
are valid for the covered geographical area. The second is a domain
mapping between NXX codes (the second digit triplet) and the valid
NPA codes associated with each NXX. The first table is used to test our
first rule that the NPA code is a valid one. The second tables allows
us to test if the local exchange code is used within that NPA.
Another very good example is the U.S. Postal Standard. The Postal
Service has a well-defined standard for mail addressing. This standard
includes a definition of each component of an address (first name, last
name, street, street suffix, city, etc.), as well as standard and nonstandard
forms for the wording of particular components. For example, there is a
set of standard abbreviations for street types the address suffix
"AVENUE" has a set of commonly used abbreviations: {"AV," "AVE,"
"AVEN," "AVENU," "AVENUE," "AVN," "AVNUE"}, but only one
("AVE") is accepted as the Postal Service standard. The Postal Service
also has a predefined domain of valid ZIP codes, as well as a mapping
between valid street name and city combinations and an assigned ZIP
code. In each case, though, an address can be tested to see whether it is in
valid form. We look at this in greater detail in Section 14.9,
14.1.5 Transforming in Standard Form
The other important feature involving data standards is the ability to

transform data in nonstandard form to standard form. While this may
sound simple, it can be quite complex because there are many ways that
a data value can violate a standard form. Using our Postal Service
example, an address may not be standard because the ZIP code has six
digits, the ZIP code is not valid for the named city, or the city named
does not exist.
Transformation into standard form can range from straightfor-
ward, in cases where the standard form is defined based on values of
other fields (such as figuring out the right ZIP code), to impossible.
from which nonstandard data is completely unrecoverable (especially

when the data values are incorrect).
14.2 COMMON ERROR PARADIGMS
Many types of what appear to be random errors actually are due to rel-
atively common problem paradigms, as we see in the next sections.
They are outlined in Figure 14.2.
Error Type Description Example
Attribute Data model is not configured for Using one data attribute holding 1
Granularity proper granularity of values. street, city, state, and ZIP code
Strict Format Format for data entry is too Insisting on a first name, middle 1
Conformance restrictive. initial
1 Semistructured Format for data entry is too Freeform text used in a description
Format permissive. field
1 Finger Flubs Errors come from mistakes in Typing INTERANTIONLA

data entry. instead of INTERNATIONAL
1 Transcription The data entry person makes spelling the name LOSHIN
1 Error a mistake when transcribing data. as LOTION
Transformation An automated process makes a Some defined function merging

Flubs mistake in data extraction and data from two tables is incorrect
loading.
1 Misfielded Data from one attribute appears The street address appears in an
Data in a different attribute. ADDRESS2 field, instead of
ADDRESSl 1
Floating Small-sized attributes cannot support Company name data spills out of
Data longer values, so data flows from the name field and into the street
one field into the next. address field
1 Overloaded The same attribute contains data Inserting both names of a couple in
1 Attributes associated with more than one value, the name field of a customer
or the attribute respresents more database when both members of
than one property of the entity. the couple should be listed
FIGURE 14.2 Common error paradigms

DATA CLEANSING 341
In this paradigm, the data model is not configured to correctly incorpo-

rate the granularity of information required. A good example is a cus-
tomer name field, which incorporates first, middle, and last names all in
one field. If the user wants to have each name separated, it would be
better to use threefields:last name, first name, middle name.
14.2.2 Strict Format Conformance
When the format of the data attributes is too restrictive, the data entry
user may not be able to correctly put the right information into the
database. Using the same example of customer name, let's assume that
we broke up one data field for name into three fields: last name, first
name, and middle initial. But the are many people who prefer to be
called by their middle names instead of their first names, in which case
the correct data should be last name, first initial, middle name. But since
there is only room for a middle initial, either both first initial and mid-
dle name are crammed into the first name field, or the middle name is
placed in the first name field and the first initial is placed in the middle
initial field, thereby reversing the customer's first two names!
14.2.3 Semistructured Formats
Semistructured data are data that map to a heuristically determinable

format. In other words, the user knows what to expect in general,
although not all the data conforms to the expected form all the time.
Therefore, there is some intuition about how to get the information. As
an example, given an account name field in a bank database, the name
field might be semistructured, in that different data entry personnel may
have transcribed the name into one of these formats.
First name, middle name, last name
Last name, comma, fist name, middle name
Last name, comma, title, first name
The result is a localized, though potentially widespread inconsistency of
format in the context of an imposed structure.
14.2.4 Finger Flubs
Typists are not infallible, and typing errors creep into data. Common
mistakes are transcribed letters inside words, misspelled words, and
miskeyed letters. In one environment that maintained information on
businesses, many of the company names in the database had the amper-
sand (&) character as part of the name (Johnson & Smith). A frequent
error that occurred was the appearance of a 7 instead of the ampersand
(Johnson 7 Smith). This makes sense, however, when you see that on
one key the &C character is shift 7. (So, the 7 might be due to a sticky
shift key!)
14.2.5 Transcription Errors
When information is being transcribed, variations in spellings or abbre-

viations or phonetic similarities may be introduced.
14.2.6 Transformation Flubs
Sometimes errors are introduced when applications extract from one

source and transform it before depositing it into the target database. If
the transformation rules are not completely correct or if there is a flaw
in the transformation application, errors will be created where none
originally existed.
For example, in an interesting case, a database of names was found
to have an inordinately large number of high-frequency word frag-
ments, such as "INCORP," "ATIONAL," "COMPA." It turned out
that the field had been reconstructed from data extracted from a legacy
database, which had a limit on the size of an attribute, forcing the data
modeler to use more than one actual field to make up one virtual field.
When the data was merged, the merging application inserted a space
between each of the original fields when they were concatenated. Since
some of the original data spanned two of the actual data fields, the
insertion of the space created a collection of fragments that did not
appear in the original data.
DATA CLEANSING 343
14.2.7 Misfielded Data
Another common error is data placed in the wrong field. In a similar

example, the street address field was insufficiently sized, so the entry
person just continued entering street address data into the city field.
14.2.8 Floating Data
When a data model is not specified well or the metadata is unclear

about what kind of data goes into whichfields,we can get errors associ-
ated with floating data where information that belongs in onefieldis
contained in different fields in different records in the database.
14.2.9 Overloaded Attributes
Either a reflection of poor data modeling or as a result of changing busi-

ness concerns, there is information stored in one data field that actually
contains more than one data value.
Examples of this are prevalent in the financial industries, where
companies are moving from an account-oriented to a customer-oriented
view. Typically, account names are the official names associated with a
financial account and may include many combination forms with many
associated "context terms." The context terms represent business rela-
tionships that might need to be exposed in a customer or parties data-
base. Here are some examples.
1. John Smith and Mary Smith
2. John and Mary Smith
3. John Smith in Trust for Charles Smith
4. John and Mary Smith UGMA Charles Smith
5. John and Mary Smith Foundation, John and Mary Smith
Trustees
In the first example, there are clearly two parties, John Smith and
Mary Smith, and the two names are separated by the connector "and."
In the second example, we also have two parties, although the seman-
tics of the connective are a little more complicated because it relies on a
business rule that says that married women take their husband's last
name.
In the third example, we also have two parties, but in this case there
is a different kind of relationship represented by the business phrase "in
Trust for." The two parties here are John Smith and Charles Smith.
In the fourth example, we now have three parties: John Smith,
Mary Smith, and Charles Smith, linked together via a different business
term ("UGMA," an acronym for Uniform Gift to Minors Act).
In the fifth example, we have three parties but more relations
between the parties. To start off, we have John Smith and Mary Smith,
and now we also have the John and Mary Smith Foundation, which is
an entity in its own right. The relationship here is that John Smith and
Mary Smith act as trustees for the foundation that bears their names.
14.3 RECORD PARSING
In all of these cases, the first step in data cleansing is the identification of
distinct data elements embedded in data records. In the most common
situations, where the data set under investigation consists of name,
address, and perhaps other important information, the data elements
may be as follows.
First names
Last names
Middle names or middle initials
Title (MR, MRS, DR)
Name Suffix (JR, SR, MD)
Position (SALES MANAGER, PROJECT LEADER)
Company name
Building
Street address
Unit address (Apt, Floor, Suite)
City
State
ZIP code
Business terms
The goal of data parsing is to absorb the data embedded in an

attribute and produce the data components that exist in the field. In
order to do this, we need four pieces.
DATA CLEANSING 345
1. The names and types of the data components expected to be

found in the field
2. The set of valid values for each data component type
3. The acceptable forms that the data may take
4. A means for tagging records that have unidentified data
components
If this sounds familiar, it should! The names and types of the data
components are the metadata for this parse model; the sets of valid val-
ues for each data component type are the same domains that we dis-
cussed in Chapters 7 and 8. The acceptable forms that the data may
take can be expressed using our data quality rules as well.
Most tools that are on the market that provide a data parsing
activity are either experience based or pattern based. Experience-based
tools take into account the vendor's experience working with large cus-
tomer name and address databases, letting the tool rely on predeter-
mined data parsing rules and inferred data domains. The pattern based
tools are more flexible in that they provide some analysis of the user's
data set to look for common patterns in the data. The resulting patterns
are presented back to the user for analysis, whereupon the selected pat-
terns may be integrated back into the parsing and cleansing stages. The
next sections describe the algorithmic components needed to build a
data cleansing system (see Figure 14.3).
Original data |~ Title Mr. " " n

First Name David
Mr. David Loshin, Proj Manager
Knowledge Integrity Inc. Parsing Last Name Loshin
123 Main Street, Suite 3101 Position Proj. Manager
Anytown, NY 11223
Company Knowledge Integrity Inc.
Address 1 123 Main Street
Address2 Suite 3101 1
City Anytown
State NY
1 ZIP 11223
FIGURE 1 4 . 3 Example of record parsing

14.3.1 Parsing Metadata
We will need some metadata intended specifically for parsing rules. The
first metadata items we need are the data element types. In other words,
we want to know what kinds of data elements we are looking for, such
as names, street names, city names, telephone numbers, and so forth.
For each data element type, we need to define a data domain for
the vaUd values. This allows us to test to see if any particular data ele-
ment is recognized as belonging to a specific domain.
14.3.2 Token Parser
The next component we need for data parsing is the parser itself: a pro-
gram that, given a data value (such as a string from a data field), will
output a sequence of individual strings, or tokens. These tokens are
then analyzed to determine their element types. The token is at the same
time fed into a pattern analyzer as well as a value lookup analyzer.
14.3.3 Pattern Matching
A different kind of parsing is based on the format of the data as

opposed to the content per se. We can go back to the example described
in Section 7.5 of the Social Security number. We can specify that Social
Security numbers take the XXX-XX-XXXX format and use this struc-
ture format as a pattern for matching strings that we want to identify as
Social Security numbers.
In general, we want a format specification for pattern matching for
the purpose of parsing. When we evaluate strings, we can characterize
their similarity to known data element structure based on how well they
conform to defined structure patterns. This can be used for comparing
datafieldsthat we expect to have a pattern, such as telephone numbers,
addresses, and identification numbers (see Figure 14.4).
We use structure to infer meaning once we have imposed afirstcut
of attribution to our strings. Typically, we would run the data element
parser initially and assign data types to data elements before a pattern
comparison stage. We then can specify structural similarity through the
definition of patterns. Patterns can range from single character strings
DATA CLEANSING 347
AccountNumber SocialSecuritylD TelephoneNumber
f f f
DD-AAAAA-DD DDD-DD-DDDD l(DDD) DDD-DDDD
TaxID NYLicense ZIP+ 4
y r f 1
DD-DDDDDDD CDD DCC DDDDD-DDDD
FIGURE 1 4 . 4 Examples of pattern matching
to sequences of strings. Our Social Security number example is one kind

of single-string pattern, as are telephone numbers.
If we are parsing data items for element type membership by struc-
tural pattern, we need a program that will match a token to a defined
pattern. We can use our structural domain rules from Chapters 7 and 8
to describe structural patterns.
We should also note that while a data value may structurally
resemble a known format, the actual data within may be incorrect for
that form. For example, we may have a string "(513) 409-F5H5" that is
similar in format to a telephone number, but the value of the string is
not a telephone number.
A very popular way for performing pattern matching is using
applications known as parser generators. A parser generator takes as
input a description of a set of patterns and generates a program that
will recognize those patterns and report back which pattern was
matched. Parser generators are usually used in building programming
language compilers, but they can be used in any situation where a well-
defined language can be defined.
Consider the problem of freeform names appearing in a name data
field. Presuming that we can assign a data element type to each of the
strings, the next step would be to determine whether the entire string
matches a specific pattern. For example, if we were able to differentiate

between last names andfirstnames, along with initials, we might define
the following patterns for recognizing individual names.
(first name) (first name) (last name)
(first name) (initial) (last name)
(initial) (first name) (last name)
(last name) (comma) (first name)
(last name) (comma) (first name) (first name)
(last name) (comma) (first name) (initial)
While not all names will conform to these patterns, a large number
will. If we can distinguish malefirstnames from femalefirstnames, we
can use the pattern system to infer gender assignment as well.
14.3.4 Value Lookup
If the token does not match a particular pattern, then we will see if that
value exists in one of the predefined data domains associated with the
parsing metadata. Note that a token may appear in more than one data
domain, which necessitates the next component, probability assignation.
14.3.5 Probability Assignation
Whenever a token value is recognized as belonging to more than one

data domain, we must also assign some probability that the token
belongs to each recognized domain. We can do this based on some
heuristic rules, which may depend on the context from which the data
value was originally taken. For example, if the token was taken from
the name field, it is more likely that the token "HOWARD" is a first
name than a city name.
14.4 METADATA CLEANSING
We need to be careful when making use of metadata for data cleansing,

since if the metadata is not clean, the cleansing operation will effectively
fail. Metadata cleansing is a process of validating that the metadata that
we are using is correct.
DATA CLEANSING 349
Our cleansing metadata is essentially composed of data domain

definitions. A way to cleanse the metadata is to perform some analysis
on the conformance of data sets to the metadata. This process includes
the following.
Looking at how frequently the metadata is used to correctly parse
tokens
Determining if there are values included in the metadata domains
that never appear in the actual data
Determining if there are more than one overlapping data domains
(that is, do two or more domains share a large percentage of values?)
Determining if there are overloaded domains (that is, are a small
number of domain values accounting for most of the token matches?)
Checking if new data values emerging in the data belong to
known data domains but have not yet been assigned
Checking if new patterns emerging in the data require changes to
the metadata patterns
The cleansing metadata repository requires an occasional review by a
domain expert to make sure that the metadata is correct.
14.5 DATA CORRECTION AND ENHANCEMENT
The parsing stage identifies those data elements that are recognized as
belonging to a specific element type and those that are not recognized.
The next stage of the process attempts to correct those data values that
are not recognized and to tag corrected records with both the original
and the corrected information.
14.5.1 Automation vs Human Interaction
It is a noble goal to attempt to completely automate the data correction

process. Many data cleansing tools try to correct data based on long-
time experience and records of the types of corrections that analysts
perform while cleaning data. This accounts for large knowledge bases
of rules that are incorporated into these products and a proliferation of
obscure rules that reflect special cases.
Unfortunately, there are a few flaws in this approach. First of all,
the effect of accumulating correction rules based on analyzing certain
kinds of data will bias the corrective process to that kind of informa-
tion. Thus, if a company's set of correcting rules is based on direct mar-
keting databases, there may be an abundance of rules for correcting
individual names but a dearth of rules for correcting business names.
The second flaw is that every organization's data is somehow dif-
ferent from any other organization's data, as are the business rules that
govern the use of that data. Relying on the business rules from other
organizations will still add value, especially if the data content is simi-
lar, but there will always be some area where humans will need to inter-
act with the system to make decisions about data corrections.
The third flaw is that data can only be perceived to be incorrect
when there are rules indicating correctness. Again, if we rely on other
sets of correctness rules, we may miss errors in the data that may pass
through provided correctness tests. An example of this in address cor-
rection is the famous East-West Highway in suburban Washington,
D.C. Because the expectation with addresses with the word "East" at
the beginning is that the word is being used as a direction prefix and not
as part of the street name itself, some applications inappropriately "cor-
rect" this to "E. West Highway," which is not the correct name.
14.5.2 Correction
Obviously, if we can recognize that data is in error, we want to be able

to fix that data. There are a few different ways to correct data, and
these all rely on some sort of intelligent knowledge base of rules and
transformations or some heuristic algorithms for recognizing and link-
ing variations of known data values. One form of correction involves
finding string values that are similar to correct values but are slightly
varied. In context, an assumption can be made that the intent is to use
the correct form.
Another form of correction is the transformation from vanity
names and aliases to real names. A vanity name is a euphemized or
alternative name that is applied to an item with an already existing
name. Place names are frequent holders of vanity names, which are a
way to project an image different from the image associated with the
surrounding area. Aliases are similar and are also associated with loca-
tions that have more official names, such as buildings or office parks.
In general, the correction process is based on maintaining a set of
incorrect values as well as their corrected forms. As an example, if the
DATA CLEANSING 351
word "International" is frequently misspelled as "Intrnational," there

would be a rule mapping the incorrect form to the correct form.
Again, this should sound familiar. We can use the domain map-
pings we discussed in Chapters 7 and 8 to create the correction rules. In
these cases, our mappings are many-to-one mappings, where the source
domain is a set of incorrect forms and the target set is of correct forms
of value strings.
14.5.3 Standardizing
As we discussed in Section 14.1, standardization is the process of trans-

forming data into a standard form. In the cleansing environment, we
use standard forms as a preparation for consolidation, where we try to
link different records together. This is different from using standard
forms for data normalization. Instead, we use standardization as a
means of reducing the search space when looking to link entities
together.
For example, consider these two customer records.
1. Elizabeth R. Johnson, 123 Main Street
2. Beth R. Johnson, 123 Main Street
In the absence of any information to the contrary, our intuition
says that these two records represent the same person, since Beth is a
nickname for Elizabeth. This is an example of the use of a standard
form: Elizabeth is the standard for Beth, so we can augment the second
record with the standard form.
There are many different data element areas that carry with them
some intuitive set of content-related standardization rules. Each cul-
ture's set of first names comes along with variation forms, nicknames,
and so on that all can relate any name to at least one standard form. For
example. Bob, Rob, Bobby, and Robbie are all different forms of
Robert. Liz, Lizzie, and Beth, may all be nicknames for Elizabeth, We
can create a mapping from known names to some standard form that
can then be used as a linking value for later cleansing phases. This holds
true for other kinds of data as well: last names, business words,
addresses, industry jargon, and transaction types are just a few element
types where this technique can be used.
It doesn't even matter if the standard form is applied in a manner
that is not consistent with real life. For example, Beth Smith's name
might not be Elizabeth, but we can assign EUzabeth to the records any-
way just in case. In this context, standardization is being used purely as
a means to a different end: enhancement and linkage.
As we have seen in other sections, we can implement standardiza-
tion using the same kind of data domain and domain mapping method-
ologies from Chapters 7 and 8. As an example, we can create a data
domain for what we will consider to be the standard male first names
and a separate domain called "alternate names" consisting of diminu-
tive, shortened, and nicknames. This allows us to have a mapping from
the alternate names to the standard names, which we will use to
attribute each name record.
14.5.4 Abbreviation Expansion
Similar to standardization, abbreviation expansion is a rule-oriented

process that maps shortened forms to expanded forms. Again, the pur-
pose of expanding abbreviations is so that different records can be
merged or linked together.
Abbreviations come in different flavors. One type of abbreviation
shortens each word in a set to a smaller form, where the abbreviation
consists of a prefix of the original data value. Examples include "INC"
for incorporated, "CORP" for corporation, "ST" for street. Other
abbreviations shorten the word by eliminating vowels or by contracting
the letters to phonetics, such as "INTL" or "INTRNTL" for interna-
tional, "PRGRM" for program, or "MGR" for manager. We call these
kinds of abbreviations contractions.
A third form of abbreviation is the acronym, where multiple words
typically seen together are characterized with a single string, such as
"IBM" for International Business Machines, or "RFP" for request for
proposal.
Abbreviation expansion makes use of either predefined mappings
from the abbreviated form to the full form or is oftentimes ignored until
record linkage and data matching takes place. In thefirstcase, abbrevi-
ation is similar to standardization, but in the second case, more free-
form abbreviations (those that are not performed in a standard way,
such as contractions) can still be linked based on approximate matching
techniques.
DATA CLEANSING 353
14.5.5 Application of Business Rules
After the standard corrections, standardizations, and abbreviations

have been addressed, there is one more stage at which business rules
particular to the application domain may be applied. In our example
from Section 14.2.9 where we looked at the overloading of data values
within attributes, there were certain business rules within the domain.
There are a number of cases where business rules can be applied. In
the case we have discussed, the application is both to determine if there
is content overloading and if so to split the overload values out of the
single attribute. Another case for using business rules is filling in miss-
ing data fields. A third area is knowledge inferences, which we will
explore in Chapter 16. In any case, these business rules should be
expressible using the rules formats we have discussed in this book.
14.5.6 Enhancement
Data enhancement is the process of appending additional data to a

record that increases the value of the data. We talk about data enrich-
ment in Chapter 16 as a greater process of intelligently merging data
from different data sources, but in this section, we look at enhancing
the information embedded in a record by appending all the knowledge
inferred from the associated parsing and cleansing.
The data parsing process attributes a data element type to each
string appearing in a record's field. That information is communicated
to the analyst by reformulating the data in a presentation indicating the
data element type associated with each string. So if, for example, we
have determined that there is a business name embedded in an attribute
called name, then we will associate the recognized business name with a
tag indicating the "business-ness" of the name. As another example, if
we determine that the value of the city field is actually a vanity name,
we would append the official name of the address's city to the record as
well. Ultimately, our record is presented back as a set of element type,
value pairs.
By appending standardization information, we provide multiple
pivot points which help streamline the later data matching and record
linkage process. This lends an advantage, which we will see especially in
the duplicate elimination and merge-purge processes.
In more general terms, we are creating an associative representa-

tion of the data that imposes a metacontext on the data. By indicating
detail about the content of the data, we are promoting bits and bytes to
more concise information, which then allows us to exploit business rule
relationships that exists in the data. This allows more interesting infer-
ences to be made about the data, as we will see in Chapter 16.
If we have found an error in some value's form, or if we standard-
ize data values, we should not just append the corrected form but we
should also log the error occurrence and the rule used in correction or
standardization. We gain insight both by having the information repre-
sented in its parsed and standardized form but also through the analysis
of the process itself. Aside from having the data corrected, for example,
we can also measure how many different ways errors existed in the
original data set and how frequently each kind of error occurred. This
yields significant value, as we can analyze the root cause of the errors
and try to prevent them from occurring in thefirstplace.
14.6 APPROXIMATE MATCHING AND SIMILARITY
In all of our discussion so far, we have been gravitating toward a more

abstract concept that is inherent in the application of data cleansing. If
all the data were clean, the processes that depend on clean data would be
extremely straightforward. For example, eliminating duplicate entries is
simple if we are sure that all entries are 100 percent correct the dupli-
cates will be trivially found after sorting the data, for example. The need
for data cleansing is purely based on the fact that errors creep into the
data set and prevent the straightforward algorithms from working.
This implies that we must use a more complex means for finding
similar data values, not just identical values. We refer to this as "approx-
imate matching." Approximate matching functions take as arguments
two data values and return a score of how similar the two submitted val-
ues are. We can characterize an approximate match based on the func-
tion used for similarity scoring and the score returned.
14.6.1 Similarity Measures
In order to provide a scoring function for any approximate matching, we

need to define what similarity means. Similarity measures are based on a
DATA CLEANSING 355
distance function the closer the distance between two data values, the
more similar those two values are to each other. The simplest form of a dis-
tance function is a Euclidean distance on a Cartesian coordinate system.
For example, across a single integral dimension, we can compute a
distance between two integers as the absolute value of the difference
between the two values. In a plane, the distance between two points,
(xl, yl) and (x2, y2) is the square root of (xl - x2)^ + (yl - y2)^.
Even with nonnumeric data, we still have to formulate some kind
of quantitative similarity measure. Because a similarity measure yields a
score that measures the closeness of two data values, there must be
some way to characterize closeness, even if the values are character
strings. In general, we will want to be able to compute the similarity
between two multiattribute records, and that means that we must have
distance functions associated with all kinds of data types, as well as our
constrained data domains.
14.6.2 Difference and Similarity
We will use a definition of a difference function d(x, y) that maps a pair

of values of the same data type to a real value between 0 and 1.0, where
a difference of 1.0 indicates no similarity, and a difference of 0 indicates
an exact match. A difference function is symmetric, which means that
d(x, y) = d(y, x). Difference functions and similarity functions are essen-
tially inverses of each other. The measure of how similar two objects are
is the same as the inverse of how different they are.
The literature also imposes a property on difference functions
known as the triangular inequality, which states that d(x, y) + d(y, z) >=
d(Xy z) This basically says that given three objects, x, y, and z, the dif-
ference between x and z is less than or equal to the sum of the differ-
ences between each of those two objects and some third object. This
becomes very useful when clustering points within the dimensional
space, since any previously measured differences can be used to look for
matches within a certain area.
When comparing any two objects, we must compare their individ-
ual components. We can represent any data record or message as a set
of (attribute, value) pairs. Each attribute represents some dimension,
and each associated value is a specific point along that dimension. Simi-
larity comparison is then basically comparing the distance in some
multidimensional space (see Figure 14.5).
Source Record 1 Search Record Weight
Name 1 HDLoshin 1 David Loshin 1 Similarity is based oo approximate string match* 1 5
Address 1 123 Main St 123 Main Similarity is based on value parsing and component matching. 1 3
City 1 Anytown 11 Anytown 1 Similahty is based on equivalence (including vanity names). 1 2
State NY NY 1 Similarity * lif both are equal; 0 otherwise. 1 1
ZIP 11223-6523 11223 ZIP-f 4 exact match 1; ZIP match ZIP-f 4 s .85; ZIP s ZIP s .75. 2
1 Telephone 509-652-9822 509-652-9822 1 Similarity s lif both ace equal; 0 otherwise. 9
INC Y N 1 Similarity s lif both are equal; 0 otherwise. 1
FIGURE 1 4 . 5 Distance and similarity
For each data type or data domain, we assign a similarity function,

and for each set of data attributes, we also provide some weight to be
factored in when computing an overall similarity or difference. So,
given a set of data records with name, address, telephone number, and
Social Security number, we can configure a similarity function that is
the composition of the weighted similarity functions associated with
each of the components. When we use weighting, though, we must be
careful to provide a scaling factor as well so that the overall similarity
value falls between 0 and 1.0.
Simply, given a set of properties, p ^ , . . . , p^, a set of difference
functions rf^,.. . , J^ associated with each of the properties, and a set of
weights, M/Q, . . . , w^y the difference between any two vectors of values
XQ, . . . , x and y^,...,y^ is
'^Wi*d(Xi, yi)
/=0
i=0
This gives us a basis for defining similarity measures. The next step
is in defining the difference function for each particular data type. We
have already seen the use of Euclidean distance for numeric domains.
Next we explore similarity measures for different data domains.
DATA CLEANSING 357
14.6.3 Thresholding
When performing approximate matching, how do we determine what is

and is not a match? With exact matching, the answer is easy either
two values are the same or they are not. With approximate matching,
though, we are not given a definitive answer but rather some point
along a continuum indicating the degree to which two values match.
With approximate matching, it is up to the user to define the point
at which two values are considered to be a match. We would specify a
threshold score, above which the two compared values are considered a
match and below which the two values are not considered a match. We
can be more precise and actually define three score ranges: match, not a
match, and questionable. The threshold above which two values are
considered a match to 1.0 is the match range. The not match range
would be between 0 and a second threshold value, below which the two
strings are definitely not a match. The range between the lower thresh-
old and the higher threshold is the questionable range. Two values
whose similarity falls within the questionable range would be candi-
dates for human intervention a user might need to specify that the
two values under investigation are matches or not (see Figure 14.6).
This process of incorporating people into the matching process can
have its benefits, especially in a learning environment. The user may
begin the matching process by specifying specific thresholds. Because
the process integrates user decisions about which similarity values indi-
cate matches and which do not, a learning heuristic may both automat-
ically adjust the thresholds and the similarity scoring, yielding a better
similarity measurement.
14.6.4 Edit Distance
The first nonnumeric data type we will look at is the character strings.
Clearly, comparing character strings for an exact match is straightfor-
ward. It is the determination of closeness for character strings that
becomes hazy. For example, we know intuitively that the last names
Smith, Smyth, and Smythe are similar. Or do we? If we are not native
English speakers, or if we had never heard all three names pronounced,
how would we know that these names all sound the same?
One way to measure similarity between two character strings is to
measure what is called the edit distance between those strings. The edit
Field Name Threshold
1 Name 75% 1
1 Address 67%
1 City 50%
1 State 100%
ZIP 85%
1 Telephone 100%
1 INC 100%
Total Threshold
FIGURE 1 4 . 6 Using thresholds for the decision process
distance between two strings is the minimum number of basic edit oper-
ations required to transform one string to the other. There are three
basic edit operations.
1. Insertion, where an extra character is inserted into the string.
2. Deletion, where a character has been removed from the string.
3. Transposition, in which two characters are reversed in their
sequence.
So, for example, the edit distance between the strings "INTER-
MURAL" and "INTRAMURAL" is 3, since to change thefirststring to
the second, we would transpose the "ER" into "RE" and delete the "E"
followed by an insertion of an "A."
Some people include substitution as an automatic edit operation,
which is basically a deletion followed by an insertion. Strings that com-
pare with small edit distances are likely to be similar, whereas those that
compare with large edit distances are likely to be dissimilar.
DATA CLEANSING 359
Edit distance is computed using an iterative dynamic programming

method. Given two strings, x = x^x^ ^ and y = ^^^2 ^m^ ^^ ^^^^
distance /"(/, /) is computed as the best match of two substrings x^x^ . . .
jc. and y^^ y, where
/(0,0) = 0
M / ) = mm\f{i-\,]) + 1),/(/,/- 1) +!,/(/- 1,7- D + ^(^^^y)]
where d(x.,y.) is 0 if x. is the same as y. and is 1 otherwise. This recur-
rence relation defines the format for the dynamic programming frame-
work, where the function f is computed at each point while traversing
through both strings, looking for the minimal value of the function to
determine the next step to take. The result is the computation of the
minimum edit distance between the two strings (see Figure 14.7).
14.6.5 Phonetic Similarity
Another measure of similarity between two character strings is the pho-

netic similarity, or how much the two strings sound alike. In the early
INTRANATIOANL | ORGANASATION
transposition substitution
INTRANATIONAL | ORGANISATION
transposition substitution
INTARNATIONAL | ORGANIZATION
substitution
INTERNATIONAL | Edit distance = 2
Edit distance = 3
HOUSNGAUTHORTY PRODUICTIVTY
insertion deletion
HOUSING AUTHORTY PRODUCTIVTY
insertion insertion
HOUSING AUTHORITY PRODUCTIVrrY
Edit distance = 2 Edit distance = 2
FIGURE 14.7 Example edit distances

twentieth century, a system called Soundex was developed to help in the

indexing of U.S. census data. The system involves encoding a name
string as a four-character code. The goal is that similarly sounding
names would be encoded as the same or similar Soundex codes.
The Soundex method works Hke this: The first character of the
name string is retained, and then numbers are assigned to following
characters. The numbers are assigned using this breakdown.
1=BPFV
2 = CSKGJQXZ
3 = DT
4=L
5 = MN
6=R
All other vowels and consonants are ignored. If the same character
appears more than once in a row, the run is reduced to a single number
value. In traditional Soundex, the code is truncated after one character,
followed by three digits. If there are not enough digits, zeros are
appended. For example, the Soundex code for "Smith" is S530, and the
Soundex code for "Smythe" is also S530, which indicates that these two
names are similar.
Soundex is flawed as an approximate matching tool for the follow-
ing reasons:
The constraint of using the first letter followed by a phonetic code
means that names that sound alike but do not share the same first
letter would not necessarily match. Consider "Dixon" (D250) and
"Nixon" (N250), which can sound similar over a poor telephone
connection.
Because Soundex uses consonants in the name string, as opposed
to phonetic units, names that have "silent" letters or letter combi-
nations with different sounds will skew a match. Consider the
name "Loshin" (L250) and the word "lotion" (L350).
Because Soundex only uses the first letter and next three conso-
nants, longer names will be encoded in a truncated form.
Soundex is geared toward English names and is not effective when
used with non-English names.
There are other phonetic systems, such as the one developed for the
New York State Identification and Intelligence System (NYSIIS) coding.
NYSIIS is similar to Soundex, except that while Soundex removes all
vowels, NYSIIS retains at least the placement of vowels by replacing
DATA CLEANSING 361
them all with the symbol "A." Unfortunately, the NYSIIS encoding is
also faulty in that it is biased toward English names. Figure 14.8 shows
some Soundex and NYSIIS encodings.
A system called Metaphone, developed in 1990, claims to better rep-
resent English pronunciation. The Metaphone system reduces strings to a
one- to four-letter code, with some more complex transformation rules
than Soundex or NYSIIS. Specifically, Metaphone reduces the alphabet to
16 consonant sound: { B X S K J T F H L M N P R O W Y } , where the zero
represents the th sound. Transformations associated with character
sequences are defined based on the different phonetic constructs. For
example, a T can reduce to an X (representing a sh sound) if it appears in
the -TIA- or ~TIO- context or to a 0 if it appears in the -TH- context.
Otherwise, it remains a T.
There are some ways to improve on the efficacy of these phonetic
representations, especially when addressing the issues described here. In
terms of similar sounding words that do not share the same initial sound,
the phonetic encoding algorithm can be applied to strings in both the for-
ward (in English, left to right) and the backward (right to left) directions.
Word Soundex NYsns
LOSHIN L250 LASAN
JACKSON J222 JACSAN
LINCOLN L524 LANCALN
FINGERHUT F526 FANGARAT
DDCON D250 DAXAN
PETROFSKY P361 PATRAFSCY
BARTLETT B634 BARTLAT
FIGURE 14.8 Sample phonetic encodings

In this way, words that are similar, like "DIXON" and "NIXON," may
have a chance of matching on the reversed phonetic encodings.
The problem with Soundex's inabiUty to deal with silent letters or
alternate phonemes is to use a more complex algorithm such as Meta-
phone. The issue of the truncated encoding is easy to address: Encode
the entire string without truncation.
Unfortunately, it is difficult to overcome the fact that these pho-
netic schemes are biased toward English names.
14.6.6 N-Gramming
Another way to measure similarity between strings in the presence of

errors is using a technique called n-gramming. We can look at any string
as a composition of its substrings, grouped consecutively in discretely
sized chunks moving from left to right. We break a word into its n-
grams by sUding a window of size n across the word and grabbing the
n-sized string chunk at each step. The chunk size is determined by the
"n" in the n-gram; in other words, if we look at substrings of size 3, we
are 3-gramming. If we select chunks of size 4, we are 4-gramming.
As an example, consider di-gramming (w = 2) the word "INTER-
NATIONAL." We end up with the following digrams:
IN
NT
TE
ER
RN
NA
AT
TI
lO
ON
NA
AL
For a string of length 13, we would generate 12 digrams. If we
were 3-gramming, we would have these chunks:
INT
NTE
DATA CLEANSING 363
TER
ERN
RNA
NAT
ATI
TIO
ION
ONA
NAL
We can use the n-gramming technique as part of another similarity
measure. If two strings match exactly, they will share all the same n-grams
as well. But if two strings are only slightly different, they will still share a
large number of the same n-grams! So, a new measure of similarity
between two strings is a comparison of the number of n-grams the two
strings share. If we wanted to compare "INTERNATIONAL" with
"INTRENATIONAL" (a commonfingerflub,considering that "E" and
"R" are right next to each other on the keyboard), we would generate the
n-grams for both strings, then compare the overlap. Using n = 2, we have
already generated the digrams for the first string; the digrams for
"INTRENATIONAL" are:
IN
NT
TR
RE
EN
NA
AT
TI
lO
ON
NA
AL
The two strings share 9 out of 12 digrams 75 percent. For a two-
string comparison, this is a high percentage of overlap, and we might say
that any two strings that compare with a score of 70 percent or higher
are likely to be a match. Given two strings, X and Y, where ngram(X.) is
the set of n-grams for string X and ngram(Y) is the set of n-grams for
string (Y), we can actually define three different measures.
1. Absolute overlap This is the absolute ratio of matching

n-grams to the total number of n-grams. This is equal to
(2 (\ngram(X) n ngram(Y)\)) -^ (\ngram(X)\ + \ngram(Y)\)
2. Source overlap This is the number of matching n-grams divided
by the number of n-grams in the source string X.
(\ngram(X) n ngram(Y)\) ^ \ngram(X)\
3. Search overlap This is the number of matching n-grams divided
by the number of n-grams in the search string Y.
(\ngram(X) D ngram(Y)\) ^ \ngram(Y)\
When comparing longer strings or data fields that contain many
string tokens, we can try different percentage thresholds and the differ-
ent overlap measures, depending on the data set and the application.
14.6.7 Other Similarity Functions
Considering that we are able to define more constrained data domains,

we should also be able to define similarity methods specific to the data
domain. For example, if we have a data domain for U.S. states, we could
define a new similarity measure between values from that domain that is
based on geographical distance instead of lexicographic distance. As
another example, values within a color attribute might be scored for sim-
ilarity based on closeness within their composite red/blue/green values.
For similarity functions within enumerated domains, we can define
a mapping between a composed pair of domain values mapped to a
similarity score. For generated domains, functions can be defined for
the comparison of any two generated values. These similarity measures
will define a similarity score between any two domain values and can be
used in combination with the similarity or difference measurement of
two data multiattribute records, as described in Section 14.6.2.
14.7 CONSOLIDATION
Consolidation is a catch-all phrase for those processes that make use of

collected metadata and knowledge to eliminate duplicate entities, merge
data from multiple sources, and other data enhancement operations.
DATA CLEANSING 365
14.7.1 Scoring Precision and Application Context
One of the most significant insights into similarity and difference mea-
surements is the issue of application context and its impact on both
measurement precision and on the matching criteria. Depending on the
kind of application that makes use of approximate searching and
matching, the thresholds will most likely change.
As our first example, let's consider a simple direct mail sales pro-
gram. While our goal would be to find duplicate entries, if a pair of
duplicates is not caught, the worst that can happen is that some house-
hold might get some extra unwanted mail. In this case, we might prefer
that any borderline matches be assumed to be mismatches so our cover-
age is greater.
For our second example, let's consider an antiterrorist application
used to screen incoming visitors. If the visitor's name matches one of the
names on the list of known terrorists, the visitor is detained, and a full
investigation is performed to determine if the visitor should be allowed
into the country. In this instance, where safety and security are con-
cerned, the worst that can happen if there is a missed match is that a
dangerous person is allowed to enter the country. In this case, we might
prefer that the match threshold be lowered and any borderline matches
be brought to the attention of the examiners so as to avoid missing
potential matches (see Figure 14,9).
The basic application in both of these cases is the same (matching
names against other names), but the precision depends on our expected
results. We can group our applications into those that are exclusive
searches, which are intended to distinguish as many individuals as pos-
sible, and inclusive searches, which want to include as many potential
matches into a cluster as possible. The direct marketing duplicate elimi-
nation would be an exclusive application, while the terrorist applica-
tion is an inclusive application.
14.7.2 Duplicate Elimination
Duplicate elimination is a process of finding multiple representations of

the same entity within the data set and eliminating all but one of those
representations from the set. In some instances, such as with a primary
key in a relational database table, duplicates are not allowed, and so it is
imperative that duplicate records be found and reduced to a single entity.
Duplicate
Elimination
Direct Mail Match Duplicate Pairs
Mailing List Threshold =
93%
In a direct mail campaign, we keep the matching

threshold high for duplicate elimination because
we do not want to mistakenly eliminate any
individuals.
List of Known
Terrorists
List of People
Entering Potential
the Terrorists
Country Entering the
Country
When filtering out dangerous individuals, there is

a greater likelihood that a name may have been
altered to avoid detection. Therefore, we keep a
low threshold for matching to make sure we do
not miss any possible matches.
FIGURE 1 4 . 9 Different degrees of search precision
When duplicates are exact matches, they can be discovered through the
simple process of sorting the records based on the data attributes under
investigation. When duplicates exist because or erroneous values, we
have to use a more advanced technique such as approximate searching
and matching for finding and eliminating duplicates.
Duplicate elimination is essentially a process of clustering similar
records together, then using the three-threshold ranges described in Sec-
tion 14.6.3. Depending on the application, as we discuss in Section
DATA CLEANSING 367
14.7.1, the decisions about which records are duplicates and which are
not may either be made automatically or with human review.
14.7.3 Merge/Purge
Merge/purge is a similar operation to duplicate elimination, except that

while duplicate elimination is associated with removing doubles from a
single data set, merge/purge involves the aggregation of multiple data sets
followed by eliminating duplicates. Data from different sources will tend
to have inconsistencies and inaccuracies when consolidated, and there-
fore simple matching is insufficient during an aggregation phase. Again,
using approximate matching can be used to cluster similar records, which
again, depending on the application, can either have a reduction phase
automated or passed through human review (see Figure 14.10).
14.7.4 Householding
Householding is a process of joining a number of records into a single

set associated with a single household. A household could be defined as
a single residence, and the householding process is used to determine
which individuals live within the same residence.
Householding is more than just finding all individuals with the
same last name living at the same address. Associated with household-
ing is a more advanced set of knowledge, such as marital status, family
structure, and residence type (single versus multifamily home versus
apartment). As in other areas that we have discussed, the goal of the
application determines the result of the householding process.
For example, a mail-order catalog company might want to ensure
that only one catalog was being sent to each residence. In that case, the
householding process is meant to aggregate records around a particular
delivery address, attempting to recognize those names that all belong to
the same address, whether or not they belong to the same family. Alter-
natively, an application that is targeting only the teenagers in a house-
hold would want to identify all members of the same family as well as
each family member's role. A third application might be to find unmar-
ried couples living together. In each of these applications, the process is
similar, but the details of which attributes are used in the process may
differ. Figure 14.11 shows an example of householding.
1 David Loshin Project Manager L
1
1 H D Loshin Product Manager 1
1 John Franklin 509-555-1259 L]-l

" V

1 509-555-1259 John Frankhn DS-187932771
!
1
Merge/purge will link

these two records by the
telephone number and
merge the information
embedded in both
records.
Figure 1 4 . 1 0 Duplicate elimination and Merge/Purge
The general algorithm for householding naturally revolves around

the address.
1. All addresses are cleansed and standardized (more about this in
Section 14.9).
oo 00 00 00
\o\ ON ON ION
ON ON ON ON
P o o O
i
T-(
r"' rH
LH
>- i
Lsi o |j> V
=3
> > *> >
1 (2
1
(2

J3
z
C
0 1 C
(A
c 1
o
- ON ON
<N > O

O
NO
o\ ON a\ <N
o o o <N
1-H 1-t
o
o A
o A
o A
o o o
o\ ON ON 0^ o
o O o o
o o o 3
O
T X
.Si C 1c3
2
c
2 t-i
UH
o b
n\
^ u
OS
o
> 2
D
C9
2. Groups are clustered based on the same address.

3. The address is enhanced based on its location status. Some loca-
tion statuses include single-family home, two-family home, multi-
family dwelling, small apartment, large apartment, storefront,
business address, and so forth. This information is available for
enhancement from third-party sources.
4. The clusters now represent sets of entities residing at the same
location. If any distinction must be made based on individual
occupancy unit (apartment number, floor, suite, etc.), it is done at
this time.
5. Within each cluster, each entity is considered as to its relationship
to other entities within the cluster. Two entities are related if
some connection can be established between them using corrobo-
rating data. This might include sharing a last name, having the
same telephone number, or a third record that contains both
names. When two entity records are determine to be related, that
relation is documented as a virtual link between them.
6. As entities are linked via relationships, a set of mini-networks of
interrelated entities is formed. Within each set, the application
may attempt to assign one of a list of roles to each entity, depend-
ing on the application.
14.7.5 Consolidation and Currency
There are other applications that use a consolidation phase during data
cleansing. One application is currency and correctness analysis. Given a
set of data records collected from multiple sources, the information
embedded within each of the records may be either slightly incorrect or
out of date. In the consolidation phase, when multiple records associ-
ated with a single entity are combined, the information in all the
records can be used to infer the best overall set of data attributes.
Timestamps, placestamps, and quality of data source are all prop-
erties of a record that can be used to condition the value of data's cur-
rency and correctness. Presuming that we can apply the data quality
techniques prescribed in this book, we can quantify a data source's data
quaUty level and use that quantification to consolidate information.
DATA CLEANSING 371
14.7.6 Consolidation and Aggregation
Another application is aggregation based on standard reduction func-

tions across the data set. This includes counts, summations of individ-
ual columns, averages, maximums, and minimums, among others.
14.7.7 Network Analysis
A third consolidation activity is a more advanced form of the network

analysis described in Section 14.7.4. In general, in any transactional
environment, entities acting within the system must communicate with
each other. As a component of the consolidation process, one applica-
tion is the determination of which entities are related to other entities
within the system (see Figure 14.12).
This process can have one of a number of goals, including a search
for strongly related entities, inference of community, or determination
of a communication chain. A simple example is the analysis of inferred
"minicommunities" within a newsgroup environment. Within a single
newsgroup, there may be many threads to which the users may post
messages. Over time, an analyst may find that certain newsgroup regu-
lars may tend to cluster together in smaller groups the same people
always participate together within threads, and other groups partici-
pate in other threads.
The general method for this kind of analysis involves some
straightforward steps and some more complex ones. Our network will
consist of a set of individual entities connected with weighted edges.
The simple part is attaching an edge between any two entities when evi-
dence of a relationship exists. If an edge already exists between a pair of
entities, the edge's weight is increased. The more complex part of the
process involves the clustering of entities grouped via highly weighted
connections. Groups where all entities are fully connected (each entity
is connected to all others within the group) with highly weighted edges
are obvious clusters, but there may be groups where the connection
weights between different entities is higher or lower, which requires
heuristics to determine cluster membership.
In this example, each node represents a party, and a Hnk between two nodes indicates an
established connection between those two nodes. The shaded areas cover networks that
indicate a "minicommunity,** where (for the most part) each member has an established
link with all (or almost all) other members.
FIGURE 1 4 . 1 2 Network analysis
14.8 UPDATING MISSING FIELDS
One aspect of data cleansing is being able to fill fields that are missing
information. Recall from Chapter 8 that there are five types of empty
attributes:
1. No value There is no value for thisfield a true null.
2. Unavailable There is a value for thisfield,but for some reason it
has been omitted. Using the unavailable characterization implies
DATA CLEANSING 373
that at some point the value will be available and the field should
be completed,
applicable value.
value is not known.
The data cleansing process will not address categories 1 or 3, since
by definition there is no way to attack those missing fields. But with the
other categories, the reason for the missing value may be due to errors
in the original data, and after a cleansing process, there may be enough
information to properly fill out the missing field.
For unavailable fields, if the reason for the omissions has to do
with the dearth of data at the time of record instantiation, then the con-
solidation process may provide enough information leverage to supply
previously unavailable data. For unclassified fields, it may not be pos-
sible to classify the value due to erroneous data in other attributes that
may have prevented the classification. Given the corrected data, the
proper value may be filled in. For unknown attributes, the process of
cleansing and consoHdation may provide the missing value.
14.9 ADDRESS STANDARDIZATION
One last area of data cleansing is address standardization. Because

address data quality is a critical business operation in many businesses,
the ability to use address data as a data merging and consolidation
point is very important. While many countries have postal standards,
very few of these standards are actually enforced. In some countries, the
allocation of addresses is not well defined, but in the United States, both
the way that addresses are assigned to new buildings and streets and the
way that addresses are formed for mailing purposes are well defined.
In fact, the U.S. Postal Service addressing standards are extremely
well defined, and they give us a good example of a template for a data
standardization process. The standard is documented in the USPS Pub-
lication 28 and is very comprehensive. In this section, we will give a fla-
vor of the degree of detail expressed in this standard by reviewing some
aspects of the USPS addressing standards. In many cases, there are
acceptable forms and preferred forms, the use of which is suggested to

enable better automated address standardization. Figure 14.13 shows
these addressing standards.
14.9.1 Recipient Line
The recipient line indicates the person or entity to which the mail is to
be delivered. The recipient line is usually the first line of a standard
address block, which contains a recipient line, a delivery address line,
and the last line. If there is an "attention" line, the standard specifies
that it should be placed above the recipient line.
Secondary address designators may

ALPHA INCORPORATED Recipient Line
be placed on the delivery address line.
123 N ELM ST Delivery Address Line
The last line contains the city, state
SPRINGFIELD, OH 31338-0228 Last Line
(using the standard abbreviation),
and a ZIP + 4 code.
ATTN Jason Smith Attention Line

ALPHA INCORPORATED Recipient Line
An attention Hne is placed above
123 N ELM ST Delivery Address Line the recipient line.
If all the information for the delivery

ALPHA INCORPORATED Recipient Line address cannot be continued in the
SUITE 3100 Alternate Location delivery address line, the secondary
Delivery Address Line address information can be placed
123 N ELM ST
in an alternate location line above
the delivery address.
Figure 14.13 The standard address block

DATA CLEANSING 375
14.9.2 Delivery Address Line
The delivery address line is the line that contains the specific location
associated with the recipient. Typically, this line contains the street
address, and should contain at least some of these components:
Primary Address Number This is the number associated with the
street address.
Predirectional and Postdirectional A directional is the term the
Postal Service uses to refer to the address component indicating direc-
tional information. Examples of directionals include "NORTH,"
"NW," "W." The predirectional is the directional that appears before
the street name; the postdirectional is the directional that appears after
the street name. While spelled-out directionals are accepted within the
standard, the preferred form is the abbreviated one. When two direc-
tionals appear consecutively as one or two words before or after the
street name or suffix, the two words become the directional, the excep-
tion being when the directional is part of the street's primary name.
When the directional is part of the street name, the preferred form is not
to abbreviate the directional.
Street Name This is the name of the street, which precedes the
suffix. The Postal Service provides a data file that contains all the valid
street names for any ZIP code area.
Suffix The suffix is the address component indicating the type of
street, such as AVENUE, STREET, or CAUSEWAY, for example. When
the suffix is a real suffix and not part of a street name, the preferred form
is the abbreviated form. The standard provides a table enumerating a
large list of suffix names, common abbreviations, and the preferred stan-
dard abbreviation.
Secondary Address Designator The secondary address unit desig-
nator is essentially a more precise form of the address, narrowing the
delivery point to an apartment, a suite, or a floor. Examples of sec-
ondary unit designators include "APARTMENT," "FLOOR,"
"SUITE." The preferred form is to use the approved abbreviations,
which are also enumerated in Publication 28.
Additionally, there are other rules associated with the delivery
address line. Numeric street names should appear the way they are
specified in the Postal Service's ZIP + 4 file and should be spelled out
only when there are other streets with the same name in the same deliv-
ery area and spelling the numeric is the only way to distinguish between
the two Streets. Corner addresses are acceptable, but it is preferred to

use the physical numeric street address. There are also rules associated
with rural route delivery addresses, military addresses, post office
boxes, Puerto Rican addresses, and highway contract route addresses.
14.9.3 Last Line
The last line of the address includes the city name, state, and ZIP code.
Besides the dash in the ZIP + 4 code, punctuation is acceptable, but it is
preferred that punctuation be removed. The standard recommends that
only city names that are provided by the Postal Service in its city state
file be used (this addresses the issue of vanity city names).
The format of the last line is a city name, followed by a state abbre-
viation, followed by a ZIP + 4 code. Each of these components should be
separated using at least one space. The standard also prefers that full city
names be spelled out, but if there are labeling constraints due to space,
the city name can be abbreviated using the approved 13-character abbre-
viations provided in the city state file.
14.9.4 Standard Abbreviations
The Postal Service provides in the appendixes to Publication 28 a set of

enumerations of standard abbreviations, including U.S. State and Pos-
session abbreviations, street abbreviations, and common business word
abbreviations. Each of these tables can be embodied as enumerated
data domains, and we can establish the standard abbreviations as a
mapping from the common word domain to the corresponding stan-
dard abbreviation table. Other standard abbreviations, such as for city
names, are included in the city state file, which can be purchased from
the Postal Service.
14.9.5 ZIP+ 4
ZIP codes are postal codes assigned to delivery areas to improve the
precision of sorting and delivering mail. ZIP codes arefive-digitnum-
bers unique to each state, based on a geographical assignment. ZIP + 4
codes are a further refinement, narrowing down a delivery location
within a subsection of a building or a street.
DATA CLEANSING 377
ZIP + 4 codes are essentially represented as a mapping from a

domain consisting of a composed street address and city name to a
domain of ZIP + 4 codes. These data are available for purchase from
the Postal Service.
14.9.6 Business-to-Business Data Elements
Business-to-business addressing is more complicated than individual

addressing, since more information is embedded in a business address.
Consequently, the postal standard explores the complexity of business
addressing by defining a set of data element types that can be included
in a business address, many of which are irrelevant when it comes to
mail delivery! These data element types are essentially a superset of the
components of an address described earlier in this section. Here is a full
listing, although we only elaborate on those that we have not yet dis-
cussed.
Name prefix (such as MR, MRS, DR)
First name
Middle name or initial
Surname
Suffix title (such as maturity (JR, SR) and professional (PHD,
DDS))
Professional title (this includes job titles, such as PROJECT
MANAGER)
Division/department name (such as ACCOUNTS PAYABLE)
Mailstop code (this is usually associated with large office environ-
ments)
Street number
Predirectional
Street name
Street suffix
Secondary unit indicator
Secondary number
Company name
PO box number
City
State
ZIP/ZIP + 4
Carrier route code. The carrier route is the Ust of addresses to

which a carrier deUvers mail. A code is assigned to each individual
route. Addresses with a carrier route code may qualify for dis-
counts on postage.
Operational endorsement. An endorsement is a marking within
the address indicating handling instructions, a special service, or a
request for service. An example of an endorsement might indicate
that delivery of the piece must take place within a certain time
period.
Key line code. The key line code is optional information printed in
the address and may contain information about a piece's presort
level or just include other information about the piece.
POSTNET barcode. POSTNET is an encoding system that
encodes delivery point and ZIP + 4 data as a barcode.
POSTNET address block barcode.
Business addresses will usually contain business words, whose
presence can tip off the data cleansing application as to whether the
address is associated with a business or not. The standard suggests that
business words not be abbreviated, but when label size restrictions indi-
cate that space is needed, the preferred method is to abbreviate the last
business word, using those approved abbreviations enumerated in the
appendix of Publication 28.
14.9.7 Address Standardization
The Postal Service gives this very straightforward definition of a stan-

dardized address, which is taken directly from USPS Publication 28: "A
standardized address is one that is fully spelled out, abbreviated by
using the Postal Service standard abbreviations shown in Publication 28
or as shown in the ZIP + 4 files." The implication, though, for the
process of data cleansing is that there must be a method to convert a
nonstandard address into one that meets the USPS definition.
What this really means is that any address standardization process
must include these components:
A method to verify that an address is in standard form
An address data element parser and tokenizer
A method to disambiguate data elements that might cause a con-
flict (such as the use of directionals as part of a street name or the
DATA CLEANSING 379
difference between the use of the abbreviation ST for street or

saint)
A method to perform the necessary lookups in the Postal Service
data files (ZIP + 4 and the city state file)
A method to reformulate the data address elements in a standard
form
14.9.8 NCOA
It is said that 20 percent of the United States population changes addresses

each year, making it very difficult to maintain up-to-date address lists. For-
tunately, the Postal Service provides National Change of Address (NCOA)
data to a number of licensees who then incorporate that data into either
application services or packages. This allows an aspect of data cleansing in
terms of currency to be done on name/address records.
14.10 SUMMARY
This chapter, one of the most comprehensive in this book, focused on

many aspects of data cleansing the process of evaluating, validating,
and correcting incorrect data. We began the discussion with a discus-
sion of standardization benefits of using a data standard and the
process of defining a data standard and conforming to it.
We then explored some common error paradigms the sources of
different kinds of errors, including attribute granularity, format confor-
mance, and a discussion of semistructured form. Other common errors
includefingerflubs,transcription errors, transformation errors, and the
problems associated with information appearing in the wrong context.
We then began to explore the process of data cleansing, including
record parsing, pattern matching, and probability assignation. We also
looked at the automation of data cleansing applications, including data
correction, data standardization, expansion of abbreviations, and the
application of business rules.
It became clear that a large part of the data cleansing operation
involves approximate matching. The ability to find errors in data may
be contingent on distinguishing between good and bad data, and that
means being able to match up incorrect data with its correct equivalent.
Approximate matching requires some kind of similarity and difference
metrics, which provides a framework for measuring the similarity

between any two values. The simplest form of similarity measurements
involve numeric distance. For string data, we looked at a number of dif-
ferent ways of measuring similarity, including edit distance, phonetic
similarity, n-gramming, as well as similarity functions defined on user-
defined domains.
The consolidation phase of data cleansing looks at the precision
used for assigning thresholds for matching based on similarity scores. If
the application requires that all matches be found, the threshold for a
potential match may be set relatively low. On the other hand, if the
application's penalty for missing a match is low, then the threshold
should be set high. Approximate matching can be used during the con-
solidation phase for eliminating duplicate records, merging data from
multiple sources, and other applications such as householding and role
assignment.
Finally, we considered the process of address standardization,
using the United States Postal Standard as the basis for standardization.
We looked at some important points from the standard and discussed
the application components that must be available for automated
address standardization.
Obviously, data cleansing is an important part of information
management, although we would hope that a good data quality pro-
gram couldfinessethe need for ongoing data cleansing.
15
ROOT CAUSE ANALYSIS
AND SUPPLIER MANAGEMENT
Being able to identify the location and magnitude of a problem is one

thing, but being able to figure out what is going wrong is another. Up to
this point, we have focused on discovering the locations of problems. In
this chapter, we concentrate on figuring out the source of the problem.
We will do this with root cause analysis, an iterative process of asking
questions that exposes the actual cause of a problem.
For data quality issues, root cause analysis is performed as a
sequence of stepping backward through the information chain to look
for the source of a problem. During this process, we may find multiple
variations on the same problem that all stem from the same root cause.
If we can trace the problem to its actual source, we can address all the
variants associated with that problem at once by fixing the problem,
not the symptoms. The next step is to understand what is causing the
problem by looking at how the processing stage or communications
channel affects the data and proposing a correction.
Not all problems originate within the organization. In many com-
panies, a significant amount of data is provided by external suppliers.
These suppliers may not always address the issue of data quality as
carefully as is done within the company, but the data quality problems
must still be addressed. When the root cause of the problem can be
tracked to data that are being supplied by an external vendor, we apply
a technique called supplier management. While it might appear that in
this situation, the quality of the data appears to be beyond your con-
trol, it is possible to take a direct approach of managing the relation-
ship with the data supplier based on concrete measurements and
statistics, leading to the establishment of Service Level Agreements.
381
This technique is a useful way of addressing the problem of low-quality

supplied data.
15.1 WHAT IS ROOT CAUSE ANALYSIS?
When we build a database or write a program, it is possible for small

implementation errors to creep into the system. When the implementer
is made aware of the problem, he or she must find and fix the error; this
process is called debugging. Root cause analysis is like application
debugging, but in our case it is applied to a bug in the data.
More formally, root cause analysis is an iterative process of analyz-
ing a problem to determine the actual cause. The most effective ques-
tion that can be asked in this process is "Why.^" At each iteration, we
ask how and why the problem appears, and the answers to these ques-
tions usually can be subject to the same questions, until the source of
the problem isfinallyrevealed.
15.1.1 Symptoms vs Cause
Most of the time, when a problem is detected based on some set of

symptoms, the symptoms are treated, but the root cause of the problem
is ignored. Obviously, when this happens, there is an appearance of
having "fixed" the problem, but the problem usually resurfaces later. In
some cases, the problem will resurface elsewhere in the system under
someone else's jurisdiction. In addition, treating the symptoms some-
times creates other system failures.
Certainly, treating the symptoms is not a scalable solution. As the
system grows, the percentage of errors associated with root causes will
grow as well, sometimes linearly and sometimes even exponentially.
Root cause analysis is a process to iteratively track back the problem to
its earliest appearance and determine the root cause at that point. To
demonstrate the usefulness of this process, let's look at a simple example
dealing with data quality.
15.1.2 Example
Consider a product pricing system in a department store. Each week,

new prices are advertised in a shopping circular that is distributed with
RCX3T CAUSE ANALYSIS AND SUPPLIER MANAGEMENT 383
the newspaper. Supposedly, the correct price for each product is posted
within the store at each product display. But when a shopper brings the
products to the checkout counter, it turns out that many products do
not ring up at the advertised sale price.
If the shopper is keeping track of the prices as they are being rung
up, it is possible to catch these errors at checkout time. When the shop-
per alerts the cashier that an incorrect price has come up, the cashier
must call for an attendant to go to the product display or the weekly cir-
cular and confirm the price. This usually takes time, but when the cor-
rect price is found, the incorrect price is voided, and the correct price is
keyed into the register. This may happen two or three times per pur-
chase. The result of this process is that the amount of time it takes for
the shopper to check out and pay for the purchase increases, resulting in
longer lines. When the store manager sees this, he instructs another
cashier to open up a new register.
If the shopper does not catch the error until after the payment has
been made, the procedure is a bit different. The shopper must go to the
customer service line and wait until a cashier can evaluate the price dif-
ference. At that point, an attendant is sent to do the same price check,
and if there is a price discrepancy, the shopper is reimbursed the over-
charge. Typically, this line is very long because aside from the multitude
of incorrect charges, there are other customer services being provided at
this location also.
The cashier's process of crediting the incorrect price and then
directly typing in the correct price is one example of treating the symp-
tom. The problem, to the cashier's eyes, is an incorrect charge that must
be fixed. Unfortunately, the cashiers are usually not in a position to cor-
rect the product price in the store database. Also, since they are paid
based on the number of hours they work and not as a function of the
number of customers they service, there is no compelling business rea-
son to care how long it takes to process each customer.
The fact that longer lines precipitate the manager's opening of
more registers is another example of treating the symptom. In the man-
ager's eyes, the problem, which we could call a throughput problem
(slow-moving lines), can be fixed by creating additional bandwidth
(additional registers), but it does not occur to the manager to examine
the root cause of the low throughput. If he did, though, he might notice
that approximately 15 percent of the purchased articles ring up at an
incorrect price.
We can see a number of important impacts from these incorrect
prices.
Customers are getting irritated that they have to wait so long in

Une. This may lead to customer attrition and blockading.
There are likely to be as many undercharges as overcharges, but
the customers are more likely to speak up about the overcharges
and ignore the undercharges. This is a direct cost to the depart-
ment store in the form of a loss in revenues.
More employees are needed to support the cashiering process and,
therefore, cannot be employed in more productive pursuits. This
includes the additional cashiers needed to cover the registers, and
the attendants to scout out the actual prices.
The next step would be to ask why the prices in the computer are
not correct. According to some of the cashiers, sometimes the prices are
not updated until after the circular has been released. Sometimes the
cashiers have no excuse. It is now up to the manager tofigureout when
the correct prices are supposed to be entered into the system and make
sure that the data entry is performed on time and correctly. This is prob-
ably the root cause of the problem, and addressing the root cause will
eventually solve the overall problem of low throughput.
m ike fuoHcUiatifm
process Associate notifies
manager of
difoence in price
between display
andcompiHsr
FIGURE 1 5 . 1 The department store example

ROOT CAUSE ANALYSIS ANO SUPPLIER MANAGEMENT 385
15.2 DEBUGGING THE PROCESS
We can think of root cause analysis as a process of detective work.

Essentially, our catalog of data quality problems represents a set of
clues, only a subset of which actually contributes to the ultimate prob-
lem. In general, we follow these steps in determining a root cause.
1. Gather data. This is discussed in this section.
2. Identify the key causal factors. We will look at this in Section
15.2.1.
3. Chart the causal factors. We look at this in Section 15.2.2.
4. Determine the root cause of the causal factors. This is covered in
Section 15.3.
5. Develop corrective actions. This is discussed in Section 15.4.
We have actually put into place all the factors needed to gather
data, ranging from the definition of expectations (Chapters 8, 9, and
10) and the insertion of measurement of key data quality assertions. In
addition, we can construct a reconciliation system into which noncon-
forming data is shunted, enhanced with the name of the assertion that
was violated.
We can break the analysis into two parts. Thefirstpart is tracing
the problem backward in the system until the problem occurrence point
is found. The second part is determining what caused the problem in
thefirstplace and devising a solution. In this section we will explore the
first part. We start with the information collected through our assess-
ment and measurement process and then trace the problem back
through the information chain until wefinditsfirstappearance.
15.2.1 Our Starting Point: Reconciliation Records
In a large system with many problems, the first question to ask is

"Where to begin?" A multitude of problems poses a daunting task for
determining root causes. In Chapter 6, we introduced the Pareto Princi-
ple in which the claim is made that 80 percent of the effect is caused by
20 percent of the causes. Our plan is to build a Pareto chart, which
highlights those areas that are responsible for the greatest percentage of
a problem and which variables are involved in those areas.
We can implement this by making use of our rules engines coupled
with our data quality rules to create a reconciliation system. The goal of
the reconciliation system is to automatically collect information about
nonconformance in the system. The resulting information collected can

be fed into a histogram, sorted by violated assumption, which will act
as the Pareto chart.
The system is composed of three parts (see Figure 15.2).
1. A reconciliation database table
2. A set of data quality rules
3. An executable rules engine
The reconciliation database is a repository for records denoting the
failure of tested records to conform to the specified rules. Each rule has
a unique identifier and can be referenced through that identifier. The
reconciliation table maintains a reference to the data being tested and
the rule that failed.
/ Data \
( Consumption J
Invalid data shunted

Reconciled data to reconciliation database
reinserted into process
FIGURE 1 5 . 2 The reconciliation system

ROOT CAUSE ANALYSIS AND SUPPLIER MANAGEMENT 387
create table recon(

reconciliationID integer,
tableName varchar(lOO),
recordID integer,
violatedRule integer
);
We can use the data quality rules defined in Chapters 7 and 8, or
we can rely on the kinds of measurements performed for the current
state assessment as described in Chapter 9. The rules can be incorpo-
rated into a rules engine as described in Chapter 12.
Every time a data record is passed through the rules engine, the
pertinent rules are fired. If the data in the record violates one of the
rules, this record is said to be nonconforming, and a reconciliation
record is created, incorporating the name of the table, the record being
tested, and the rule that failed.
When we have completed some amount of validation, our reconcil-
iation database will have a collection of references to records that do
not conform to our assertions. The reconciliation database provides
two benefits. The first is that analysts wishing to correct database
records have an enumeration of the incorrect records and the reason
that each failed (that is, tagged back to the violated rule). The process of
correction is simplified because the reason for the nonconformance is
highlighted.
The second benefit is that the reconciliation records can be sorted
according to the rules that were violated, and this can be used as input
to our Pareto chart. We can group nonconformities by count by vio-
lated rule, which we can also weigh based on business factors to pro-
vide the histogram representing the problems that call out for attention.
The fact that we have attached to each nonconformance the name
of the rule that was violated means that we have a head start on identify-
ing the causal factors. The dependent data values described in the vio-
lated rules are the targets for analysis, and aggregating nonconformance
by rule violation not only provides a place to start in the analysis but also
points us in the right direction for tracing back the nonconforming data.
15.2.2 The Problem Trace-Back
The next step is charting the causal factors. In our case, the causal fac-
tors are likely to be manifested as incorrect data values, and to chart
these factors we must trace the provenance of the incorrect data back to
its origin. Luckily, we already have a road map for this trace the
information chain.
The Pareto chart will direct the selection of a particular problem to
address. Once this selection has been made, the way to proceed is to
start tracing the problem backward through the information chain. At
each point in the information chain where the offending value can be
modified, we can insert a new probe to check the right values' confor-
mance with what was expected.
There are two important points to note. The first is that since the
value may be subject to modification along the information chain, the
determination of what is a correct versus incorrect value is needed at
each connecting location (such as the exit from a communication chan-
nel and entrance to a processing stage) in the information chain. The
second is that at each location in the information chain, any specific
data value may be dependent on other values sourced from other loca-
tions in the chain. If one of those source values causes the nonconfor-
mity, it is that value that is picked up as the one to continue with the
trace-back.
15.2.3 Identifying the Problem Point
Eventually, this kind of value tracing will result in finding a point where
the value was correct on entry to the location but incorrect on exit from
that location. We can call this location the problem point because that is
the source of the problem.
Finding the problem point is also interesting if we were to recali-
brate some of our other analysis processes. As we traced back the value,
we may have encountered the source of other problems along the infor-
mation chain. For example, the appearance of an incorrect product
price might affect the customer checkout process but also the daily
bookkeeping process. This means that what we originally considered to
be two problems is actually only one problem, and the impacts and
costs associated with the two problems can be aggregated.
As we trace backward through the information chain (see Figure
15.3), we can decide whether we want to move forward along other
paths and check the effect of the traced nonconformity. This allows us to
subsume both the impacts and the costs associated with low data quality
(as we analyzed using our tools from Chapters 4, 6, and 9) at earlier
FIGURE 1 5 . 3 Tracing the problem back through the information chain
points in the information chain. This "cost accumulation" determines

the recaUbration and may also point out that many significant problems
associated with low data quality can be attributed to a few root causes.
15.3 DEBUGGING THE PROBLEM
Having narrowed the analysis to the problem point, the next step is to
determine the cause of the problem. We can segregate problems into
chronic problems (those that have been around for a long time and
ignored) and acute problems (those that have cropped up recently and
are putting new pressures on the system). Depending on the type of
problem, there are different approaches to debugging.
Often, organizations choose to address chronic problems only
when the cost of the solution is disproportionate to the growth of the
system. On the other hand, organizations address acute problems
because their appearance disrupts normal operations. Debugging the
problem may be as simple as determining that incorrect data are being
input, but this may be difficult when faced with legacy application code
that is not actively being supported.
Most likely, the root cause falls somewhere in between. To track

down the source of a chronic problem, it is beneficial to ask some of
these questions.
How long has this been a problem? This will provide some insight
into the history of the problem, who was working on the applica-
tion or data when it first appeared, or whether anyone really
knows when the problem first appeared.
How long has the code at this processing stage been in produc-
tion? The answer might help in determining whether this problem
has always been there or is not related to the implementation.
What has prevented any previous root cause analysis? If there has
never been a good reason to track down this problem before,
what has changed in the environment to increase this problem's
priority?
If the problem is an acute problem, it may be due to some recent
change in the system environment. These questions may be helpful.
How recently has the application code been modified? If the
change in the application code correlates to the appearance of the
problem, that is a good place to start looking.
What areas of the code have been changed? Again, this focuses on
the source of the problem.
Has the input to this processing stage changed? Perhaps it is not
the processing that is flawed but the source data. If so, that may
be the focal point.
Has the delivery of data to this processing stage been disrupted? It
might be that neither the code nor the data have changed but the
process by which the data is provided to the code.
Are other data sets upon which this processing stage depends
made available? For example, if the data input is sourced from a
database that was traditionally updated before a certain time and
the scheduling of the updates changed, the processing stage may
be using out-of-date input data.
Has the supplier of information into this processing stage changed
recently? Perhaps the new supplier's data is not of as high quality
as the previous supplier's data.
Has the business process changed recently? In this case, it is the
fact that what was expected out of the application changed, but
perhaps the appUcation and/or the data have not been brought
into synchronization with the business change.
The answers to these questions will lead to other questions. The
debugging process is one of narrowing down the set of possible causes
by the process of elimination. Eventually, the precise point of the non-
conformity will be isolated, and that leads to the next question: How do
we fix the problem?
15.4 CORRECTIVE MEASURES RESOLVE OR NOT?
Now that the problem location has been isolated, it is the job of the
analyst to work with the system engineers to determine what needs to
be done to correct the problem. If we go back to our department store
example, we see that the location of the root cause was at the product
price update process. If the product price database were updated at a
consistent time with the correct prices, the effect would be seen at the
cash registers.
We have two goals at this point: further analyze to propose a cor-
rection to the problem and register what we have learned, and then
decide whether to implement the correction. Knowledge about the trace
of the problem backward through the system and the accumulated
knowledge about the specifics of the problem and suggestions for cor-
rection both fall into the area of enterprise knowledge that should be
absorbed under enterprise management. This information should be
registered within a centralized repository, so if the decision is made not
to correct the problem at this time, the history of that decision can be
reviewed at a later date.
15.4.1 THE CORRECTION REPORT
The collected root cause analysis information should be represented in

a format that enables the problem's correction. The most efficient
method to effect a correction to a problem is to provide a correction
report that completely specifies these elements (see Figure 15.4).
A concise description of the problem. This can be used as a sum-
mary for high-level review. The bulk of the other components of
the report will supply a detailed description.
High-Level Description
Characterization
Local Information
Chain
Suggested Correction
Date
Analyst
Impact Table Impact Cost
|;;fs:^;?|f'i-|''::;S-^^
FIGURE 1 5 . 4 The correction report template

A characterization of the problem. Is it an error associated with

incorrect input data, or is there an incorrect computation embed-
ded in the code?
The location of the problem. We have just pinpointed the location
of where the errors or nonconformities first occur, both within the
information chain as well as inside a particular data set or code
module.
The suggested correction. This captures the knowledge learned
from the root cause analysis process and provides insight to the
next set of interested parties in how to proceed.
The date of the analysis
The name of the person providing the report
A description of the impacts and costs associated with this prob-
lem. This is a summary of the results of the trace-back process,
where information is collected about the overall impact of specific
nonconformities.
This information can be embedded in its own set of database
tables, which provides a way to archive the analyses.
create table problemReports (
reportid integer,
reportDate date,
shortDesc varchar(lOOO),
problemType integer,
location integer,
correction text,
analyst varchar(100)
);
The problemType attribute refers to a domain of defined problem
types. An auxiliary table logs the impacts.
create table problemlmpacts (
reportID integer,
impactType integer,
cost float
);
The impactType attribute refers to a domain of defined impacts, sim-
ilar to those described in Chapter 4. The cost attribute can be used to tally
the individual associated costs by correction report or by impact type.
15.4.2 Resolve to Resolve
In our department store example, the corrective measure was described

relatively easily, but the description of the correction does not necessar-
ily justify its implementation. There is actually a decision point when
we reach this stage for a number of reasons, including the following.
The cost offixingthe problem is greater than the continued treat-
ment of the symptoms.
The resources required are not available.
It is not always clear how to solve the problem or whether it is
worthwhile.
Taking a system offline to correct it may result in significant losses.
As before, the ultimate decision to make the correction is based on
a return on investment justification. The projected cost for making the
correction needs to be offset against the continued cost of living with
the problems, and a decision can be made based on whether and when a
breakeven point occurs.
To extend our management of the correction reports, we should
also keep track of the decision of whether or not to implement the cor-
rection, and if so, the party responsible for that correction.
15.5 SUPPLIER MANAGEMENT
It is possible that the trace-back determined that the problem originated

outside the organization. This means that the problem actually origi-
nates with a supplier to the organization, either as an external data
provider or as an external application provider. At this point, we can
address the problem through the process of supplier management.
SuppUer management is a term that covers the range of interactions
between a product or service vendor and a purchaser. Vendor-purchaser
relationships are forged before the sale is made, but afterward there is
likely to be less oversight applied to the supply process. The result is that
the level and quality of the provided service may suffer in the long run.
As with all manufactured goods, information products and information
services are subject to quality controls. It is just that few organizations
implement these controls. This is due to the same reasons that internal
information quality is not measured, as discussed in Chapters 3 and 4.
The key to supplier management is being able to specify a set of

requirements that must be met by the supplier in order to satisfy the
purchaser. The same way that an automobile manufacturer can set
specifications on the products that are produced by OEMs for inclusion
in a finished automobile, information-based businesses can set specifi-
cations for supplied information. For simplicity, our discussion will
focus on the data supplier, but the process can be applied to service
providers in a similar fashion (see Figure 15.5).
A company can set those specifications the same way internal
requirements are defined: through the use of defined metrics based on
the dimensions of data quality. We can use our data quality rules to
specify assertions and expectations about the data as it enters the enter-
prise. The only difference is where the data is being tested: after it has
All supplied data

ts suiêa to a
validation tst
before it is
imported into the
enterprise.
Enterprise
Data Validation
Buffer Zone
Data is only gated

dbfottgbifit
conforms to die
esqpected levels of
data quality.
FIGURE 1 5 . 5 Enforcing rules at the supply site

been injected into the information chain or before it enters the system.
Obviously, it is better to prevent bad data from entering the system than
having to deal with it once it is in the system.
Supplier management is best implemented as a "compliance codi-
cil," which is included as a component of the original agreement between
the supplier and purchaser. The agreement should include a provision for
a periodic review of the levels of data quality, a minimum threshold for
acceptance, and a definition of how the threshold is computed. The com-
pliance codicil should also specify the penalty when the data provided by
a supplier do not meet the purchaser expectations, as well as a set of
actions to be taken whenever a serious violation of the requirements has
occurred.
The threshold can be expressed as a function of the measured con-
formance to the set of requirements. In Chapter 10, we looked at the
definition of data quality requirements, as well as a way to collect and
manage requirements in a set of metadata tables. Each requirement is
specified as a rule, along with a description of that rule's implementa-
tion. It is the responsibility of the purchaser to provide a copy of these
rules to the supplier so that the desired levels of data quality are under-
stood equally well by both supplier and purchaser.
One interesting way to make sure that both sides are complying
with the agreement is for each side to develop the testing application for
requirements testing, agree that each implementation correctly validates
the information, and then exchange those applications. This means that
when the supplier is running its validation engine, it is using the appli-
cation written by the purchaser, and vice versa.
15.6 SUMMARY
In this chapter, we looked at completing the data quality improvement

cycle. We have looked at analyzing the economic impact of low data
quality and the costs associated with those impacts. We are able to
define sets of data quality rules and perform a current state assessment
based on measuring conformance to our expectations with respect to
the dimensions of data quality. We can use the results of these measure-
ments to create a requirements document. We can analyze our data sets,
looking for reference data and metadata, and we now can analyze the
system tofindthe root cause of data quality problems.
We can now provide a full analysis of the data quality of our sys-
tem and create a framework for selecting and implementing improve-
ments in implementable units of work. In the next chapters, we look at
improving the value of our data through enrichment, and then we dis-
cuss the actual means to implement our data quality improvement
program.
16
DATA ENRICHMENT/ENHANCEMENT
Data enrichment (also known as data enhancement) is a process of

increasing the value of a pool of information by combining data from
multiple sources. The goal of enrichment is to provide a platform deriv-
ing more knowledge from collections of data. We can use enrichment as
a means for learning more about entities in a set of databases to create a
"value-added" information product, to investigate fraudulent behavior,
or even to increase the quality of a given data set.
A simple enrichment example is in targeted marketing. Let's
assume that a preferred customer is one who spends more than $100 a
month on a set of products and whose credit history indicates strong
predilection toward prompt payment of bills. We can use data enrich-
ment to combine customer sales data with customer credit data to build
a data mart that can be used to make special offers to these preferred
customers.
Another example involves investigating fraud. Suppose that when
more than a certain amount of money is paid to members of a profes-
sional organization, then there is a suspicion of fraud. In this case, we
would merge health insurance claim information with professional
billing information to search for this kind of behavior.
A third example involves enabling efficient financial transaction
interchange. In the financial world, different companies have different
representations for securities trades, but there are unique identifiers for
each stock, option, bond, and so forth. Through combining incoming
transaction messages with internal data tables, we can enrich financial
electronic data interchange messages to enable automated straight-
through processing.
399
In any case, data quality is critical to successful enrichment. In this

chapter, we discuss how data standardization, clustering techniques,
and the use of data quality rules can be used to express enrichment
directives. We will also look at some enhancement examples and how
they can be affected by poor data quality.
16.1 WHAT IS DATA ENHANCEMENT?
Let's imagine this brief scenario. You are a sales manager, and your
entire compensation is based on commissions on the sales that you and
your team make. Imagine that the director of sales handed each member
of your sales staff a list of sales leads that contained names, addresses,
and telephone numbers. Each sales representative might have some
chance of closing a sale with any one of those leads, but it would be
anybody's guess as to which lead is more likely to convert to being a
customer than any other. Without any additional information, each call
is a crapshoot.
Now imagine if the list contained each sales lead's annual salary,
the amount of money they've spent on similar products over the last five
years, a list of the lastfivepurchases the lead has made, and the propen-
sity of that lead's neighbors to purchase your product. Which of these
lists would you rather have.^
That second list contains enhanced data. The value of the original
list has been improved by adding the extra personal and demographic
information. By having the extra data added to the original set, the sales
team can increase its effectiveness by prioritizing the sales leads in the
order of propensity to buy.
The value of an organization's data can be greatly increased when
that information is enhanced. Data enhancement is a method to add
value to information by accumulating additional information about a
base set of entities and then merging all the sets of information to pro-
vide a focused view of the data.
16.2 EXAMPLES OF DATA ENHANCEMENT
In any situation where there is an opportunity for enhancement, the

goal in to produce a data set whose added value can help optimize a
business process. In the customer relationship world, the goal is to pro-
DATA ENRICHMENT/ENHANCEMENT 401
duce a set of customer profiles that can provide both a framework for
more efficient sales efficiency and a streamlined mechanism for cus-
tomer service. In the sales analysis world, this may imply enhancing
point-of-sale data to understand purchase patterns across the organiza-
tion's sales sites. In the health/pharmaceuticals industry, a goal could be
to understand the interactions between different drugs and to suggest
the best possible treatments for different diseases.
There are different ways to enhance data. Some of these enhance-
ments are derived enhancements, while others are based on incorporation
of different data sets. Here are some examples of data enhancement,
which are shown in Figure 16.1.
16.2.1 Temporal Information
A relatively simple way to improve data is to add a timestamp. Under-

standing aspects of the "ebb and flow" of transactions over a specific
time period adds significant value. Also, being able to track historical
data is useful in understanding trends.
16.2.2 Auditing Information
One frequent data enhancement is the addition of auditing data to a

record in a database. For example, in a customer support database,
each time a customer has a discussion with a customer support repre-
sentative, not only is the conversation noted but also the name of the
representative with whom the customer spoke, along with a timestamp.
16.2.3 Contextual Information
The context in which a piece of information is manipulated may yield

an opportunity for enhancement. Location, environment, path of access
are all examples of context that can augment data. For example, an e-
commerce company may monitor Web accesses to determine if the users
are visiting Web pages from a computer at work or at home. Contextual
enhancement also includes tagging data records in a way to be corre-
lated with other, external pieces of data during later analysis.
0 <u
>
n
1t "C
a
C3
u
TD
C
c
^
u
^
C/5
-Si o
*- -Si a 4>
Si
g ô '-%
^G
- a -rS
13
^ 'I-'
^ s (d
u .H o
0 ^
v^
>s
S
1> T3
ON
s o
^^ U^
C!
^ C c
^2 (A ^
-^
S2 ^0 Cfl ^0
CA
9
NJ
0
9
^ ^ HJ HJ
"2 "O
"H "O
"3 -o
'> >
(d
>
rt >
(D
Q Q Q
i'~\
O
ON
OS
"3
I I
O
3
I 3
c to 00
a> OS
2
11
3
I O
i
u
1
o
U
3
1 -Si
"a,
6
X
g
O
CO
I J3
I I I U
O I o (0
UJ
cc
D
G
16.2.4 Geographic Information
There are a number of geographic enhancements. Data records contain-

ing addresses are initially enhanced through address standardization.
Addresses are cleansed and then modified to fit a predefined postal stan-
dard, such as the United States Postal Standard. Once the addresses
have been standardized, other geographic information can be added.
This includes regional coding, neighborhood mapping, latitude/longi-
tude pairs, or other kinds of regional codes.
16.2.5 Demographic Information
For customer data, there are many ways to add demographic enhance-
ments. Demographic information includes customer age, marital status,
gender, income, ethnic coding, to name a few. For business entities,
demographics can include annual revenues, number of employees, size
of occupied space, and so on.
16.2.6 Psychographic Information
Psychographic information is used to break the target population down

into its component lifestyles defined by how each individual behaves.
This includes product and brand use, product and brand preferences,
organization memberships, leisure activities, vacation preferences, com-
muting transportation style, shopping time preferences, and so on.
16.2.7 Inferred Information
Having a base set of data whose quality we can trust gives us the oppor-
tunity to aggregate, drill, slice, and dice that data. When we can infer
knowledge based on that data, we can augment the data to reflect what
we have learned.
16.3 ENHANCEMENT THROUGH STANDARDIZATION
When we talk about standardization, we refer to making sure that a

data record or a message conform to a predefined expected format.
That format may be defined by an organization with some official
authority (such as the government), through some recognized authori-
tative board (such as a standards committee), through a negotiated
agreement (such as electronic data interchange (EDI) agreements), or
just through de facto convention (such as the use of hyphens to separate
the pieces of a U.S. telephone number as opposed to other national
formats, which might use commas or periods).
Converting data to a standardized format is an extremely powerful
enhancement. Because a standard is a distinct model to which all items
in a set must conform, there is usually a well-defined rule set describing
both how to determine if an item conforms to the standard and what
actions need to be taken in order to bring the offending item into con-
formance.
Aside from the data quality benefits that we discussed in Chapter
14, the value of data standardization lies in the notion that given the
right base of reference information and a well-defined rule set, addi-
tional data can be added to a record in a purely automated way (with
some exceptions). A very good example is ensuring that a U.S. street
address has been assigned the correct ZIP code and inserting the correct
ZIP code when it is missing.
This is not to say that automated standardization is simple. As we
discussed in Chapter 14, sometimes transforming into standard form
can be relatively complex. There is significant benefit when a standard
can be translated into a set of rules from which a standardization
process can be formed. Using our example from the previous para-
graph, we can standardize the ZIP code portion of a U.S. street address
through a process that is determined based on the specification of the
assignment of ZIP codes.
The entire country is divided into geographical regions designated
by a five-digit ZIP code. Each region includes a number of the streets
within a political geography, and the Postal Service maintains a data-
base of which ZIP code regions contain which streets in each city (along
with address ranges). So, given a street address within a named city and
state, the determination of the correctness of the ZIP code is the same as
the process for standardizing the address by adding a ZIP code.
1. Determine if the city is a valid city name within the state. This
corresponds to a query in a city-state mapping table.
2. If so, determine if the street name is a valid street name within
that city. Again, this corresponds to a query in a database map-
ping between streets and cities.
3. If the street is valid, check to see if the address (the street num-
ber) is in a range that is valid for that street. Do another mapping
lookup, and this mapping should also reflect the ZIP code map-
ping as well. A test will compare the found ZIP code, and an
assignment will just use that found ZIP code.
Note, by the way, how this process uses the data domain mappings
we discussed in Chapter 7! But what happens if one of these lookups
fails .^ The default would be to resolve the ZIP code to the closest level in
the geographical hierarchy. For example, if the street is valid but the
number is not, then assign the ZIP code for the street. If the street does
not exist within the city, assign a default ZIP code for that city. While
the result is not always correct, it may still be in standard form.
16.4 ENHANCEMENT THROUGH PROVENANCE
This enhancement is a generalization of the temporal and auditing

enhancements just described. A provenance of an item is its source. An
interesting enhancement is the operation of associating a provenance
with a data record, which would consist of a source code and a time
stamp (marking the time at which the data was updated).
16.4.1 SIMPLE PROVENANCE
A provenance can be as simple as a single string data field describing the

source or as complex as a separate table containing a time stamp and a
source code each time the record is updated, related through a foreign
key. Because there is a complete audit trail for all tagged data records, this
second approach allows for a more complete itinerary to be compiled.
One interesting benefit is that later analysis can show the value of
different sources of data, as well as different pathways or traces of mod-
ifications to data. For example, if it can be shown that one particular
original data source consistently provides information that is never
updated, it can be inferred that that provider's information is more reU-

able than that of a provider whose records are consistently made cur-
rent. Another example might allow for measures of volatility with
respect to information acquired through different tracks.
16.4.2 Audit Trails
The more advanced form of this enhancement supplies not only a

source for creation or an update of a record but also additional activity
fields. When we combine source, time, and activity information into a
record, we can trace back all occasions at which a piece of information
was touched, giving us the opportunity to truly understand how activi-
ties cause data to flow through a system.
16.5 ENHANCEMENT THROUGH CONTEXT
In today's Web-based environment, there are ways to not only track

when a user is touching a data set but also where the user is sitting when
the activity occurs. In these cases, we can enhance data with what we
call context information. This kind of information includes the virtual
location from which the activity takes place (visiting a particular Web
domain name), a physical location (from a home computer versus an
office computer), as well as other data that can be collected directly
from the user through direct interaction.
In this kind of environment, substantial marketing benefit can be
derived, since this context information can be fed into a statistical
framework for reporting on the behavior of users based on their loca-
tions or times of activity. For an e-commerce concern, this information
can be quite revealing and can be used for strategic benefit. As an exam-
ple, a Web-based business can determine that many of its customers
browse through catalog entries during the daytime while at work but
peruse the content for a subselection of entries at home in the evening.
The business can retool the Web site to provide different kinds of pre-
sentations during work hours or leisure hours, which will encourage
users to purchase products.
16.6 ENHANCEMENT THROUGH DATA MERGING
An effective method for data enhancement is combining the informa-

tion that is stored in more than one database into an enhanced data-
base. These kinds of operations are performed all the time, mostly
represented as database joins. Frequently, these joins exist for some
operational purpose, and the enrichment of the data is purely done for
an operational reason. For example, straight-through processing sys-
tems may require record data enrichment as data are passed from a
source organization to the target so that the next step in the process can
be determined and executed (see Figure 16.2).
How is data merging done? It is basically a matching operation:
select a number of attributes that most likely characterize the entity,
(we'll call them the characterizing attributes) and then search for all
records that have the same attributes. Those records form a candidate
set for enhancement all of those records may contribute data to a
new enhanced piece of information. This process works well when
merging new records into an already existing data set but may be less
effective when applied directly to different data sets. But hold on to that
thought we will explore this issue more in Section 16.7.
Strategic value can be attained through combining data sets. In this
case, the talent in creating a data enhancement environment is less in
the SQL skill and more in the qualitative analysis of what information
needs to be merged and what other techniques can build a more effec-
tive data merge. Several scenarios exist in which data merging may be
necessary.
16.6.1 Corporate Mergers
When two companies merge, eventually they will have to merge their
customer databases, employee databases, and base reference data. Con-
solidating customer records prevents potentially embarrassing market-
ing snafus (such as having more than one sales representative contacting
the same customer on the same day). More interesting, additional busi-
ness intelligence can be collected on those customers that existed in both
original data sets.
f "1
OS 1 1
OS 1 1
OS 1 1 2 ^
u 1
1 **; 00 2 1
t1 1 1 1 > wi a 1
1 a 4> 4> 1
a> 1 1 JD "O 1
^ 1 1 u 1 u 11
t S 2:!
^ 1^ 1 1 ON 1 1 TD 1 1
OS 1 1
o
Q a 1 ON 1 1
I1^
^^ 1 1
u 1 00 1 1
Q 1 5^ 1
T^ 1 1
to 1
ON 1 OS 1
o1 11 o) 11 11
o 1 1
1
0 1 1
OS
o 11
1 1 lit
1
"3
n
00
C3N
I^
I 1
00 1
ON 1 1
*o 1 1
I^ 11
NO
1 f2
d 1
3
V
f^
00
1
t^ 1 1
1
A 1 1
f^
S> 1
1
1
1
1 C3 (A >
O H-u
CJ
^^ 1
1 1
< 6 ^ 1 1 1 V jd o
ON ON 1 <<
- o
ON 1 1
CO 1 1
OS ON 1
^1 ON 1 ^1 S
OS 1 o 1 ON 1
-^ i ON ON 1 ON 1
<N 1
(3 ON 1

1 1 oo
'^ 1
00 1
^-^ 1
1
O

J^
l^
00
1
1
1
1
1
> 00
rH
1
1
ON 1
A 1
<s 1
(A
2i SO 1
^
a <N K. 1 <N 1
Q
z ON 1
1 ^ ^ tt
1>
i 2 2 1 i 2 1
O ^1
bS
r
1 u<
1
b
i ^ M >
8 S-S W)
a
r r tn
^
ill
^1
M
OS
Mil ^ 1 ^
1 "
8 1 c: 1
1 ^ 1 o 1 Q
1^ 1 1 ^
1 o
B 1 o a 1 ^
1 V
1
1 73 i
i 1 ^ 1 Si 1 ^ 1 MH 1
1 (A 1
1 *"* 1 J5 a - g 1 c:
1\2
1 Q^ 1 V 1 1 ^ 1
^ 1 ^ 1 w
1 ""* 1 1 *^ S !> 8 1 ^
^
1 '^ 1 ^ 11 "^
^ 1
1
r 1 ^1
V ^ &
1 V i$ (d
HI 1 T-
UJ
1 1 PQ t " S 1 CO
J v__ ,^ il
16.6.2 Affinity Programs
An affinity program attempts to sell a product customized to a particu-

lar group of customers, based on a preexisting characteristic shared by
all of them. One example is a university alumnus credit card with a pic-
ture of a university's main library on the card and a program through
which some small percentage of a customer's annual charges are
donated to that university. Once the affinity product has been created,
an affinity program can be effected by merging a customer lead list,
such as a bank's current account holders with average balances greater
than $5,000, with an affiliation database, such as the alumni directories
from the Ivy League schools.
Other affinity programs include health and life insurance tie-ins,
product discounts, and service discounts. In each of these cases, there is
some aspect of data merging.
16.6.3 Cooperative Marketing
Many companies take advantage of synergies when partnering with

other organizations in cooperative marketing campaigns. Frequently,
the originators of these campaigns attempt to influence the decision to
purchase one product, based on a customer's previous experience with
another product. A good example of a cooperative marketing campaign
is when a long-distance telephone provider teams with an airline to give
the customer a bonus of frequent flyer miles if that customer opts for
the offered long-distance service.
Typically, organizations that try to build cooperative marketing
programs will do so in a "brute force" manner by combining data sets
and blasting out offer letters. While this may a workable means of
acquiring new business, a more refined method exists that can create a
more effective marketing campaign. This method combines data
enhancement techniques with database merging to provide a narrowed
focus for marketing.
16.6.4 Data Cleansing
Because customers move, get married or divorced, change their name,

and so forth, very often, large customer databases tend toward entropy
in turns of accuracy. Not only that, when people are involved in the
data entry process, mistakes can be made. This leads to the existence of
duplicate records for many individuals or corporate entities. As seen in
Figure 16.3, which contains real data extracted from the U.S. Securities
and Exchange Commission's EDGAR database, many of these dupli-
cates are not always simple to recognize. Linking the duplicate records
together is a way to cleanse the database and eliminate these duplicates.
16.6.5 Householding
Another process, called householding, attempts to consolidate a set of

individuals to a single grouping unit based on the database record attri-
bution. A household consists of all people living as an entity within the
same residence. The simplest example is that of consolidating husband
and wife records to a single household.
Households can be differentiated by demographics as well as geo-
graphies, and in one household there may be different subsidiary roles,
such as primary earner and primary decision maker (these two may be
different individuals). More complex examples include the identifica-
tion of dependents, categorizing dependents by class, as well as separat-
ing out boarders or transient residents. Householding can be used to
improve demographic analysis, marketing optimization, as well as tar-
geting particular defined roles in the household.
16.6.6 Medical Diagnosis and Treatment
In the health industry, data merging is performed both for diagnosis and
for determination of treatment. A collection of medical professionals
may pool their individual patient information (most likely having been
anonymized first) as a cooperative means for building up a knowledge
base particular to a set of illnesses. The information, which consists of a
patient's history, diagnosis, and treatment, must be enhanced to fit into
the collaborative data model.
For diagnostic purposes, a new patient's history and profile are
matched against other patients' histories in the hope of finding a close
match that can help determine the cause of a health problem. Once a num-
ber of matches have been found, the matched patients' treatment proto-
cols are examined to suggest a way to treat the new patient's problem.
i!
o
I ( ^ "^^ o vo t ^ oo ro '
J <^ r ^ ^ ^ u
O O O ( S (N <S 0 ' i O O O ,-(
I <N r^^ N cv| ( N ( N <N 1 ^ <> <^ ;/5
i 0^ 04 QdS P^ P^ P^ P^ PIH Pd P^ > u j p^ p^ pe: u u u SQ
I U j UL) U j UJ UJ U j PL] u j u j pg I t o t o t o [jj
I C/1 t/5 t o t o t o t / i C/5 t o t o t o I )J J J ^ < CO CQ r^
! D D D C CO D D ! ;
I U 11^ U U U3 U U ^ D ^ ^c^ OH
t/5 r n r o <J
I D D D D D D ID :
iXXXg^
* til PJ U4 H
b^ûU U U PQ i55|
CQ CQ CQ CJ
SPS
I (1) tL) tL] U j PL) U j PL] t iwSi * ^ r"< ?*^ tiJ
GGb^
1< *? to
:x X X X x x x ^ U U U oti
<^<
Ip p p p pppf >aa(yx <<<<
PL] ' ^ ^ ;
I ^ ^ ^ ^ ^ ^ ^ . '^z z z ^ : i pi^ >
|< < <3 < < < < ^ )B s a ^ <i :< :<^ ^
> > > ^
Oi Oj CL, O
ID D D DD D P :
oO 00 00 0^ S P2 U
OOO2;
J3
PfS P ^ pe4
ÔQOOOQQOOSQOOOOOOOOO o8 o^ o8 I (d PL) uj Tl
PEH
D D D D zqqi
OOOP^
tto
o tto
o to
0/5 tto
o
^ pq PL] I
Pl^ P^ P^ I
6
;uoo IO U O U < <: < <
[ i ] UJ t i j tL) to to to <
UN t i l PL4 PUi s
^ u ^ u
< <: <: <
sss; o
CO
auoo
I & & s g Q Q Q Q
Z Z Z Z III' S!
XXX
ti] ti] ti]
CO
CD
to to to to xxx; t-
Z Z Z Z < << UJ
i Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q <<<:<: p^ pe: p^ I cc
'-ipLiPLiujtgujtdpLjtDtgtLju^pLiujuj
iO^P^Pl^P^pt^PfiP^P^pdP^PdPl^P^P^
uuu u
p^ pfj pe^ 0^
PLJ tj PL] I
PU 0H PU I
D
UJ t i ] PL] PLJ
PDDDDDDDDDDDDDDDDDDDD
tOQ/5tOtOtO;/5tOtOQ/5tOtOtOtOtOtOtOtOtOtOtOtO ssss 1
!<<;<<
tJU t i ] PL] I
16.6.7 Fraud Detection
In many areas where there is a potential for fraud, data merging is used
as a way to both identify fraudulent behavior patterns and use those
patterns to look for fraud. Opportunities exist for fraud in all kinds of
businesses, such as transaction based (telephone and mobile-phone ser-
vice), claim based (all kinds of insurance), or monetary transaction-
based (where there are opportunities for embezzlement, for example).
In fact, there are many areas of crime that call out for data matching
and merging: money laundering, illegal asset transfer, drug enforce-
ment, and "deadbeat dads," to name a few.
16.7 DATA MATCHING, MERGING, AND RECORD LINKAGE
Simple data merging between records is a process of choosing a set of

attributes for characterization, using a base reference record and search-
ing through a set of records for those records whose attributes match the
reference record. Thos records are candidates for linkage, and the other
attributes are examined to determine the closeness of a match.
When any two records can be associated based on a set of chosen
attributes, we say that those two records are linked. Usually, record
linkage is performed only when the chosen attributes match exactly. But
simple record linkage is limited for the following reasons.
Information sources are in different formats. It is easy to say that
two records match when they come from the same table with the same
attributes (with the same column names, etc.). But when the records
come from different data sets, it may not be clear which are the right
attributes to match. For example, in one data set, there may be a col-
umn called ACCT_NUM, and in a second data set the column repre-
senting the same information is called ACCOUNT^NUMBER.
Information is missing. As we have discussed many times in this
book, there is significant importance in properly identifying the kind of
information that is not there in other words, the different kinds of
null values. When a characterization attribute's value is missing, all
other attributes may match and still prevent two records from linking.
Record Linkage is imprecise. While two records might actually
represent the same entity, a slight difference in one (or more!) attribute's
value will prevent those records from being linked.
Information sources are out of synchronization. Customer infor-
mation in a database can stay relatively static, even if the customer her-
self is not. People move, get married, divorced. The data that may be
sitting in one database may be completely out of synch with informa-
tion in another database, making positive matching difficult. For exam-
ple, recently I moved from Washington, D.C. to New York City. It took
a year before any of the online telephone directories had my listing
changed from the D.C. address and telephone to the New York one.
(Incidentally, by the time those listings had changed, I had already
moved to a new address!)
Information is lost. The actual database joins may be constructed
in a way that important information that is originally contained in one
of the data sets is lost during the merge. For example, when my baby
daughter receives a letter asking her to switch her long distance service
in return for an extra 5,000 frequent flyer miles, the information that
my wife and I just purchased a child's seat for her is apparently lost.
When the limitations of standard linkage are combined, it eventually
causes inefficiencies and increased costs. Here are some real examples:
A large frequent traveler program had fielded many complaints
because patrons who had visited member hotels were not credited
with the points corresponding to their purchases. It turns out that
the company providing the frequent traveler program had recently
merged with a few other hotel chains, and the number of different
data formats had ballooned to more than 70, all out of synchro-
nization with each other.
A bid/ask product pricing system inherited its prices from multiple
market data providers, the information of which was collected
and filtered through a mainframe system. When the mainframe
did not forward new daily prices for a particular product, the last
received price was used instead. There was significant embarrass-
ment when customers determined that the provided prices were
out of synchronization with other markets!
Very frequently with direct mail campaigns, multiple mailings are
sent to the same household or to the same person, and a large
number of items are sent to the wrong address.
Current customers are pitched items that they have already
purchased.
Some silly data merging mishaps can be embarrassing to the orga-
nization and cause customer distrust. For example, if a service company
can't keep track of who its customers are, how can they be trusted to
supply the right level of service as well as provide a correct bill every
month .^
16.8 LARGE-SCALE DATA AGGREGATION AND LINKAGE
Databases and data warehouses are not the only place where data merg-
ing is an important operation. In our ever-growing World Wide Web, any-
one can publish anything, and that information can be aggregated as well.
Closer to the fact, many Web sites act as front ends to different underlying
databases, and clicking on a Web page is effectively the invocation of a
query into a database, and the presentation of information posted back to
the client Web browser is a way of displaying the result of the query.
If we configure an application to act as a client front-end replacing
the standard browser but still making use of the HTTP protocols to
request and accept replies, we can create a relatively powerful data
aggregation agent to query multiple Web sites, collect data, and provide
it back to the user. Effectively, the World Wide Web can be seen as the
world's largest data warehouse. Unfortunately, the problems that
plague standard record linkage as described in Section 16.7 are magni-
fied by at least an order of magnitude. The reason for this lies in the rel-
ative free-form style of presenting information via the World Wide Web.
16.8.1 Semistructured Data
For a large part, Web data are "semistructured data." While database
records and electronic data interchange messages are highly structured,
information presented on Web pages conforms to the barest of stan-
dards, some of which are bent based on the selection of Web browser
targeted. On the other hand, there are some Internet "motifs" that
appear regularly. For example, business home pages most often contain
links to other sections of a Web site, with a contact page, an informa-
tion page, a products page, a services page, a privacy policy page, and a
"terms of service" page. Information on each of these pages also tends
to follow a certain style.
For example, a corporate data page will often have a list of the top
managers in a company, followed by a short biography. We can even
drive down to finer detail in our expectations in the corporate
biographies, we can expect to see some reference to college and gradu-
ate school degrees, an account of prior work experiences (with dura-
tions), and professional affiliations. Yet while the format of these
biographies is not standardized at all, in general we learn the same kind
of stuff about each manager in each company.
DATA ENRICHMENT/ENHANCEMENT 41 5
Personal Web pages will also conform to some semistructure.

These pages will include names, contact information, perhaps some
work history, interests, and so on. But there is not standard format for a
personal Web page. Perhaps we can take a cue from some of the more
popular Web page editing tools and, understanding their templates and
motifs, use that as a guide for identifying particular pieces of data that
are relevant or interesting.
16.8.2 Web Sites That Are Really Databases
To borrow an idea from Phillip Greenspun's book Philip and Alex's

Guide to Web Publishing, some Web sites are really front ends to data-
bases, and the pages that are published reflect queries into the database.
For example, searching an online book store for a particular author
"Smith" is really asking (in pseudo-SQL), "Select all records where
author's name is 'Smith.'"
Because different providers configure their replies to requested data
in different formats, we have the same set of issues as we have seen before.
What differentiates this situation is that the amount of information that
can be collected from each site and the multitude of sites that can be vis-
ited. The result is a very large-scale data aggregation framework.
16.8.3 Web Sites That Are Broadcasters
In contrast to database-oriented Web sites, some content providers are

broadcasters of information that can be valuable for enrichment. There
are two clear examples of this kind of Web site news providers and
financial information providers. In both cases, data that are broadcast
may have different value to different visitors, and any aggregation
application must ultimately be able to allow the end user to identify
what data items are important and which are less important.
16.8.4 The Breakdown in the Data Warehouse Analogy
The data warehouse analogy partially breaks down when it comes to

the actual aggregation and presentation of information at the presenta-
tion site. A critical value that can be derived from the aggregation of
information is its coagulation, enrichment, and presentation in terms of

critical business intelligence that is meaningful to the user.
One reason for this is simple similar types of information avail-
able across the Web are represented in different ways. Whether using
different storage formats, different names, different visual representa-
tions, the data that are presented at different Web sites and that can be
collected and delivered to the client location can be made much more
valuable when that data is massaged into business intelligence guided
through client-defined requirements.
Another reason for the breakdown in the analogy has to do with
governance (or more clearly, the lack thereof). Anyone can publish any-
thing on the Internet, yet there is no governing body that oversees the
quality of any information published. Therefore, it is up to the user to
characterize his or her data quality requirements to which the informa-
tion should conform before the information is presented.
16.8.5 Aggregation and Linkage
The key to making the best use of aggregated Web data is that the pre-
sentation and packaging of the information is effectively an exercise in
data enhancement. Before the information is presented to the client, any
data quality rules, business rules, validation filters, or trigger rules
should be applied. This can only be done when there is a method for
linking data items from different sources coupled with the definition
and usage of business and data quality rules, as we have explored in
other sections of this book.
16.9 IMPROVING LINKAGE WITH APPROXIMATE MATCHING
One way to counter the limitations of standard record linkage is the use
of a technique called approximate, or approximate matching. Standard
linkage requires that the sets of values in the characterizing attributes
all must match exactly. But in databases, just as in real life, sometimes
things are not always as they seem. Approximate matching relaxes the
exact matching requirement, giving us the chance to find those elusive
near (and sometimes not-so-near) matches.
16.9.1 Approximate Matching
What constitutes an approximate match? There are two aspects to

approximate matching. The first is the distance, or similarity measure,
and the second is the threshold.
In Chapter 14, we learned that "approximate" can be measured in
terms of what is called a distance measure. In review, this term is inher-
ited from mathematics from geometry problems that identify how close
two points are in a given multidimensional system by measuring the dis-
tance between them. There are many kinds of distance measures, rang-
ing from integer comparisons to traditional geometric measures (pull
out those high school math texts!) to measures based on the number of
alphabetic symbols that are shared between two character strings. In
most cases, then, a distance measure provides a measure of closeness
between any two data items of a particular type or domain.
To further level thefield,we like to scale all distance measurements
so that they are between 0 and 1. In essence, we are using a percentage to
indicate a measure of closeness. This leads to the second element of
approximate matching: the threshold. Because the matches are not exact,
depending on the degree of the problem, we must have some point at
which we determine two values to be a match or not. This point is defined
by a threshold value the value above which a similarity measure is
deemed a match.
Even for the same set of data values, thresholds may change based
on the application. For example, if we are trying to prevent terrorists
from entering the country, our criteria for an approximate match might
be looser than if we are trying to eliminate duplicates from a direct mar-
keting mailing.
16.9.2 An Example of Approximate Matching
We can best illustrate approximate matching by looking at an example.

Once I received invitations to an upcoming data warehouse seminar.
One was addressed this way.
David Loshin
Prod Manager
Jones Smith Co
750 X Ave
New York, NY 10019
And the second one was addressed this way,

H D Loshin
Product Manager
Jones Smith & Co
750XAveFil2
New York, NY 10019
By examining these two records, one can easily see that they would
not have been linked using exact matching. In a approximate matching
scheme, a similarity score would be assigned to eachfieldin these name
and address records. Let's say these are the scores.
Record 1 Record 2 Score
David Loshin H D Loshin 70% (D matches David,

Loshin matches Loshin)
Prod Manager Product Manager 90%
Jones Smith Co Jones Smith & Co 95%
750 X Ave 750XAveF112 80% (due to the floor
designation)
New York, NY 10019 New York, NY 10019 100%
Since all the scores are relatively high, even outside of the context
of the material being sent, our intuition would say that these are likely
to be the same person. Now adding in one more piece of information,
which is the psychographic item of "interest in data warehousing," this
might tip the balance to an automated process to automatically link the
two records. One more interesting note: I received the same invitation
at my home address!
16.9.3 Approximate Matching Rules
For any attribute where we want to use approximate matching, we

must provide both a similarity measure and a threshold value. When
performing approximate matching over a set of characterizing attrib-
utes, there must be a similarity measure and threshold for each pair, as
well as an overall aggregation function (one that collects the individual
similarity measures) and threshold value.
DATA ENRICHMENT/ENHANCEMENT 41 9
We discussed in Chapter 8 that there were other data quality rules.

Here we formally add approximate matching rules to that rule set. A
similarity measure can be defined in two ways either within the sys-
tem (using code definitions for calculating similarity) or as a reference
to an external processing phase. We can assume that there are some
built-in similarity measures, such as linear (how close two numbers
are), the geometric similarity (for multidimensional systems), and edit
distance for character strings. A approximate matching rule for a spe-
cific attribute is then defined using a domain (or base data type), a simi-
larity measure, and a threshold. Here is an example that defines
similarity in terms of the relative closeness of two values, indicating a
match when the score is greater than 75 percent.
ScoreMatch: APPROXIMATE MATCH{1 - (ABS(Valuel -
Value2)/MAX(Valuel,Value2))}:threshold=0.75
In this example, we specify an approximate matching rule we call
ScoreMatch, with an embedded similarity score computed by scaling
the difference between the two values to a percentage and subtracting
that from 1. We also specify a threshold of .75, which means that for
any compared values, if the result of applying the similarity measure
yields a result greater than 0.75, the two values are deemed to be a
match.
A set of characterizing attributes can be combined using an aggre-
gation function over all those attributes under an approximate aggrega-
tion rule.
APPROXIMATE AGGREGATE CustomerAgg -.{((ScoreMatch
(Customers.Loyalty, Leads.Loyalty)*3) + (NameMatch(Customers.Last-
Name, Leads.LastName)) + AddressMatch(concat(Customers.Address,
Customers.City,Customers.State),concat(Leads.Address,Leads.City,
Leads.State))/5}:0.81
In this aggregation, we are scaling the score match to have a weight
that is three times greater than the other two match rules that complete
the set. Included in an aggregation rule must be an expression designat-
ing the approximate rules to be included, the weighting of each of the
scores, a function describing the aggregation, and the threshold above
which two records match. Also, within each approximate match rule
we indicate the data values that are to be matched.
16.9.4 Where Approximate Matching Changes the Outcome
Approximate matching gives businesses the opportunity to be more effec-

tive in its data merging operation because it weakens the requirements for
determining a match of two entities by allowing some degree of heuristic
capability. Approximate matching allows for a wider net to be cast when
merging data from different data sets, accounting for differences in the
way that the information was collected, entered, and so forth.
Being able to identify different representations of the same entity
(customer, address, business, etc.) in the same large data set is beneficial
for many reasons:
You can gain greater insight into your customers when more
records can be linked together.
For direct marketing campaigns, costs can be reduced by eliminat-
ing double and triple mailings.
The same fraudulent behaviors may manifest themselves in differ-
ent patterns, which can be detected only through approximate
matching.
Data sets can be updated with more recent information.
Important customer information can be derived from the results
of approximate linkage (for example, lifestyle changes such as
marriages, new home purchases, the birth of children).
A single data set of record can be used to validate other data sets.
16.10 ENHANCEMENT THROUGH INFERENCE
Having discussed any number of mechanical enhancements such as the

ones described in this chapter, it is time to discuss one set of enhance-
ments that is less mechanical and more intuitive enhancement
through inference. An inference is an application of a heuristic rule that
essentially creates a piece of information where it didn't exist before.
Even though inferencing represents the application of intuition, it is
done so in a way that can be automated.
An example of an inference for credit card databases might be if
there are two names associated with the account number, if more than
70 percent of the dollar amount of purchases is made by one of the two
account holders, and if that account holder is the primary decision
maker for that account. Using this rule, we can populate an attribute
for the primary decision maker with either a definitive value or an

"indeterminate" null value.
Inference rules usually reflect some understood business analysis
that can be boiled down to a set of business rules. We have explored a
way to define these kinds of inferences we called them derivation
rules in Chapter 8. These are rules that specify how certain attribute
values are determined by operations applied to other attributes.
Combining inference rules with data merging yields a powerful
process for enhancement. For example, in a householding operation,
determining those parties that share an address is one thing; isolating a
role for each party within the household is more difficult. But imagine a
rule called the "power of purchase" rule: When any purchase incorpo-
rating products that are bought directly for named individuals is paid
for by a single person, we can infer that the nonbuying individuals are
dependent on the purchaser. Travel tickets, school tuition, personalized
gifts, health insurance, trust accounts, address labels, personalized
mementos (like bronzed baby shoes), and automobile insurance are all
examples of this kind of purchase. This rule suggests two roles: the
dependent and the spending authority.
We already have a structure for the definition of enhancement rules
based on the rule definitions from Chapter 8. An enhancement can be
characterized using new attributes, and these attributes can have
derived values.
16.11 DATA QUALITY RULES FOR ENHANCEMENT
Another step in intelligent enhancement is adding data quality rules. As

we have seen in Chapter 8, we can define rules that both validate asser-
tions as well as modify or transform data values. We can use these rules
for enriching data in one of two ways. If the data record (or message)
completely conforms to the data quality rule set, then we can tag the
record with a new field indicating the location in the information chain
where the record was vaUdated and a "stamp of approval." If the
record does not conform to the rule set, the record can be enriched with
a new field including the rule (or rules) violated and the location in the
information chain where the validation failed.
When the information is evaluated at a later time, it is then possible
to identify those locations in the information chain that are good "data
quality performers" and those weak links in the chain for data quality.
One interesting aspect of enhancement is the possibiUty of inferring

information based on patterns of poor data quality. Patterns of certain
data quaUty problems may appear chronically, which may hint toward
particular root causes. In addition, the fact that these patterns may be
taggable to particular sources enables a certain degree of inference
possibiUties.
For example, we can log the types of data quality errors that appear
and then measure their frequencies. When there are frequently occuring
errors, we can evaluate the sources of those errors to see if there are any
interesting patterns. If so, we can enhance incoming data by identifying
its potential for certain errors, thereby characterizing its quality level and
its improvement history.
16.12 BUSINESS RULES FOR ENHANCEMENT
The next step in intelligent enhancement is the addition of business

rules. In our previous data merging example, we snuck in a business
rule to help our matching process. That rule was the addition of the
property "Interested in Data Warehousing" to our data records for the
purposes of linkage. While it is unlikely that two people with the last
name of Loshin work with the same company in the same building, it is
entirely possible that two people with the last name Smith and the first
name John may work in the same 50 floor New York City skyscraper.
That is why adding additional business rules to the merging process can
improve its usability. Another benefit of adding business rules is that
you can impose more advanced logic on where and when not to link
records.
16.12.1 What Kinds of Rules Are Used?
We can use rules when the result of the merging/linkage operation

depends on content embedded in the merged record. For example, if the
long-distance company that sent the offer to my baby daughter had a
rule that states, "Do not send this offer if the frequent flyer has had a
child seat purchased within the past six months," they would have
known that a child is most likely not going to have spending authority.
A baby is the kind of direct marketing "mismatch" that one would
want to eliminate from a targeted campaign. By eliminating these mis-
matches, a company can increase its return on investment for marketing

strategies.
Business rules will typically take the form of an "if-then" state-
ment, with a condition that must first be fulfilled before the guarded
operation can take place. Once two or more records have been linked in
the merging process, the content of those records is examined to see if
the condition is true. So, to be more formal with our long-distance/fre-
quent flyer rule, we might want to add these rules:
1. For each frequent flyer record, enhance the record by determining
whether theflyeris "head-of-household," "second-in-command,"
or "non-decision-maker." This can be arbitrarily done by assign-
ing "head-of-household" to the flyer whose credit card was used
to purchase tickets.
2. If the flyer is a non-decision-maker, do not send a long-distance
offer to that flyer.
Another option is to modify the campaign based on rules. When
more than one flyer resides at the same household, the offer can be
changed to distribute the bonus miles among allflyerswithin the house-
hold if the decision maker opts for the new long-distance service.
16.12.2 Where Rules Change the Outcome
In all these occasions for data merging, business rules can improve the
outcome. When merging customer databases, derived analytical data
enhancements such as "decision maker" can be inferred. Cooperative
marketing programs can be made more efficient, thereby increasing
response rate. Affinity programs can be improved if rules characterizing
the target's tendency to respond can be inferred. And, of course, data
cleansing operations such as de-duplification and householding are
improved when rules are added.
16.13 SUMMARY
In this chapter, we looked at data enrichment as well as some examples of

kinds of enrichment that can be applied to data messages and records. This
includes temporal information, auditing information, contextual informa-
tion, geographical information, demographic information, psychographic
information, as well as different kinds of inferred information. We also

revisited the notions of standardization as a means for enrichment.
We devoted a large part of the chapter to data merging, since it is
through the merging process that we can derive more knowledge from
our data sets. Sometimes regular joins are insufficient, and we must rely
on more sophisticated means of linking together records. This includes
approximate matching, for which we "enriched" our own set of data
quality and business rules defined in Chapter 8. Finally, we looked at
ways that business rules can be used for enhancement.
Combining information from multiple databases is not a new idea,
but the advances in technology and system speed open up new vistas for
applying intelligent techniques. Using a focused methodology for
adding value through the data enhancement and enrichment process,
companies can greatly improve the effectiveness of their customer rela-
tionship, marketing, customer service and sales programs.
17
DATA QUALITY AND BUSINESS RULES
IN PRACTICE
In Chapters 7 and 8, we discussed the definition of data domains, map-

pings between those domains, and data quality assertions and directives
guiding the specification of data quality requirements. In Chapter 13,
we looked at the discovery of data quality rules out of legacy data, and
in Chapters 14 and 16, we looked at some of the more complex issues
revolving around data quality. In this chapter, we look at some of the
more pragmatic issues associated with our data quality rules: the actual
implementation of those rules for specific data quality purposes.
We begin this chapter by reviewing our data quality rules in a con-
text that demonstrates ways to actually use these rules. Specifically, we
have talked about the way that these rules partition the data into two
sets: conforming records and nonconforming records. There is already
a means for specifying set partitions in databases the query language
SQL, which is used not only for performing operations on databases
but to turn rules into executable queries.
The rest of this chapter is designed to show how the implementa-
tion of a data quality rules process can add significant value to different
operating environments. We will examine the following areas.
Data quality and the transaction factory
Data quality and the data warehouse
Data quality and electronic data interchange (EDI)
Data quality rules used to drive content-dependent user interfaces
Hopefully, this chapter will bring together a lot of the technical
details and tie up loose ends while bridging the mechanics of assess-
ment, user requirements analysis, and rules specification into a viable
operational context. Figure 17.1 shows the complete rule system.
425
Data quality rules

edited via Web
interface
Data is provided into Conforms to

validation engine Data Target
DQ expectations
Corrected data
Does not conform to DQ expectations
FIGURE 1 7 . 1 The complete rule system
17.1 TURNING RULES INTO IMPLEMENTATION
In relational databases, a query language called the Structured Query

Language (SQL) is used to extract data from the database based on a set
of criteria. Those criteria are represented as constraints in the query, and
the SQL statement represents the set of records that conform to those
constraints. For most of the rules that we have discussed, we can describe
the set of data records that do not conform to the rule using SQL.
In this section, we begin with a brief review of the SQL s e l e c t state-
ment; if you are already familiar with SQL, you can skip that section. The
DATA QUALITY AND BUSINESS RULES IN PRACTICE 427
rest of the section reviews the data quality rules, then shows how the set
representing the nonconforming rule can be specified in SQL.
17.1.1 Review of SQL s e l e c t Statement
Here is a quick refresher on the syntax of the SQL s e l e c t statement.

select [all | distinct] <select_list>
from [<table_name> | <view_name> ]
[,[<table_name> | <view_name> ] . . . ]
[where <search_condition>]
[group by <column_name> [, <coluinn_name>]. . . ]
[having <search_conditions>]
[order by {<column_name> | <select_list_nuinber>}
[asc I desc]
[,{<column_name> | <select_list_number>}
[asc I d e s c ] ] . . . ]
When a query is executed, the list of data items returned is called
the result set. If there are no records that match the query, the result set
is of size 0. For our purposes, we will focus on the tables and columns
selected and the where conditions. The group by, having, and order
by components are not used in this chapter.
For the < s e l e c t _ l i s t > , you can either use * to ask for all the
columns, a list of column names, or operations applied to column
names (such as 3* salary). The <table_name> list can consist of multi-
ple table names (or view names), and in many SQL dialects, an alias can
be assigned to the table name and the selected columns can refer back to
that alias. For example, this query
s e l e c t e.empname, e . s a l a r y from employees e;
assigns the alias name e to the table named employees, and the selected
columns refer to the aUased employees table.
When using the qualifier d i s t i n c t , this limits the query to elimi-
nate duplicate entries from the result set. The where clause is used to
specify search conditions. SQL provides these operations to be used for
constructing where conditions.
Comparison operators, such as <, <=, >, >=, =, <>. Example:
Where e. s a l a r y >= 20000
Boolean combiners or negation (AND, OR, NOT)
Where e. s a l a r y >= 20000 AND e. s a l a r y < 50000
Ranges (BETWEEN, NOT BETWEEN)

Where e. salary between 20000 AND 50000
Lists (IN, NOT IN)
Where e. t i t l e in ( ' STAFF MEMBER' , ' MANAGER' , ' PRESIDENT' )
Null values (IS NULL, NOT NULL)
Where e. salary i s null
Character matches (LIKE, NOT LIKE)
Where e.telephone not l i k e (212)%
We can use the result of a s e l e c t statement to generate a list that
can be used as a subquery for the in component of a where clause.
Select e.empname, e.salary from employees e
Where e . t i t l e in
( s e l e c t t i t l e from managerpositions);
In this example, there is a separate view that contains the records
describing those positions held by managers. We select the titles associated
with manager positions, and then we select out of the employees table
those employee names and salaries for employees who are managers.
17.1.2 Naming of Rules
In each of the following sections, we focus on the context of specific

rule types as discussed in Chapters 7 and 8. When we defined a rule set,
we also will want to assign a unique name to each defined rule.
In general, our rule specification will incorporate an assigned rule
name, followed by a colon and then the specification of the rule.
<rule name>: <rule text>
This way, we can integrate multiple rules into a single rule set that
is to be appUed to a data set. In the following sections, for clarity, we
ignore the rule names.
17.1.3 Null Value Representations
In Chapter 8, we discussed the fact that there are different kinds of null
values, including this list:
1. No value There is no value for this field a true null.
2. Unavailable There is a value for this field, but for some reason
it has been omitted. Using the unavailable characterization
DATA QUAUTY AND BUSINESS RULES IN PRACTICE 429
implies that at some point the value will be available and the field
should be completed.
applicable value.
value is not known.
We allow the user to define named aliased representations for the
different null types. This is the syntax for this definition.
Define <nullname> for <nulltype> as <string
representation>
In this definition, <nullname> is the alias name, <nulltype> is one
of the varieties of nulls described here, and <string representation>
is the character string used for the null. For example, this definition
specifies the null representation for an unavailable telephone number.
Define NOPHONE f o r UNAVAILABLE as "Phone Number Not
Provided"
To implement the management of null value representations, we

will need some metadata tables. The first is the null value type table,
where we consolidate the different kinds of null value types exists in the
system, what their names are, and their description.
Create table nulltypes (
Name varchar(lOO),
Nullid integer primary key,
Description varcharClOOO)
);
Next, we will need a table to manage the defined aliases. In this
table definition, the nullrepid is a uniquely assigned id, nuUtype is a
foreign key to the nulltypes t a b l e , and nullrep is the actual repre-
sentation of the null value.
Create table nullreps C
Name varchar(lOO),
Nullrepid integer primary key,
Nulltype integer,
Nullrep varchar(lOO)
);
430 ENTERPRISE KNOWUEDGE MANAGEMENT
17.1.4 Null Value Rules
There are two kinds of null value rules: those that allow nulls and those
that disallow nulls. Let's look at those that allow nulls first.
If the rule indicated that system null values are allowed, with no
qualification, there is really no validation, since the assertion states that
the absence of a value is allowed. Alternatively, if the rule specified that
only a particular kind of defined null value was allowed, then the vali-
dation is a little more complicated.
Our earlier definition of the null value rule allowed the restriction
of use to defined null values, such as the following.
Attribute employees.phone_number allowed nulls
{GETPHONE, NOPHONE}
More formally, our rule syntax will specify the name of the
attribute, the keywords allows n u l l s , and a list of null representation
types allowed.
Attribute <table>.<attribute> allows n u l l s
{<nullreptype> [, <nullreptype> . . . ] }
The restriction on the type of nulls allowed is meant to disallow the
use of any other type. Therefore, the validation test of this rule is the
test for violators, and we are really testing that no real nulls are used.
This is the SQL statement for this test.
Select * from <table> where <attribute> is null;
Note that in older database systems with no system null, or in text
(delimiter-separated file) data sets, blanks may appear is place of an
"official" null, so we may add this vaHdation test, which grabs the first
character from the attribute and tests for a blank.
Select * from <table> where substring(<attribute>, 1, 1)
If the rule is being obeyed, the size of the result set should be 0. If not,
the result set represents those records that violate the rule.
Selecting records with null values or counting the number of records
with null values is now a bit more complicated, since the null values
allowed are those drawn from the null representation list. Since we have
accumulated those null representations in our null representation table,
the query can be specified either in the general sense (grabbing all records
with nulls) or in the specific sense (getting all records with a "not avail-
able" null type) by using the null representation in the nuUreps table. To
locate all records with nulls, we use this SQL statement.
Select * from <table> where <attribute> in
( s e l e c t nullrep from nullreps where nullreps.name in
C<nullreptype> [, <nullreptype> . . . ] ) ) ;
To specify the selection of those records associated with a specific
null representation, we restrict the null representation list to the specific
representation we care about.
Select * from <table> where <attribute> in
( s e l e c t nullrep from nullreps where
nullreps.name = <nullreptype>);
17.1.5 Non-Null Rules
Similarly, when we specify that an attribute may not be null, the valida-
tion is the same as for the defined null case. This is the syntax of our
original rule.
Attribute employees.empid n u l l s not allowed
The validation test is as follows.
Select * from employees where phone_number i s n u l l ;
Along with the corresponding test for spaces.
Select * from employees where substring(phone_number, 1,
1) = ' ;
If the rule is being obeyed, the size of the result set should be 0. If
not, the result set represents those records that violate the rule.
17.1.6 Value Restrictions
Value restriction rules limit the set of valid values that can be used
within a context. In our definition of a value restriction, we assigned an
alias to the restriction and specified the expression that defines the
restriction. Here is an example.
Restrict GRADE: value >= 'A* AND value <= 'F* AND value
!= 'E'
In general, the form for the value restriction rule includes the key-
word r e s t r i c t , a name assigned for the restriction, and a conjunction
of conditions, where the operators associated with the conditions may
be any drawn from predefined set (+, -, *, /, etc.) or from a user-defined
set of functions (which must be provided for execution, of course).
Restrict < r e s t r i c t i o n name>: <condition> [(AND | OR)
<condition> . . . ]
We actually use this restriction for defining value ranges that could
be associated with functional domains, as well as the restriction as a
representation of the restriction of values for a specific attribute. This is
the format for the latter specification.
Attribute <table>.<attribute> r e s t r i c t by < r e s t r i c t i o n
naine>
This indicates that no values that show up in the named attribute
may violate the named restriction. The validation test for this restric-
tion is a test for violators of the restriction, and the query is composed
by searching for the negation of the restriction. We apply DeMorgan's
laws to generate the negation.
NOT (A AND B) => (NOT A OR NOT B)
NOT (A OR B) => (NOT A AND NOT B)
The where clause of the validation test is generated from the nega-
tion. This would be where clause from our previous example.
NOT ( v a l u e >= 'A' AND v a l u e <= 'F' AND v a l u e != ' E ' )
After the application of DeMorgan's law, it becomes this.
(VALUE < 'A' OR VALUE > 'F* OR VALUE = 'E')
In general the validation test becomes this.
Select * from <table> where NOT <condition l i s t > ;
Again, if the rule is being obeyed, the size of the result set should
be 0. If not, the result set represents those records that violate the rule.
17.1.7 Domain Membership
Recall that we maintain our defined enumerated domains in a set of

database tables. The first table is the domain reference table, which is a
metadata table for all named domains.
c r e a t e t a b l e domainref (
name varchar(30),
dtype char(l),
domainid integer);
The actual values are stored in a single data table, referenced via a
foreign key back to the domain reference table.
domainid int e ger,
value varchar(128));
A domain membership rule specifies that the data values that pop-
ulate an attribute be taken from a named data domain. This is the for-
mat for this rule.
< t a b l e > . < a t t r i b u t e > taken from <domain name>
The validation test for domain membership tests for violations of
the rule. In other words, we select out all the statements whose attribute
has a value that does not belong to the domain. We can do this using a
subselect statement in SQL.
SELECT * from < t a b l e > where < a t t r i b u t e > not i n
(SELECT v a l u e from domainvals where domainid =
(SELECT domainid from domainref
where domainref.name = <domain name>));
not, the result set contains those records that violate the rule. With a
nonempty result set, we can also grab the actual data values that do not
belong to the named domain.
SELECT < a t t r i b u t e > from < t a b l e > where < a t t r i b u t e > not i n
(SELECT v a l u e from domainvals where domainid =
17.1.8 Domain Nonmembership
In a domain nonmembership rule, we indicate via a defined rule that no

values from a specified domain are allowed within an attribute. More
formally, the domain nonmembership rule includes the attribute name,

the keywords not taken from, and the name of the domain.
<table>.<attribute> not taken from <domain name>
The validation test checks for any occurrences of records whose
named attribute contains a value taken from the named domain.
SELECT * from <table> where <attribute> i n
(SELECT value from domainvals where domainid =
not, the result set contains those records that violate the rule. With a
nonempty result set, we can also grab the actual data values taken from
the domain that should not be in the attribute.
SELECT <attribute> from <table> where <attribute> in
17.1.9 Domain Assignment
The domain assignment rule specifies that all the values from the
attribute are implicitly to be included in the named domain. This rule,
which is useful in propagating domain values to the metadata reposi-
tory, must include the keywords define domain, the name of the
domain, the keyword from, and a list of attributes from which the
domain values are gleaned.
Define Domain <domain name> from
(<table>.<attribute> [, <table>.<attribute> . . . ] )
For each of the named attributes, we must have a select statement
to extract those values that are not already in the domain.
SELECT <attribute> from <table> where <attribute> not in
We then must merge all these values into a single set and for each
of the values in this set, create a new record to be inserted into
domainvals table. Each new record will include the identifier of the
domain and the new value, and a SQL insert statement is generated.
Here is the pseudocode for the entire process.
Dom_id = s e l e c t domainid from domainref;
Nonmember s e t S = n u l l ;
For each t a b l e t , a t t r i b u t e a i n a t t r i b u t e l i s t AL do:
S = S union
SELECT a from t where a not i n
(SELECT v a l u e from domainvals where
domainid =
where domainref.name =
dom_id));
For each v a l u e v i n nonmember s e t S do:
I n s e r t i n t o domainvals (domainid, v a l u e )
Values (dom_id, v ) ;
17.1.10 Mapping Membership
Mappings are maintained as metadata using two data tables. The first is
the mapping reference table, into which we store the name of the map-
ping, the source domain identifier, the target domain identifier, a
description, the source of the data, and an assigned mapping identifier.
c r e a t e t a b l e mappingref (
name varchar(30),
targetdomain integer,
mappingid integer
);
The second table actually holds all pairs of values associated with a
particular domain mapping.
c r e a t e t a b l e mappingpairs (
mappingid integer,
);
Mapping membership is similar to domain membership, although

for mapping membership we need to make sure that both elements in a
pair of values exist in the mapping. The format for a mapping member-
ship rule includes a pair of attributes and the name of the mapping to
which it belongs.
(<tablel>.<attributel>, <table2>.<attribute2>) Belong to
Mapping <inappingname> (Propagate | No Propagate)
The validation test is to find those records whose named attributes
have pairs of values that do not belong to the named mapping. This
involves selecting out those records whose composed attributes are not
in the named mapping. This select statement shows how this is done
when both attributes are in the same table, using the string composition
operator "II".
Select * from <table> where
<tablel>.<attributel> || <tablel>.attribute2> not in
(select sourcevalue || targetvalue from mappingvals
where mappingid =
(select mappingid from mappingref where
name = <mappingname>));
This select statement extracts those records that violate the map-
ping membership rule. We can also form a SQL statement to extract
those value pairs that violate the mapping relationship.
Select <tablel>.<attributel>, ".", <tablel>.<attribute2>
from <tablel> where
<tablel>.<attributel> | | <table2>.attribute2> not in
( s e l e c t sourcevalue | | targetvalue from mappingvals
where mappingid =
( s e l e c t mappingid from mappingref where
17.1.11 Mapping Assignment
The mapping assignment rule specifies that all the value pairs from two
specified attribute are implicitly to be included in the named mapping.
This rule, which is useful in propagating mapping values to the meta-
data repository, must include the keywords define mapping, the name
of the mapping, the keyword from, and a list of attributes from which
the mapping values are gleaned.
Define mapping <inapping name> from

(<tablel>.<attributel>, <table2>.<attribute2>)
First, we extract the value pairs from those records that have
attribute pairs not already in the mapping. Here is the single table SQL
statement.
Select * from <tablel> where
<tablel>.<attributel> | | <tablel>.attribute2> not in
where mappingid =
For each of the value pairs in this set, create a new record to be
inserted into the mappingvals table. First, though, we must make sure
that the values actually belong to their corresponding domains. If we
allow the mapping assignment rule to propagate domains, then domain
values that are not already in their respective domains are propagated.
If we do not allow domain propagation, we must validate domain
membership for each of the values
Each new record will include the identifier of the mapping and the
new value, and a SQL insert statement is generated. Here is the
pseudocode for the propagation process, again, using the single table
example.
map_id = s e l e c t mappingid from mappingref where
name = <mappingname>;
Nonmember set S =
Select <tablel>.<attributel>, ",", <tablel>.<attribute2>
from <tablel> where
<tablel>.<attributel> | | <table2>.attribute2> not in
where mappingid =
For each value pair (a,b) i n nonmember set S do:
If (a not in s e l e c t values from domainvals
Where domainid = sourcedomain) then
Insert into domainvals
Columns (domainid, value)
Values (sourcedomain, a ) ;
If (b not in s e l e c t values from domainvals
Where domainid = targetdomain) then

Insert into domainvals
Columns (domainid, value)
Values (targetdomain, b);
Insert into mappingvals
columns (mappingid, sourcevalue, targetvalue)
Values (map_id, a, b);
For the no-propagation process, instead we must test for violation
of the domain membership for the source and target domains and issue
an error message when the attempt at inserting a new mapping value
cannot proceed.
17.1.12 Mapping Nonmembership
Mapping nonmembership specifies that a record's value pair must not

belong to a defined mapping. The format for specifying this rule must
include the attribute names, the key words it does not belong to, and
the name of the mapping.
(<tablel>.<attributel>, <table2>.<attribute2>)
does not belong to Mapping <mapping name>
The validation test for this rule selects out those records where the
attribute pair does belong to the named mapping. Here is the SQL when
both attributes are taken from the same table.
Select * from <tablel> where
<tablel>.<attributel> | | <tablel>.attribute2> in
where mappingid =
W e can also grab all the pairs so that we can investigate the viola-
tions.
Select <tablel>.<attributel>, <tablel>.attribute2> from
<tablel> where
<tablel>.<attributel> || <tablel>.attribute2> not in
(select sourcevalue || targetvalue from mappingvals
where mappingid =
(select mappingid from mappingref where
17.1.13 Completeness
Completeness governs the proper population of a record's fields, possi-

bly depending on other fields' values. A completeness rule is a condi-
tional assertion that specifies a condition followed by a list of attributes
that must have values. Formally, a completeness rule includes the key-
word if, a condition, and a list of attributes that must be non-null if the
condition is true.
If (<condition>) then complete with (<table>.<attribute>
[, <table>.<attribute> . . . ] )
The rule is observed if, for all records, the condition is true. Then
the attributes in the attribute list are non-null. The validation test
checks for nulls if the condition is true. Here is the test in SQL.
Select * from <table> where <condition>
and (<attribute> i s null
[, or <attribute> i s null . . . ] ;
This selects out the set of violating records. If the record set size if 0,
there are no violations.
17.1.14 Exemption
An exemption rule specifies a condition and a Ust of attributes, and if

the condition evaluates to true, then those attributes in the list are
allowed to have null values. Exemption rules, like completeness rules,
are conditional assertions.
Formally, an exemption rule indicates a condition, the keyword
exempt, and a list of attributes, possibly qualified by the kinds of nulls
allowed.
If (<condition>) then exempt
(<attribute>[:<nullreptype>] [, <attribute>
[:<nullreptype>] . . . ] )
In this rule, the attributes in the attribute list may only be null (or
take on specific null representations) if the condition is true. Therefore,
we can infer that the rule is violated if the condition is false and the
attributes contain null values. Therefore, we apply DeMorgan's laws to
the condition and extract those records with nulls in the specified fields.
Here is the SQL to extract nonconformers when there are no null repre-
sentations specified.
Select * from <table> where not <condition> and
(<attribute> i s null
[, or <attribute> i s null . . . ] ;
If we have qualified the attributes in the attribute list using a null
representation, then the where condition is slightly modified to incorpo-
rate the null representation. Here is the same SQL statement with the
null test replaced.
Select * from <table> where not <condition> and
<attribute> in
( s e l e c t nullrep from nullreps where
nullreps.name = <nullreptype>);
17.1.15 Consistency
Consistency rules define a relationship between attributes based on field

content. For example, we can specify a consistency relationship between
the value assigned to an employee's position and that employee's salary.
IF (Employees.title == "Staff Member") Then
(Employees.Salary >= 20000 AND Employees.Salary <
30000)
A consistency rule must specify a condition, followed by an asser-
tion.
If <condition> then <assertion>
The condition may refer to a number of attributes and the asser-
tion may refer to a number of attributes. Records are valid if the condi-
tion is true and the assertion is true. It says nothing about the attributes
referenced in the assertion is the condition is not true. The test for vali-
dation finds those records where the condition is true, but the assertion
is violated. This involves creating a where clause with the condition
ANDed with the negation of the assertion.
Select * from <table> where <condition> AND NOT
<assertion>;
The records set selected are those that violate the rule.
17.1.16 Derivation
A derivation rule is the prescriptive form of the consistency rule. In a

derivation rule, if a condition evaluates to true, then a consequent dic-
tates how another attribute's value is defined. A derivation rule defines a
data dependency between the attributes in the condition and the
attribute in the consequent.
A derivation rule specifies a transformation and can be used in
extraction, transformation, and loading processes. The formal specifica-
tion for a derivation rule specifies a condition followed by an attribute
assignment.
If <condition> then
<table>.<attribute> = <expression>
The expression is some operational combination of data pulled
from other attributes, which may be scalar (refers to a single value) or a
scalar result of an operation on a vector (a SUM or a maximum value
pulled from a column).
The execution representation of a derivation rule tests for the con-
dition and, if the condition is true, generates an update statement
changing the value of the named assigned attribute as a function of the
expression.
Update <table>
Set <attribute> = <expression>
Where <condition>;
If the rule specifies different tables and attributes, the execution is
more likely to be transformed into intermediate code that extracts the
data from the sources, performs the operations, and creates an update
(or insert) into the target table.
17.1.17 Functional Dependencies
A functional dependency between two columns X and Y means that for

any two records Rl and R2 in the table, iffieldX of record Rl contains
value X andfieldX of record R2 contains the same value x, then iffieldY
of record Rl contains the value y, thenfieldY of record R2 must contain
the value y. In other words, attribute Y is said to be determined by
attribute X. Functional dependencies may exist between multiple source
columns. In other words, we can indicate that a set of attributes deter-

mine a target attribute.
Essentially, a functional dependency rule is like a consistency rule
that covers all records in the table. Based on this understanding, we can
see that aside from the standard database use of functional dependence
to ensure normalization, functional dependencies may actually repre-
sent a governing rule about what is represented in the data in other
words, a business rule!
A functional dependency must specify the defining attributes and
the defined attribute, separated by the keyword determines.
(<table>.<S_attribute> [, <table>.<S_attribute>])
Determines (<table>.<T_attribute>)
We will refer to the attributes on the left-hand side of the keyword
determines as the determining attributes and the attribute on the right-
hand side the determined attribute. The validation test for this rule
makes sure that the functional dependence criterion is met. This means
that if we extract the determining values from the set of all distinct
value set pairs, then that set should have no duplicates. First, we extract
the composed pairs of determining, determined attributes from the
table. Then we extract the determining attributes from the table. The
count of the distinct determining attributes must be the same as the
count of the combined attributes or there will be duplicate determining
attributes, which constitutes a violation.
This process provides those determining attributes that are non-
conforming. The next step is to extract those records where the noncon-
formance occurs. We can take the attribute combinations that were
duplicated and extract those records where the determining attributes
match those duplicates.
17.1.18 Primary Key
We can actually ratchet down the constraint from defining a specific pri-
mary key to just defining a key on a table. A key is a set of one or more
attributes such that for all records in a table, no two records have the
same set of values for all attributes in that key set. We specify the key
assertion by listing the attributes that compose the key, along with the
keyword key, and the table for which those attributes are a key.
<table>,<attribute [, <table>.<attribute> . . . ] key for

<table>
The test to validate that this attribute set is a key is to count the
number of records in the table, then count the number of unique occur-
rences of those combined attributes.
Select count(*) from <table>;
Select count ( d i s t i n c t <attribute> [,<attribute> . . .])
from <table>;
The two numbers returned from these two statements should be
equal. If we actually want to extract those multiple record occurrences
with the same set of values that should have been unique, we can use
SQL such as this SQL statement (which can be used for a single
attribute key).
SELECT *
FROM <table> AS t l , <table> AS t 2
WHERE t l . < a t t r i b u t e > = t 2 . < a t t r i b u t e > ;
This Statement looks for instances in the table where a pair of val-
ues taken from a pair of records in the table are the same. If the key
really is unique, the result set will still contain one entry for each record
in the original set, since the selected record will match its value against
itself in both instances of the table. On the other hand, if the value is not
unique, there will be an additional pair of records that share a key.
17.1.19 Uniqueness
The uniqueness rule is a cousin to the primary key rule. The difference
is that we may indicate that a column's values be unique without using
that column as a primary key. We indicate that an attribute's value is to
be unique with a simple assertion.
<table>.<attribute> i s Unique
Our first test for validation is to count the number of records and
then count the number of distinct attribute values. Those counts,
retrieved using these SQL statements, should be the same.
Select count(*) from <table>;
Select count ( d i s t i n c t <attribute>) from <table>;
The two numbers returned from these two statements should be

equal. If we actually want to extract those multiple record occurrences
with the same set of values that should have been unique, there is a sep-
arate known primary key.
SELECT < t a b l e > . < a t t r i b u t e > , < t a b l e > . < a t t r i b u t e >
FROM <table> AS t l , <table> AS t 2
WHERE t l . < a t t r i b u t e > = t 2 . < a t t r i b u t e > and t l . < p r i m a r y >
<> t2.<primary>;
In this case, it is helpful to have a predefined primary key, which can
then be used to extract out the offending pairs more easily than in the
earlier case of defined keys.
17.1.20 Foreign Key
A foreign key represents a relationship between two tables. When the

values in field f in table T is chosen from the key values in field g in
table S, field S.g is said to be a foreign key for field T.f. A foreign key
assertion must indicate that a field in one table is a foreign key for a
field in another table.
We specify that a set of attributes is a foreign key by specifying the
attributes that act as the foreign key and the attribute that the key
indexes.
(<target table>.<attribute>) foreign key for
(<source table>.<attribute>)
The referential integrity constraint, which is implicit in the defini-
tion of a foreign key, indicates that if table T has a foreign key matching
the primary key of another table S, then every value of the foreign key
in T must be equal to the value of some primary key in table S or be
completely null. Therefore, a foreign key assertion specifies a consis-
tency relationship between tables for all non-null foreign keys that
appear in table T, there must exist a primary key with the same values in
table S.
We test this assertion by making sure that all values that appear in
the foreign keyfieldof the source table exist in target table.
Select * from <source table> where <attribute> not in
(Select distinct <attribute> from <target table>);
This SQL statement pulls out all the values that are supposed to be
used as a foreign key but do not appear in the target table. These are vio-
lating records, and the question is whether the violation occurred because
the foreign key value in the source table was incorrect or whether the for-
eign key is really missing in the target table and should be there.
17.1.21 Alternate Implementation The Rules Engine
In this section, we have explored the use of SQL as a way of both exe-
cuting and representing data quality and business rules. We should not
feel obligated to only use this mechanism as a tool for databases. Even
though a rule is described using SQL, the implementation need not be
restricted to an environment where the data sit in a relational database.
The use of intermediate data representations, the use of standard data
structures such as arrays, linked lists, and hash tables, and other pro-
gramming tricks will allow for the execution of these rules in a runtime
environment separated from a query engine.
We like to think that the kinds of rules engines that we saw in
Chapter 12 can be called upon to create intermediately executing rule
validation objects. We can encapsulate the interpretation and operation
of a validator as an operational execution object that can be inserted
into a processing system, and we use this idea in the next section.
As long as the rules engine can be handed a "map" of the data that
pass through it either through the use of metadata for table schemas
that can be provided into the rules engine or a data type definition and
message format schema, such as that which can be described using
markup systems like XML the rules engine can manage the testing
and validation of data quality rules when integrated into the processing
stream. In this way, coordinating with actions that can be linked in
together with the rules engine, a content-oriented workflow system can
be enabled.
17.2 OPERATIONAL DIRECTIVES
In Chapter 8, we also looked at a number of operational directives

that were used either to specify the model of the information chain or
an insertion of an operational procedure into the operation. In this
section, we look at the relation of these directives to an actual process-

ing environment.
17.2.1 Information Chain Specification
The specification of an information chain is embodied through the iden-

tification of the stages of processing within the system and the informa-
tion channels between these stages. The actual benefit of this kind of
specification occurs when the specification is in much greater detail,
incorporating inputs and outputs that can be fully specified using a data
model or a data interchange definition. This benefit allows for the inser-
tion of a validation engine at specific points in the information chain, as
will be described in Section 17.2.2.
Suffice it to say, the information chain specification defines a data
flow model for information throughout the system and an interface
specification for data transfer. The directives specify either a processing
stage or an information channel, as well as a qualified name describing
the effects of information passing from one point in the model to
another.
Each location in this graph can be identified as a source for data, a
target for data, a processing stage for data, or a communications chan-
nel. The implementation of this model is the construction of a runtime
model of the information flow, where each named object is incorpo-
rated into a full map of the system. Each location refers to the corre-
sponding actual location in the information chain, and at any point
where data is written or read, the data model is maintained. The rest of
the directives in this section will refer to the actual processing locations,
as referenced via this information map.
We will store the information chain in a metadata table. First, we
have a small data domain specifying the kinds of information chain
objects.
Define infochain_objtype {"SOURCE", "TARGET", "CHANNEL",
"STAGE"}
We then have a table describing the objects in the information
chain. Each record contains a type, a name, a source and a target (if it is
a channel), and an assigned object id. In addition, each record has a
description field, which will cover the description of the actual location
in the processing system that corresponds to this information chain
object.
DATA QUAUTY AND BUSINESS RULES IN PRACTICE 447
Create table infochain (

Objtype varchar(lO),
Objname varchar(50),
Desc varchar(lOOO),
Source integer,
Target integer,
Objid integer
);
Actually, this is a good opportunity to specify some of our data
quality rules.
Infochain.objtype belongs to infochain_objtype;
I f (infochain.objtype = "CHANNEL") then complete with
(infochain.source, i n f o c h a i n . t a r g e t ) ;
I f (infochain.objtype <> "CHANNEL") then exempt
(infochain.source, i n f o c h a i n . t a r g e t ) ;
17.2.2 Measurements
Our goal of a measurement directive is to (1) create an opportunity to

insert a rule validation engine at a specified point in the information
chain, (2) take a measurement, and (3) record the measurement to a
separate data resource.
To this end, we must define another set of data tables, which
describes measurements taken based on a rule set executing at a loca-
tion within the information chain. These tables can be used for as sim-
ple a process as collecting conformance percentages to one as complex
as maintaining a reference to all violating records (or messages).
This database is used to keep track of the date and time the mea-
surement was taken, the location in the information chain where the
measurement was done, the number of records (or messages) tested, the
rule set referenced, and the number of passes and fails per rule being
measured. The first table maintains the record of which measurements
are being taken and their location on the information chain. This corre-
sponds to the location for the measurement and the rule being measured.
Create table measures (
Measureruleid integer,
Locationid integer,
Rule integer,
);
The next table models each actual measurement.

Create table measurements (
Measurementid integer,
Measurementdate date,
Ruleset integer,
Numtested integer,
Passed integer.
Failed integer
);
For the details, we refer to a more descriptive data set.
Create table measurementdetails (
Measurementid integer.
Rule integer,
Numtested integer.
Passed integer,
Failed integer
);
The measurement identifier refers back to the governing measure-
ment in the measurements table. Each record in measurementdetails
refers to a particular rule in the referenced rule set, the number of
records or messages that were tested, and the numbers of passes and
failures.
A measurement directive indicates the location where the measure-
ment is to take place and the rule that is to be tested. The implementa-
tion of this is the creation of a rule tester, integrated with the specific
rules and inserted in the information chain at the named location. The
rule tester will execute this pseudocode for each record.
Passed[l:num_rules] = 0;
Failed[l:num_rules] = 0;
Tested[l:num_rules] = 0;
For each record do
For each rule do
If this record applies to this rule, then {
Tested[this rule]++;
If this rule passes then
passed[this rule]++;
Else
Failed[this rule]++;
}
total_tested = sum(tested[l:num_rules]);
total_passed = sum(passed[l:num_rules]);
total_failed = siim(failed[l:niimrules]) ;
Insert record into measurements table;
For each rule,
Insert detail record into measurementdetails table;
17.2.3 Triggers
A trigger directive indicates the invocation of a named operation at

some point in the data-flow chain. We may include a condition to trig-
ger the operation, or the operation may take place specifically when a
record or a message hits a certain point in the data-flow chain. A trigger
directive is implemented similarly to the measurement directive: an
operational object inserted into the actual processing system, as mir-
rored in the placement in the information chain.
Create table t r i g g e r s (
triggerid integer,
Locationid integer,
Rule integer,
action integer
);
Note that we have glossed over the details of the actions being
taken. We assume that a separate system can be created to manage
described actions, and we are referring to those actions.
The operational object, when integrated into the actual processing
system as mirrored by the information chain, will test the rule, and if
the rule evaluates to true, the action is taken.
17.2.4 Transformations
Although we have specified transformations as separate directives, we

can treat them as a syntactically different but semantically equal version
of the derivation rule as described in Section 17.1.16.
17.2.5 Updates
Updates refer to the initiation of an ongoing process and the restarting

of a process, as related to both triggers and measurements. We might
want to integrate a rule-oriented reload of a data mart from a data
warehouse, based on a trigger indicating that the conformance to a data
quality rule set has finally measured above the acceptance threshold.
These update directives can be integrated as inlined code or as periodic
refreshes using the system's timer and scheduling utilities.
17.2.6 Approximate Match
In Chapter 16, we defined two additional rules. The first of these rules,
the approximate match definition, defines a similarity score between
any two values, along with a threshold indicating whether two values
are assumed to match. The formal definition of an approximate match
rule must include a defined function, which should evaluate to a real
value between 0.0 and 1.0 and a threshold value above which a score
indicates a match, which is also a real value between 0.0 and 1.0.
APPROXIMATE MATCH
C<real function>C<valuel>, <value2>)):threshold=0.75
The definition of the function can include predefined operators
(such as the same ones used for expressions) as well as user-defined
functions (which of course, must be provided to build a working appli-
cation). The actual implementation of this rule converts the rule defini-
tion into a function. Here is an example of the conversion of this rule
match_defl: approximate match (1 - (ABSCValuel-
Value2)/MAXCValuel, Value2))):threshold=0.75
into C code:
#define MAX(_a,_b) (_a > _b ? _a :_b)
int match_defl(float _vall, float _val2) {
if (1- (abs(_vall - val2)/MAXCvall. _val2))) >= 0.75)
return(l);
else
return(O);
}
We can instead implement this rule as a function returning the score

itself and testing the score outside of the function. This in fact is more
interesting, since the function can then be used with the next rule.
17.2.7 Approximate Aggregate
The approximate aggregate rule collects a sum of scaled results associ-

ated with approximate match scoring. An approximate aggregate asso-
ciated with a pair of records will be a sum of a set of weights applied to
a collection of approximate matches, divided by the sum of the weights,
along with a threshold indicating a score above which the two records
are assumed to be a match.
Approximate a g g r e g a t e ((<matchA>(<valueAl>, <valueA2>) *
<scaleA>
[ + <matchX>(<valueXl>, <valueX2>) * <scaleX> . . . ] )
/ (sum (<scaleA> [, <scaleX> . . . ] ) ) : <threshold>
This rule is implemented also as a function, which incorporates the
approximate match rules used within. The result is always a floating
point number between 0.0 and 1.0.
17.3 DATA QUALITY AND THE TRANSACTION FACTORY
Now that we have a clearer story of how these data quality and busi-
ness rules can be used in an executing system, let's look at some specific
implementation paradigms. Our first is using data quality rules to help
guide the execution in a transaction factory.
In a transaction factory, a transaction, such as a stock trade or a
product purchase, takes place, and the factory processes the transaction
so that a desired result occurs. The transaction factory takes transaction
records as well as other data as raw input, produces some product or
service as a "side effect," and perhaps may generate some output data.
Each transaction factory can be represented as an information chain,
with individual processing stages taking care of the overall transaction
processing.
Data quality and business rules are integrated into the transaction
factory in two ways. First, transaction factories are built assuming that
the data being inserted into the process is of high quality and meets all
the expected requirements. Second, the actual processing that is per-

formed is based on the content of the inserted data. Together, we are pre-
sented with three opportunities to use data quality and business rules.
17.3.1 Input Validation
Because the factory is built assuming that the input meets a high stan-
dard, problems in continuing operations occur when the input does not
meet that standard. If a bad record can adversely affect the streamlined
processing, then it is beneficial to capture, remove, and correct offend-
ing data before it ever reaches the internals of the factory.
In this instance, data quality rules would be defined to match the
input requirements. A validation engine would be inserted at the entry
point for all data inputs. In addition, we must also define a reconciliation
process to direct the way that invalid records are treated (see Figure 17.2).
The implementation consists of these steps:
1. Define the data quality rules based on the input requirements.
2. For each rule, associate a degree of criticality, which will dictate
the action taken if the rule is violated.
3. For each rule and degree of criticality, determine the action to be
taken. Typically, we can make use of at least these actions.
a. Ignore: If the violation is not business critical, we may want
to log the error but let the value pass.
b. Auto-correct: We can define a derivation rule that uses the
current value of afieldto determine the actual value, if that is
possible. If a record is auto-corrected, it can then be gated
through into the system.
c. Remove for reconciliation: If the violation is severe enough
that it will affect processing, and it cannot be (or should not
be) auto-corrected, the record should be shunted off to a rec-
onciliation database, enhanced by the rule that was violated,
the data source, and a timestamp.
4. An implementation of a rule validator is built and inserted at the
appropriate entry points of data into the factory.
5. The reconciliation database needs to be constructed and made
operational.
6. A reconciliation decision process is put into effect.
Input dau is
checked for
conformance with
expectations.
Data
Data entered via input screens
z
If the data docs
not pass mustei;
the user is notified
and the correct
If the information
information is
is valid, it is
requested.
propagated into
the database.
FIGURE 1 7 . 2 Using rules for input validation
17.3.2 Supplier Management
Inserting a validation engine at the source of input data also generates a

benefit of logging the errors coming from each information supplier, as
well as calculating the compliance percentage each time data passes
through the engine. By augmenting the validation engine with a mea-
surement directive, we can also accumulate statistics about data quality
conformance, which can be presented to the supplier if the data does not
meet the requirements. Through the implementation, we can build an
automated means for generating a compliance report (see Figure 17.3).
All supplied data

is subject to a
validation test
before it is
imported into the
enterprise.
Rules
Data Supplier Enterprise
Data Validation Buffer Zone
Data is only gated

dirough if it
conforms to the
e3cpected levels of
data quality.
FIGURE 1 7 . 3 Using rules for supplier management
17.3.3 Content-Directed Processing
Last, at each decision point in the processing, it is possible that the deci-
sion is based on the content of the data record as it passes through a
point in the information chain. Again, we can associate actions for data
routing based on the values embedded in the records (or messages) as
they pass through locations in the information chain. We can then cre-
ate rule application engines that make use of these rules (and their asso-
ciated actions) at the corresponding location within the information
chain, directing the information flow.
17.4 DATA QUALITY AND THE DATA WAREHOUSE
We have already discussed how complex and pervasive the data quaUty
problem is with respect to data warehousing. In Chapter 3, we looked
at an example where we would use data quality and business rules for
what we call "data warehouse certification." Certification is a means
for scoring the believability of the information stored in a data ware-
house. We certify a data warehouse as being fit for use when the data
inside conform to a set of data quality expectations embodied in a set of
rules. Given these rules, we assign a score to the quality of the data
imported into a data warehouse for certifying warehouse data quality.
In this chapter we can see how this can be implemented. A set of
rules is developed as a quality gate for data imported into the data
warehouse. We will associate with each rule a validity threshold (as a
percentage) based on the users' expectations of quality. An engine is
configured to incorporated those rules and execute the validation tests
as data is prepared to be entered into the warehouse (see Figure 17.4).
Data Quality
Rules
Record is valid
Data Warehouse InpuQi Rules Engine Data
Vi Ktpomag
Warehouse
Record
Thresholds /Warehouse\
exceeded \ certified
Thresholds not met
FIGURE 1 7 . 4 Data warehouse validation

As records are fed into the engine, any relevant rules (that is, any
rules that refer to values of attributes defined within the record) are
tested. We create a measurement object, which tallies the successes and
failures associated with each rule and outputs the results to the mea-
surements tables. We incorporate a trigger to notify the users whether
the warehouse has been certified or not.
For any record, if no rules fail, the record is said to be valid and is
successfully gated through to the warehouse. For each rule that does
fail, a record is generated to be inserted into the measurements tables
with the information about which rules were violated. The record is
then output to a reconciUation system, as we described in Section 17.3.
The violating record can also be passed through to the warehouse, but
now it should be timestamped and marked as having not conformed to
the users' expectations, and this information can be used when per-
forming analysis.
After you've imported the data, each rule's validity value is com-
puted as the ratio of valid records to the total of records. A data quality
certification report delineating all validity percentages is generated. If
all validity percentages exceed the associated thresholds, the warehouse
is certified to conform to the users' data quality requirements. Other-
wise, the warehouse is not certified, and until the percentages can be
brought up to the conformance level, the warehouse cannot be said to
meet the data quality requirements.
Since we have integrated the measurement process into the ware-
house loading system, we will have a periodic (or as periodic as the
warehouse is loaded) measure of how well the database conforms to the
data quality requirements. In order to qualify the warehouse after a
failed certification, the records output to the reconciliation system must
be analyzed for the root cause of the failures, as explored in Chapter 15.
After reconciliation, the data is resubmitted through the rules engine,
and the validity report is generated again. This process continues until
certification is achieved.
17.5 RULES AND EDI
In Chapter 3, we discussed some of the issues associated with electronic

data interchange (EDI) and data quality. We focused on the use of XML
(the Extensible Markup Language) as a framework for both structure
definition (for the definition of data standards) and rules for validating
messages based on defined standards. We discussed well-formedness

and validity within the XML environment.
In the context of XML, a document is well formed if it conforms to
the syntax of XML. This means that an XML parser can successfully
parse the document. Also, in order to be well formed, a document must
meet a number of other conditions. For example, no attribute (a name-
value pair that attributes a tag) may appear more than once on the same
start tag, all entities must be declared, and nonempty tags must be
nested properly. Another condition is that neither text nor parameters
may be directly or indirectly recursive. An XML parser can determine
whether a document is well formed or not.
An XML document is valid if it is well formed and it contains a
DTD and the document obeys the rules specified in the DTD. This means
that element sequence and nesting conforms to the DTD, all attribute
values that are required are present, and their types are correct.
What is still needed for information validation within an EDI envi-
ronment is content-based validation. This can be achieved when using
our data quality and business rules to specify content-oriented asser-
tions about the information contained within the messages. We can
direct the processing of XML statements via processing instructions
within an XML document, and that is the hook through which we can
integrate the data quality rule engines.
17.6 DATA QUALITY RULES AND AUTOMATED UIS
One of the major goals of a data quality program is the enforcement of

strict rules conformance before information passes into the system.
Since many data collections are populated through manual data entry, it
would be preferable to create a framework that limits the kinds of
errors that can be introduced at data entry time. We can use our data
quality rules to help automate the generation of a user interface (UI)
based on maintaining a high level of data quality.
17.6.1 Selection Criteria
The definition of data domains, mappings between those domains, and

the assignation of rules associating attributes with domains and map-
pings prove to be an invaluable resource in UI generation. When we
declare that an attribute draws its values from a defined domain, we

automatically limit the possible values that the user may enter. If we
already have those valid values, why not integrate those as the selection
choices within a drop-down box whenfillingin an online form?
17.6.2 Dependence-Based Data Entry
More interestingly, when we have declared that two attributes' values

must conform to a predefined mapping, we impose a dependence on the
values that can populate the target attribute based on the selection for
the source attribute. We can create the drop down for the source
attribute's value, and when that value is selected, only those values for
the target attribute that are valid under the mapping are presented to
the user for selection.
This introduces a more interesting concept associated with the use
of data quality rules for UI generation. The definition of a set of data
quality rules may create a hierarchical arrangement of dependence
between the different data attributes within a database table or a mes-
sage. For example, associating two attributes with a domain mapping
creates a dependence for the target attribute on the source attribute. In
other words, we can ensure that the relationship between these two
attributes as expressed by the data quality mapping membership rule
can be enforced from the data entry point if the value for the target is
not collected until the value for the source has been provided!
This simple statement is actually quite profound. It covers all the
relationships expressed within data quality rules and basically states that
we can ensure a higher level of data quality from the start if we abide by
the dependences embedded in the data quality rules. In any rule that
incorporates a conditional assertion, we can infer a dependence. For
example, the completeness rules defines that if a condition that involves
the value or values of data attribute(s) is true, then that record is com-
plete only if other attributes are non-null. This means that there is a con-
ditional dependence on the attributes referenced in the condition. The
same holds true for exemption rules, consistency rules, and so on.
To carry our example a bit further, let's presume that there is a
completeness rule indicating that if the data attribute OWNS_CAR is
"y," then the record is complete with the CAR_REGISTRATION^
NUMBER attribute. From our discussions, we know that if the user
enters a "y" in the OWNS_CAR field, then the record will not be com-
plete unless the other attribute is given a value. This tells us when the
field must be filled in and when it doesn't have to be filled in. This is
shown in Figure 17.5.
17.6.3 Building the Dependence Graph
The insight we gain from the idea of dependence is that when many
attributes in a table depend on other attributes in the same (or another)
table, we can group our requests for data by what is called a depen-
dence class. All attributes are assigned into a dependence class with
some integer-valued degree. The zero-th degree includes all nondepen-
dent attributes. The first degree includes all attributes that depend only
on zero-th degree attributes.
More formally, a dependence class of degree i is the set of all attrib-
utes whose values depend on attributes of a degree less than /. At each
stage of data collection, we only need to ask for data for the current
dependence class, and we can impose our rules on the attribute values
being collected at that point because all the conditions for our rules are
known to be able to be examined. The relation between different
dependence classes can be summarized in a dependence graph.
1 productCode DOES NOT ALLOW NULLS
2 If (substr(productCode,l,l) != 'X') then exempt {color}
3 If (color == 'GREEN' OR color == 'RED') then (size != 'XXL')
4 If (substr(productCode,l,l) == 'A') then (qty >= 100)
FIGURE 1 7 . 5 Examples of data dependence

17.6.4 Creating the Ul
To create a user interface, we will generate input forms that are data-
directed. The first form to be presented asks for all zero-th degree
attributes. Each subsequent form can be crafted to incorporate only
those attributes in the next dependence class and can be attributed in
format with the data attributes associated via the data quality rules.
To continue our previous example, if the user did indicate that he or
she owned a car, the next form could explicitly say, "You have indicated
that you own a car. What is the registration number for that car?" and
then prompt for the answer. This offers more context to the user than the
traditional form of emptyfields.Figure 17.6 shows this process.
17.7 SUMMARY
In this chapter, we have examined the ways to create executable systems

driven by the data domains, domain mappings, and data quality rules
discussed in Chapters 7, 8, and 16. We have seen how the rules can be
turned into SQL queries, and we have opened the discussion for how
the rules can be integrated into a rules engine such as the kind discussed
in Chapter 12.
We also have a notion as to how we can integrate approximate
matching and scoring rules into "fuzzy" cleansing applications through
the combination of approximate rules with other rules. Including mea-
surements and triggers allows us to describe a full-fledged data quality
application before writing a stitch of code.
FIGURE 17.6 Data dependence-driven UI creation

Finally, we looked at some implementation paradigms of data

quality rules. This chapter brings the story full circle from the begin-
ning, where we first learned about the quantification of a data quality,
assessing the cost of low data quality, the dimensions of data quality,
through domains, mappings, rules, and measurements. At this point,
we can make use of our analysis tools to completely assess, develop a
return-on-investment model, specify, and implement a data quality
improvement program. The next chapter provides a complete overview
of the entire process from start to finish.
18
BUILDING THE DATA QUALITY
PRACTICE
In Chapter 1, we talked about the ideas of knowledge management and

how data quality management is the basis of a successful knowledge
organization. Our goals in this book have been to elucidate the topic of
enterprise knowledge management from first principles: building a
quality core knowledge infrastructure composed of high-quality meta-
data, data domains, data quality rules, and business rules.
To bring things full circle, we must provide a concise guide to the
data quality approach to knowledge management. This last chapter
brings together the methods and processes that we have discussed in
this book, augmented with some of the operational details needed in
building a successful data quality practice.
18.1 STEP 1: RECOGNIZE THE PROBLEM
Thefirststep is problem recognition. The problem is only a data quality

problem when it becomes clear that some issue associated with data
quality has an adverse effect on the business. This boils down to seeing
evidence of a data quality problem andfinding(at a high level) the areas
that are affected by that problem.
18.1.1 Evidence of a Problem
In Chapter 4, we looked at these issues that might indicate the evidence

of a data quality problem.
463
Frequent system failures and system interruptions

Drop in productivity versus volume
High employee turnover
High new business to continued business ratio (high customer
turnover)
Increased customer service requirements
Decreased ability to scale
Customer attrition
When these issues are evident in an organization, it is a good sign
that somewhere, poor data quality is having an effect. The next step is
to determine what kinds of problems exist and where.
18.1.2 Determining the Problem
Through employee and customer interviews, the analyst's next step is to

determine whether the noted evidence points to any particular problems
with data quality. Typically, corroborative evidence can be collected via
anecdotes, "crisis events," or customer service logs.
Part of this stage involves reconnoitering the management struc-
ture. Knowledge of the identity of the senior-level managers with
authority over the information resource is important when establishing
the need for a data quality program.
18.2 STEP 2: MANAGEMENT SUPPORT

AND THE DATA OWNERSHIP POLICY
No data quality program will be successful without the support of

senior managers. Because poor data quality affects both the operational
and the strategic sides of the business, addressing a systemic data qual-
ity program takes on a strategic importance that must be validated by
the participation of the management of the business.
There are two parts to this phase. The first is making the proper
presentation that highlights the importance of data quality within the
organization. The second is the definition of and the securing the
enforcement of a data ownership policy.
BUILDING THE DATA QUALITY PRACTICE 465
18.2.1 Gaining Senior-Level Management Support
The best way to win the support of senior management is to cast the
problem in terms of how the business is affected or can be affected by
poor data quality. A presentation can be crafted that incorporates the
following notions.
The reliance of the organization on high-quality information
The evidence of existence of a data quality problem
The types of impacts that low data quality can have
The fact that managing data quality is the basis of a knowledge
organization
A review of the anecdotes regarding poor data quality
A review of data ownership issues
The implementation of a data ownership policy
A projection of Return on Investment (ROI)
There are two goals of this presentation. Thefirstis to encourage an
awareness on behalf of the senior managers of the importance of data
quality. The second is the authority to craft a data ownership policy.
18.2.2 Data Ownership Responsibilities
Our data ownership policy will define a set of data ownership roles and
assign responsibilities to those roles. Here are some of the responsibili-
ties we discussed in Chapter 2.
Data definition
Authorization of access and validation of security
Support the user community
Data packaging and delivery
Maintenance of data
Data quality
Management of business rules
Management of metadata
Standards management
Supplier management
18.2.3 The Data Ownership Policy
The data ownership poUcy is the document guiding the roles associated
with information and the responsibilities accorded those roles. At the
very least, a data ownership policy should enumerate these elements.
1. The senior-level managers supporting the enforcement of the
policies enumerated
2. All data sets covered under the policy
3. The ownership model (that is, how is ownership allocated or
assigned within the enterprise?) for each data set
4. The roles associated with data ownership (and the associated
reporting structure)
5. The responsibilities of each role
6. Dispute resolution processes
7. Signatures of those senior level managers listed in item 1
18.2.4 Complicating Notions
Keep in mind that there are complicating notions that are working as
forces against the smooth transition to a knowledge organization,
including these.
Questions of information value
Privacy issues
Turf and control concerns
Fear
Bureaucracy
18.2.5 Defining the Data Ownership Policy
These are the steps in defining the data ownership policy (see Figure
18.1).
1. Identifying the interested parties or stakeholders associated with
the enterprise data. This includes identifying the senior-level
managers that will support the enforcement of the policy.
2. Cataloging the data sets that are covered under the policy
3. Determining the ownership models in place and whether these
are to continue or whether they will be replaced or modified
4. Determining the roles that are in place, those that are not in
place, assigning responsibilities to each role, and assigning the
roles to interested parties
5. Maintaining a registry that keeps track of policies, data owner-
ship, roles, responsibilities, and so forth
18.3 STEP 3: SPREAD THE WORD
Another key factor to the success of a data quality program is evange-

lism: making sure that the stakeholders (including the users, managers,
and implementers) are aware of the value and importance of data qual-
ity. Here we discuss two aspects: creating a data quality education
forum and what topics should be included.
18.3.1 Data Quality Education
We suggest a training program that can be broken into two parts. The
first part covers the "business" aspects of data quality, such as the eco-
nomic analysis, the cost of low data quality, assessments, and building
ROI models. The second part covers the implementation issues, such as
data domains, mappings, data quality and business rules, measure-
ments, data cleansing, correction, and enhancement.
18.3.2 Training in Data Quality Systems
A training program in data quality should incorporate at least these

concepts.
1. Creation and use of information
2. Storage of information
Determine ^V ^ Maintain a >

Identify \ / \ / Determine roles and assign V _ _ ^ registry that tracks
r^lr^holXr.
stakeholders YM Catalog data sets U W ownership responsibilities J V ownership
models
,^^ to roles ^^ \ ^ ^ poliqr ^/
FIGURE 1 8 . 1 Defining the data ownership poUcy

3. Data ownership
4. Quality concepts and the quahty improvement cycle
5. Understanding the economic impact of data quality issues
6. Dimensions of data quality
7. Aspects of reference data domains
8. Data quality and business rules
9. Metrics for measuring and assessing data quality
10. Metadata
11. Data quality requirements analysis
12. Data cleansing and standardization
13. Error detection, correction, and root cause analysis using data
quality rules
14. Data enhancement
18.4 STEP 4: MAPPING THE INFORMATION CHAIN
Understanding how information flows through the organization is the

next step in improving data quality. The information chain (see Figure
18.2) is a chart describing data movement through the enterprise and is
used as the basic map on top of which our other analyses are performed.
An information chain is composed of processing stages and com-
munications channels. To map the information chain, first identify all
the processing stages, then the channels of communication between
them, and then label each unique component in the map.
18.4.1 Identifying Processing Stages
We look for those stages that create, read, write, send, or process data.
Here is a review of the processing stages discussed in Chapter 4.
Figure 18.2 Mapping the information chain


ing stage.
stage.
required is called a decision-making stage
made at a decision-making stage is executed, which may affect
other processing stages or a data delivery stage.
9. Data consumption Because the data consumer is the ultimate
user of processed information, the consumption stage is the exit
stage of the system.
18.4.2 Identifying Communication Channels
An information channel is a pipeline indicating the flow of information

from one processing stage to another. A directed information channel is
additionally attributed with the direction in which data flows.
18.4.3 Building the Information Chain
The information chain is a graph, where the vertices are the processing
stages and the communication channels are directed edges. Every vertex
and every edge in the information chain is assigned a unique name.
18.5 STEP 5: DATA QUALITY SCORECARD
The data quality scorecard summarizes the overall cost associated with
low data quality and can be used as a tool to help determine where the
best opportunities are for improvement.
18.5.1 Impacts of Low Data Quality
Low data quality can have impacts that affect the way the operational
and strategic environment run. In Chapter 4, we explored these impacts
of low data quality.
Detection of errors
Correction of errors
Rollback of processing
Rework of work already completed under erroneous circum-
stances
Prevention of errors
Warranty against damages caused by nonconformities
Reduction of customer activity
Attrition and loss of customers
Blockading on behalf of angered ex-customers
Delay of decisions
Preemption of decision making
Idling of business activity while waiting for strategy to be defined
Increased difficulty of execution
Lost opportunities
Lack of alignment between business units
Increased acquisition overhead associated with information
products
Decay of information value
Infrastructure costs to support low data quality
18.5.2 Economic Measures
Each of these impacts must be associated with some economic measure

that can be quantified in terms of actual costs. Our starting points for
cost estimation for either an impact or the benefit if the impact were
removed are the following.
Cost increase This measures the degree to which poor data

quaUty increases the cost of doing business.
Revenue decrease This measures how low data quaUty affects
current revenues.
Cost decrease This measures how an improvement in data qual-
ity can reduce costs.
Revenue increase This measures how improving data quaUty
increases revenues.
Delay This measures whether there is a slowdown in productivity.
Speedup This measures the degree to which a process's cycle
time can be reduced.
Increase satisfaction This measures whether customer satisfac-
tion, employee satisfaction, or shareholder satisfaction is
increased.
Decrease satisfaction This measures whether customer satisfac-
tion, employee satisfaction, or shareholder satisfaction is
decreased.
18.5.3 Building the Scorecard
To create a data quality scorecard (see Figure 18.3), we follow these steps.
Map the information chain to understand how information flows
within the organization.
Interview employees to understand what people are doing with
respect to data quality issues.
Interview customers to understand the kinds of customer impacts.
Isolate flawed data by reviewing the information chain and locat-
ing the areas where data quality problems are manifested.
FIGURE 1 8 . 3 Creating the data quality scorecard

Identify the impact domain associated with each instance of poor

data quality.
Characterize the economic impact based on the uhimate effects of
the bad data.
Aggregate the totals to determine the actual economic impact.
18.6 STEP 6: CURRENT STATE ASSESSMENT
The data quality scorecard highlights the effect of low data quality on
the bottom line. We must perform a current state assessment to collect
enough data to understand the nature of the actual data quality prob-
lems (see Figure 18.4).
18.6.1 Choosing Locations in the Information Chain
We begin the current state assessment by picking particular locations in

the information chain as targets for measurements. We take into account
these characterizations when selecting a location for measurement.
1. Critical junction Any processing stage with a high degree of
information load is likely to be a site where information from dif-
ferent data sources is merged or manipulated.
2. Collector A collector is likely to be a place where information
is aggregated and prepared for reporting or prepared for storage.
3. Broadcaster A broadcaster is likely to be a processing stage
that prepares information for many consumers and, therefore,
may be a ripe target for measurement.
Choose ^ ^ Describe ^ S , .
Choose a / Measure \ / Prepare
locations f expectations \
subset of the
in the for data quaHty at each 14 data j - 1 assessment
data quality
information ^ location for each J \ quality / \ report
dimensions
chain X ^ dimension ^^
FIGURE 1 8 . 4 Performing the current state assessment

4. Ease of access While some processing stages may be more

attractive in terms of analytical power, they may not be easily
accessed for information collection.
5. High-profile stages A processing stage that consumes a large
percentage of company resources might provide useful measure-
ment information.
18.6.2 Choosing a Subset of the Data Quality Dimensions
The next stage of the current state assessment is selecting a subset of the
dimensions of data quality for measurement. We suggest that at a mini-
mum, selecting at least one dimension from each of our five classes of
data quality dimensions.
18.6.3 Measuring Data Quality
In Chapter 9, we reviewed many different ways that each of the dimen-

sions of data quality defined in Chapter 5. The real directive at this
point is to choose a way to measure the conformance to expectations
associated with the selected dimensions of data quality.
18.7 STEP 7: REQUIREMENTS ASSESSMENT
After completing a current state assessment, we have some ideas of the

scope and magnitude of the different problems, and we want to narrow
the scope of each specific problem into a manageable unit of work. We
can do this by prioritizing each problem, assigning responsibility, and
creating data quality requirements.
18.7.1 Distribution of Cost and Impact
At this stage, we combine the result of the data quality scorecard, which
attributed the information chain with the impacts and costs associated
with low data quality, and the current state assessment, which attrib-
uted the information chain with the measured levels of data quality. The
result is an assignment of a percentage of the overall cost and impact to

each of the problems so that we can order them by an assigned priority.
18.7.2 Assignment of Responsibility
Simultaneously, it is important to refer to the data ownership policy and

assign ultimate responsibility for each individual problem. That person
will then be tasked with defining the project that brings the level of
quality in line with the expectations.
18.7.3 Data Quality Requirements
Each specification of a data quality requirement must include the fol-

lowing.
A unique identifier for the requirement
A name for the requirement
The name of the responsible party
The location in the information chain where the requirement
is applied
A reference to the dimension of data quality that is being
measured
A description of the measurement method
If possible, the measurement rule
The minimum threshold for acceptance
The optimal high threshold
The scaling factor as a percentage of the overall system data quality
18.8 STEP 8: CHOOSE A PROJECT
We cannot expect to solve the entire enterprise data quality problem at

once, and so the best way to attack it is piece by piece. We have ana-
lyzed the cost of low data quality, measured the different problems that
exist, determined the source of the individual problems, and assigned a
percentage of the cost, impact, and responsibility to each problem.
Now is the time to select a problem to solve.
Choosing the first problem to address requires some care, since the
results of implementing the first project can make or break the practice.
The first project should reflect these ideas.
The problem to be solved has a noticeable impact.
Solving the problem results in measurable cost savings.
There are little or no political issues that need to be addressed.
There is senior management support.
Access to the problem space is open.
The problem can be solved.
Remember: The goal of the first data quality project is to ensure
the continued operation of the data quality program. A failure in solv-
ing the first project will probably result in the demise of the program, so
choose wisely.
18.9 STEP 9: BUILD YOUR TEAM
The analysis is complete. The project has been selected. Now it is time
to solve the problem. Each solution project will need to have a team of
people that can execute the different parts of the job. In this section, we
describe these roles.
18.9.1 Project Manager
The project manager's job is to define the set of tasks to be performed,

create a schedule for the implementation, assign the tasks to the differ-
ent team members, and enable the team members to do their jobs.
18.9.2 System Architect
In solving data quality problems, the role of the system architect is both
system analyst and system historian. The architect must be able to
understand the way the systems work in the context of the problem and
work with the rest of the team members to craft a solution that accom-
modates the already existing system environment.
18.9.3 Domain Expert
Our choice of a problem to solve was based on the types of impacts and
costs associated with the problem. In order to understand how modifi-
cations to the process will change the impacts, it is critical to have an
expert from the business/analytical side to vet any proposed changes. It
is also the role of the domain expert to document the user requirements
and work with the rules engineer to transform those requirements into
data quahty and business rules.
18.9.4 Rules Software Engineer
The rules engineer will work with the domain expert to translate user
requirements into data quality and business rules. The rules engineer is
also tasked with managing and configuring any automated rules-
oriented processing, and this includes evaluation of tools, rule defini-
tion, and application integration. This engineer will also implement any
software needed for integration with the main system, as well as any
standalone applications.
18.9.5 Quality Assurance and Root Cause Analysis
Of course, no data quality team would be complete without a Q/A engi-

neer. This team member is tasked with both qualifying the implementa-
tion and performing the root cause analysis to pinpoint problem location.
18.10 STEP 10: BUILD YOUR ARSENAL
While the kinds of tools that exist for data quality were discussed in
Chapters 14 and 16, we did not specify or recommend that any particu-
lar product. The reason for this is that each problem is slightly different,
and different combinations of product capabilities may be needed.
On the other hand, it is likely that the project will need at least
some of these components for implementation.
Data cleansing
Data standardization
Database checking/validation
Rules definition system
Rules execution system
Approximate matching
Some of these products may be offered in bundled form a data
cleansing tool may have a rules manager bundled, for example. The
important thing is to recognize the kind of application functionality
that is needed and plan accordingly.
18.10.1 Build vs Buy
There are many fine products in the market that can make up part of a
data quality solution, but that doesn't mean that a solution can be inte-
grated every time. It is possible that much of the componentry needed to
configure a solution is available, but sometimes there are reasons why it
doesn't make sense to actually buy the product(s). Also, the cost of
these products can range from hundreds to hundreds of thousands of
dollars. For example, if a cleansing application needs to be run only
once before a new data validation system is installed, it may be more
cost effective to invest a small amount of money in a simple tool and
augment it with constructed applications.
18.10.2 Selecting Data Cleansing Applications
There are different kind of cleansing applications, ranging from name

and address cleansing to pattern analyzers and rule-oriented cleansing.
The selection of a cleansing application should be based on these factors.
The kind of data to be cleansed. As we have said, there are differ-
ent tools that act on different kinds of data. Some products are
domain based, which means that they have been optimized for
specific vertical applications, such as bank customer data or inter-
national addressing. Other products are very general but require a
lot of configuration.
The number of times the cleansing is to be performed. Our
premise in this book is that if we can maintain a high level of
quality of data that enters the enterprise, we will have lessened the
need for static cleansing. On the other hand, if we can determine

that it is less expensive to address the symptoms than fixing the
problem, cleansing may be more frequent.
18.10.3 Selecting Data Standardization Applications
There may be a question as to the need for a special data standardiza-

tion application. For the most part, standardization applications are
applicable only where there is a recognized standard, such as postal
addressing, or where the use of an EDI protocol is required.
18.10.4 Selecting a Conformance/Reconciliation System
As we have discussed a number of times in this book, hooking up a

database validation system will require some auxiliary system. This
incorporates funneling the results of the rule validation to a reconcilia-
tion system, as well as the reinjection of corrected information back
into the system.
Frequently, a workflow system is used for part of a reconciliation
system. A workflow application allows the user to specify the process
through which a nonconforming data record is examined and recon-
ciled. Workflow systems can range in cost quite widely as well, and sim-
ple workflow capability can be implemented using relatively traditional
technology (such as a database system and e-mails, for example).
18.10.5 Selecting a Rule Management Application
Our intention is to ensure that the definition of data quality and busi-
ness rules can be entrusted directly to the user. To do this, a rule editing
and management interface can be used. This interface should provide
an intuitive mechanism for nonexperts to define rules about the infor-
mation they care about. This kind of application can be integrated with
the rule execution system.
18.10.6 Selecting a Rule Execution Application
The rule execution system will take as input the rules defined via the
definition and management system and generates executable engines
that implement those rules. The rule execution system should conform
to the kinds of specifications discussed in Chapter 12.
18.10.7 Approximate Matching
In Chapters 14 and 16, we saw the need for approximate matching as a

component of a cleansing and enrichment program. Hopefully, the data
quality tools we evaluate will contain some kind of approximate match-
ing component, but if not, it will be necessary to incorporate an
approximate matching tool into the arsenal, too.
18.11 STEP 1 1 : METADATA MODEL
The next step to success is to define the metadata model if one does not
already exist. Use the guidelines described in Chapter 11 to help set this
up. Remember that we can incorporate enterprise reference data into
our metadata system, so it is worthwhile to consider the storage of
domains and mappings within this framework as well.
18.12 STEP 12: DEFINE DATA QUALITY RULES
We have done our impact analysis and our current state assessment,
both of which have fed the requirements definition process. With our
application components and metadata system in hand, we are now
ready to define our data quality and business rules.
18.12.1 Identifying Data Domains
Either through interviews or automated algorithms, it is time to identify

and register data domains. We have defined data domains in Chapter 7,
provided a rule framework in Chapter 8, and discussed the discovery of
domains in Chapter 13.
18.12.2 Identifying Mappings
As with domains, it is possible to identify mappings through expert

interviews or through automatic analysis of the data. The discussion of
mappings is also found in Chapters 7, 8, and 13.
18.12.3 Defining Rules
Once the domains and mappings are in place, we can begin to succes-
sively build our rule base by collecting the important assertions about
the expectations of data quality. Using the current state assessment and
requirements as input, we can target those areas that are most amenable
to rule definition. We use the rule editing and management application
(see Section 18.10.5) to define the rules.
18.13 STEP 13: ARCHAEOLOGY/DATA MINING
Without a comprehensive "archaeological analysis" of the data, we

may not be able to get the complete story of embedded data domains,
mappings, and rules. We can use data mining applications to explore
and discover this kind of embedded knowledge. We can use these data
mining algorithms to look for data domains, mappings, and especially
data quality rules that are embedded in data.
Clustering
Association rule discovery
Classification and regression trees
Decision trees
Link analysis
The use of these techniques may reveal that there are data quality and
business rules that are not consistently adhered to and, because of that,
simple discovery techniques will not uncover them. Once these kinds of
rules are discovered, they can be evaluated by the domain expert and, if
vetted, can be incorporated into the defined rule set.
18.14 STEP 14: MANAGE YOUR SUPPLIERS
We have a set of measurements based on the current state assessment

and the requirements analysis stage. For those requirements that are to
be imposed on external data suppUers, a suppUer management process
must be put in place. Again, with senior management support, the exe-
cution of a supplier management program begins with the specification
of minimum expected measures of data quality for data imported into
the enterprise.
As we described in Chapter 15, the supplier management program
specifies the rules that are being asserted about the expectations of the
data, along with a set of penalties for nonconformance. With our vali-
dation system to be implemented at the entry point for data, we should
be able to get hard measurements of conformance, which can be used in
turn to fuel the supplier management process.
18.15 STEP 15: EXECUTE THE IMPROVEMENT
Finally, we are at the execution stage. At this point, we are prepared to

actually determine how to incorporate the solution to the data quality
problem.
18.15.1 Architecting the Solution
The solution architecture is an augmentation of the current system

architecture. Armed with the information chain, we know how data
move through the system. When we focus on solving a particular prob-
lem, we isolate that section of the information chain and look at the
software architecture at that section.
Assuming that we are integrating some kind of data validation and
filtering system, we must determine the best place in the current system
architecture to integrate the validation engine and reconciliation sys-
tem. The validation engine will probably be placed along some set of
communication channels within the information chain at the point
where data is input into a processing stage. The reconciliation system
may be a separate system for each problem or the same centralized sys-
tem, augmented for each implemented use.
18.15.2 Performing Static Cleansing
If we are creating a data quality application, we will want to make sure

that we are not leaving any bad data lying around. To this end, we can
run a static data cleansing over the existing data to make sure we have a
clean start.
18.15.3 Integrating and Testing the Rules
The rules have been defined and reviewed. Now is the time to integrate
the rules into an executable system and build the environment to run
tests and validate the validator. A test environment should be created
that draws its input from the same source as the application will in pro-
duction, and that environment may be used for testing. This is the
opportunity to evaluate the validity of the rules in an operational con-
text so any necessary changes may be flagged and made before moving
to production.
This component will incorporate all vaUdation rules and all pre-
scriptive rules. If rules are being used to perform a data transformation,
the integration must take place at this time also.
18.15.4 Building the Nonconformance Resolution System
The nonconformance resolution system is the application built from the

reconciliation management system as the back end to the rule valida-
tion system. The governance of the reconciliation process is based on
the data ownership policy. Depending on the way the responsibility is
distributed according to both the policy and to the requirements analy-
sis stage, the nonconformance resolution system will generate workflow
tasks to those in the responsibility matrix.
18.15.5 Integrating the Rules System with the Production System
At the point where the level of acceptance is reached during the testing
phase described in Section 18.15.3, the decision to move into produc-
tion can be made. This means integrating the rules and nonconfor-
mance resolution system in with the original system architecture and

switching control into that system.
We should also be careful to integrate the production validation
system with the measurement methods used during the current state
assessment. Any place where we had a well-defined measurement
process and application in place should be tied back together with the
production rules engine.
18.16 STEP 16: MEASURE IMPROVEMENT
One of the most critical pieces of the data quality program is the ability
to demonstrate success at improving data quality. We already have a
baseline for measuring improvement the current state assessment. At
that point in the process, we have identified a subset of the particular
areas that critically impact the data quality within the enterprise, and
we have gathered measurements based on defined metrics.
When we integrate the rules system, we made sure to also integrate
these measurements. The reason is that we have successfully built a vali-
dation system to improve the data quality. This should be reflected in the
locations and metrics we chose for the current state assessment. In other
words, we can deliver a strict determination of measured improvement by
continuing to perform the same measurements from the current state
assessment. If we really have improved the data quality, we can document
it with real evidence, not just anecdotal stories about improvements.
We can use the methods of statistical process control, discussed in
Chapter 6, to document historical improvement. We can then make use
of that method to assign new thresholds for additional improvements by
resetting the upper and lower control limits based on user specification.
18.17 STEP 17: BUILD ON EACH SUCCESS
Finally, we have reached the point of success. We have measured,

assessed, analyzed, set requirements, architected and implemented a
system, and documented improvement. The success of any small project
only contributes to the greater success of the entire program. Therefore,
each small success should be used as leverage with the senior-level spon-
sors to gain access to bigger and better problems.
This is the time to address the sponsoring managers, armed with

the measurements of success, to declare war on the next set of data
quahty problems. As the enterprise knowledge is accumulated in the
central information repository, the entire system and the users thereof
should derive increased benefit. The ultimate victory is when a systemic
measurable improvement is achieved. But winning the war involves
winning small battles, and reaching a higher level of enterprise data
quality requires dedication, perseverance, and a true belief in the value
of enterprise data quality.
18.18 CONCLUSION
At the beginning of this book, we talked about data quality as a hazy

and ill-defined concept but one that can severely hamper a company's
ability to do business. Our goal has been to demonstrate that data qual-
ity is not an unclear concept but something that can be quantified, mea-
sured, and improved, all with a strict focus on return on investment.
The goal of becoming an organization that leverages its data
resource into a source of enterprise knowledge can be achieved but not
without a firm base of high-quality data. Hopefully, this data quality
approach to knowledge management will become the pedestal upon
which the pillar of corporate knowledge willfirmlystand.
INDEX
Abbreviation expansion, 352 example of, 417-418

Abstraction, 292 improving linkage with, 416-420
Accessibility, information policy, 122, n-granuning, 362-364
225-226 phonetic similarity, 359-362
Access of data, 31 rules, 418-419
Access rights, 265 similarity defined, 354-355
Accuracy, data values, 113,217 thresholding, 357
Acquisition overhead, 92 Approximate rules, 197
Actors, 235-236,240-241 Arithmetic operations, 146
Addresses Array, 258
abbreviations, 376 Assertions
business-to-business, 377-378 as business rules, 171
cleansing, 334 conditional, 170
delivery line, 375-376 definitions, 170
last line, 376 foreign key, 191-192,444-445
NCOA, 379 key, 190
parsing, 333 prescriptive, 170
recipient line, 374 primary key, 190-191,442-443
standardization, 334, 373-379 proscriptive, 170
zip codes, 376-377 types of, 169
Affinity programs, 409 Assessment process, review of, 235-238
Agglomerative clustering, 320, 323-324 Association rules, 330-331
Aggregate operations, 147-148 Attributes
Aggregation control chart, 132
approximate, 451 granularity, 340, 341
consohdation and, 371 overloaded and split, 314-316, 340,
enhancement through data, 414-416 343-344
hnkage and, 416 schema metadata, 261
metadata, 266 value restriction rules, 175-176
AUas, 257 Attrition
Application Development Trends, 281 costs, 89-90
Appropriateness, data presentation, 118,221 customer; 13, 76
Approximate aggregate, 451 Audit trails, 406
Approximate matching, 450-451
defined, 416-417 Binary types, 146
difference function and similarity, 355-356 Blockading costs, 90
edit distance, 357-359 Boundary conditions, 246-247
485
486 INDEX
Broadcasters, Web sites as, 415 householding, 367-370

Browser, metadata, 275-276 merge/purge, 367
Bureaucracy, data ownership and, 30-31 network analysis, 371-372
Business rules Constraints, 246-247
See also Data quality rules; Rule-based Contact information, metadata, 254-255,256
systems Context, enhancement through, 406
application of, 353 Control chart(s)
assertions as, 171 building a, 131-132
defined, 8-9,281-282 example of, 135-136
enhancement through, 422-423 interpreting, 138-139
knowledge management and, 15-17 role of, 128
management of, 32 types of, 132-135
"Business RulesAutomating Business Conversion operations, 148
Policies and Practices** (Kramer), 291-292 Cooperative marketing, 16,409
Corporate mergers, 407
CART algorithm, 328-329 Correction costs, 87-88
Cartesian coordinate system, 355 Correction report, 391-393
Catalog, 260 Corrections, making, 349-354
c chart, 134 Cost-effect of low data quality (COLDQ)
Centralized ownership, 38 adjusting for solution costs, 97
Change, responsiveness to, 292 dataflowsand information chains, 76-83
Character types, 145 evidence of economic impact, 74-76
Chief information officer (CIO), 41 example, 97-98
Chief knowledge officer; 41 impacts, 83-84, 85-93
City block distance, 322 measuring, 85
Clarity of definition, data model, 103,210 steps in, 73, 93-97
Clustering Costs, types of
agglomerative, 320, 323-324 acquisition overhead, 92
defined, 320-321 attrition, 89-90
distance and similarity, 321-322, 354-364 blockading, 90
divisive, 320, 322-323 correction, 87-88
generalized rule discovery and, 325-326 decay, 92-93
hierarchical versus nonhierarchical, 320 delay, 91
K means algorithm, 326 detection, 87
value ranges using, 324-325 idling, 91
Collections, 258-259 increased difficulty, 91
Competition, in knowledge-intensive infrastructure, 93
industries, 15-16 lost opportunities, 91
Completeness, data values, 114,218 misalignment, 92
Completeness relation rule, 186,439 organizational mistrust, 91-92
Composed domains, 154 preemption, 91
Composed mappings, 158 prevention, 88
Comprehensiveness, data model, 103-104, reduction, 89
210-211 rework, 88
Conditional assertions, 170 rollback, 88
Conditional operations, 147 spin, 89
Confidence, breakdown in, 14 warranty, 88-89
Connections, 260 Currency
Consistency consolidation and, 370
data presentation, 120-121,223-224 data values, 115-116,219
data values, 114-115,218-219 discovered domains, 313-314
relation rule, 187-188,440 Current state assessment (CSA)
Consolidation choosing locations in information chain,
aggregation and, 371 207-208
currency and, 370 choosing subset of dimensions, 208
defined, 364 identifying data customers, 204-205
INDEX 487
identifying sentinel rules, 208-209 essentialness, 105-106,212

mapping information chain, 205-207 flexibility, 104,211
review of, 237 homogeneity, 107-108,213
role of, 203-204 identifiability, 108-109,214
Customer! s) measuring data quality of, 210-217
acquisition, 13-14, 75 naturalness, 108. 214
attrition, 13, 76 obtainability, 109,214-215
identifying, 204-205 precision of domains, 107,213
interviews, 95 relevance, 110,215
record parsing, 333 robustness, 104-105,211
semantic consistency, 110-111,216
Data simplicity, 110,215-216
administrator, 42 structural consistency, 111-112,216-217
consumers, 27-28,42 Data owners, types of
custodian, 42 compilers, 34
extraction, transformation, and loading of, consumers, 34
60-61 creators, 33-34
flow, 79-80, 94 decoder, 35-36
migration, 59-60 enterprise, 35
mining, 63-64 funding organization, 35
processing stages, 27 global, 37
producers, 27 packager, 36
provider, 42 purchaser/licenser; 37
registrar, 41-42 reader; 36
steward, 42 subject, 36-37
trustee, 41 Data ownership
validation, 54-55 bureaucracy and, 30-31
Databases centralized, 38
data quality and, 55-57 data quality and, 32, 43^6
data quality requirements, 249 decentrahzed, 38
extraction, transformation, and loading of fear and, 30
data, 60-61 information factory, 26-28,451-454
legacy, 57-59 privacy and, 29
migration of data, 59-60 responsibilities of, 31-33,245
sources, 263 turf and, 29-30
Web sites as, 415 value and, 28-29
Data cleansing Data ownership policy
address standardization, 373-379 cataloging data sets, 41
approximate matching and similarity, creating a, 38-40
354-364 enforcing, 44-45
correction and enhancement, 349-354 identifying and assigning roles, 41-42
defined, 333 identifying stakeholders, 40
enhancement through, 409-410 in an improvement program, 51
metadata, 348-349 maintaining registry, 43
name and address, 333-334 Data presentation
record parsing, 344-348 appropriateness, 118,221
updating missing fields, 372-373 correct interpretation, 118-119,221
Data domains. See Domains flexibility, 120,222
Data enhancement/enrichment. See format precision, 120,222
Enhancement/enrichment formats and attributes, 117-118
Data-flow chains, 192-193 measuring data quality of, 221-225
Data models null values, 121,224
attribute granularity, 106-107,212 portability, 120,222-223
clarity of definition, 103,210 representation consistency, 120-121,
comprehensiveness, 103-104,210-211 223-224
defined, 102 storage, use of, 121,224-225
488 INDEX
Data quality value manipulation operators and functions,

See also Assertions; Dimensions 174,175
databases and, 55-61 Data types
data warehouse and, 61-63 base, 145-146
defined, 47-49 defined, 145
electronic data interchange and, 65-67 metadata, 256-259
horror stories, 2-7 sets versus, 148-149
improvement program, 49-53 Data values
Internet and, 67-71 accuracy, 113,217
knowledge management and, 7-10 completeness, 114,218
measuring, of data domains, 219-220 consistency, 114-115,218-219
measuring, of data models, 210-217 currency/timeliness, 115-116,219
measuring, of data presentation, 221-225 measuring data quality of, 217-219
measuring, of data values, 217-219 null values, 114,217-218
measuring, of information policy, 225-228 Data warehousing, 12
ownership and, 32,43-46 breakdown in analogy, 415-416
reasons for caring about, 10-15 certification example, 62-63
requirements, 245-250 results offlaweddata, 61
Data Quality for the Information Age validation, 455-456
(Redman), 101 Data Warehousing Institute, 12
Data quality practice, steps for creating Date operations, 148
building arsenal, 476-479 Datestamps, 146
building team, 475-476 Debugging
build on success, 483-484 the problem, 389-391
choosing a project, 474-^75 the process, 385-389
current state assessment, 472-473 Decay of information, costs of, 92-93
data mining/analysis, 480 Decentralized ownership, 38
data quality rules defined, 479-480 Decimal types, 146
data quality scorecard, 470-472 Decision making, 11
implementation, 481-483 Decision trees, 327-330
information chain mapping, 468-469 Defects, freedom from, 48-49
management support and ownership policy, Definitions, role of, 170
464-467 Delays, 91
measure improvement, 483 Dependence class, 459
metadata model, 479 Deregulation, affects of, 16-17
recognizing problem, 463-464 Derivation relation rule, 188, 420, 441
requirements assessment, 473-474 Description, metadata, 255
spreading the word, 467468 Descriptive domains, 151-152,177
supplier management, 481 Descriptive mappings, 158
Data quality rules, 45 Detection costs, 87
See also Rule-based systems Difference function and similarity, 355-356
compilation of, 198 Dimensions
domain mappings, 179-185 categories of, 101-102
domain membership, 177-179 data models, 102-112
enhancement through, 421-422 data presentation, 117-121
in-process, 192-195 data values, 113-116
management of, 197-198,270-275 domains, 116-117
naming, 428 importance of, 123-124
null value, 171-174 information policy, 122-123
operational, 195-196 Director of application development, 42
ordering of, 199-201 Director of informationflow,42
relation, 186-188 Director of production processing, 42
table, cross-table, and cross-message Distance and similarity, 321-322, 354-364,
assertions, 188-192 416-417
validation of, 198-199 edit, 357-359
value, 174-176 Divisive clustering, 320, 322-323
489
INDEX
Documents, data quality requirements, Employee interviews, 94

250 Employee turnover, low data quality and, 75
Domain discovery Enhancement/enrichment, 353-354
benefits of, 302-304 aggregation and linkage, 414-416
defined, 301, 302 approximate matching and, 416-420
expertise identification of, 305 business rules foi; 422-423
keeping domains current, 313-314 context and, 406
membership analysis by value matching, data quality rules foi; 421-422
308-309 defined, 399,400
overloaded and spHt attributes, 314-316 examples of, 399,400-403
pattern analysis identification of, 310-312 inference and, 420-421
representing data domains, 304 matching, merging, and record linkage,
set analysis identification of, 305-307 412-413
super- and subdomains, 312-313 merging of data and, 407-411
Domain mappings, 179 provenance and, 405-406
See also Mapping discovery; Mappings standardization ad, 404-405
assignment, 436-438 Enumerated domains, 150-151,177-178
enumeration, 181-183 Enumerated mappings, 156-158
functional, 180-181 Enumeration
managing, 269-270 domain mapping, 181-183
mapping assignment, 183-184 metadata, 257
mapping membership, 184,435-436 of readers, 266
mapping non-membership, 184-185,438 of writers, 267
mappings as functions, 185 Errors, common, 340-344
Domains Essentiahiess, data model, 105-106,212
assignment, 178,436-438 Euclidean distance, 322,355
composed, 154 Exact-match distance, 322
defined, 116 Exemption relation rule, 187,439-440
descriptive, 151-152,177 Expectations, determining, 238-240
descriptive power of, 149-150 Expertise identification, 305, 318
descriptive versus enumerated, 116 Extensions, 244-245
enumerated, 150-151,177-178 External data consumers, 28
example, 159-161 Extraaion, transformation, and loading (ETL)
knowledge management and cataloging, 154 of data, 60-61
managing, 268-269
measuring data quality of, 219-220 Feai; data ownership and, 30
membership, 179,432-433 Fingerflubs,340,342
membership rules, 177-179 Fimess for use, 48
as metadata, 163,256-259 Flexibihty
non-membership, 179,433-434,438 data model, 104,211
operations on, 152-153 data presentation, 120,222
as reference data, 161 Floating data, 340,343
stewardship, 117,220 Floating numbers, 146
ubiquity, 117,220 Foreign key assertions, 191-192,444-445
usage agreement, 116,219-220 Format conformance, strict, 340,341
Dynamic measurement, 229 Format precision, data presentation, 120,222
Formats, semistructured, 340,341
Economic framework. See Cost-effea of low Fraud detection, 411
data quality Function. See Mappings
Edit distance, 357-359 Functional dependency rule, 189-190,
Education, of staff, 45-46 441^42
80-20 rule, 128-129 Functional domain mapping, 180-181
Electronic Data Interchange (EDI) Fuzzy rules, 197
defined, 65
structure versus content, 67 Gartner Group, 8,10
XML, 65-67,456-457 Gottesdienei; Ellen, 281
490 INDEX
Granularity In-process rules, 192-195

attribute, 340, 341 Integration, 287, 292
data model attribute, 106-107,212 International Standard Organization (ISO),
rule-based system, 287 313
Greenspun, Phillip, 415 Internet
data queries, types of, 69-70
Handler, metadata, 256 role of, 17, 67-71
Hard impacts, 84 Web sites, 415
Help source, metadata, 256 Interpretation, data presentation and correct,
Homogeneity, data model, 107-108,213 118-119,221
Householding, 367-370, 410 Intrinsic, 257
Invariants, 246-247
Iconic representation, metadata, 255 Isolation, 292
Identifiability, data model, 108-109,214
Idling costs, 91 Jacobson, Ivar, 240
Impact(s), 83 Joins, 262
domains, 85-86,95
hard, 84 Key assertions, 190,442-445
operational, 86, 87-90 Key discovery, 326-327
review of, 236-237 Keywords, metadata, 255
soft, 84 K means algorithm, 326
strategic, 86,90-93 Knowledge discovery (data mining), 63-64
tactical, 86,90-93 Knowledge inferences, 353
Improvement program steps Knowledge management
current state assessment and requirements, business rules and, 15-17
52 cataloging domains and, 154
identifying needs of improvement, 52 data quality and, 7-10
implementation of plan, 53 defined, 7-8
ownership policy, 51 Kramer, Mitchell, 291
selecting project for improvement, 53
senior-level management endorsement, Load programs, 261-262
49-51 Location, metadata, 255
training, 51 Logical operators, 147
Indexes, 264 Maintenance, data, 32
Inefficiency, 10-11 Mapping discovery
Inference, enhancement through, 420-421 benefits of, 317
Information chains defined, 316-317
choose locations in, 207-208 expertise identification of, 318
creating, 207 one-to-one mappings, 318-319
examples of, 80-83 representing data domain mappings,
mapping, 77-78, 93-94, 205-207 317-318
processing stages, 205 Mappings, 154
representing the, 78-79 See also Domain mappings
review of, 236 characteristics of, 155-156
specification, 446-447 composed,158
Information channels, 77, 206 defined, 155
Information factory, 26-28, 451-454 descriptive, 158
Information policy enumerated,156-158
accessibility, 122, 225-226 as functions, 185
measuring data quality of, 225-228 as metadata, 163
meta data, 122, 226 operations on, 158-159
privacy, 122, 227 as reference data, 161
redundancy, 123,228 types of, 156
security, 123,227 Marketing
unit cost, 123,228 cooperative, 16,409
Infrastructure costs, 93 micro-,16
INDEX 491
Massachusetts Institute of Technology, Total Object-Oriented Software Engineering

Data Quality Management group at, 101 (Jacobson),240
Measurement directives, 193-194, 447-449 Object type mapping, 258
Measuring data quality Obtainability, 109,214-215
of data domains, 219-220 Open Information Model (OIM), 254
of data models, 210-217 Operational data consumers, 27-28
of data presentation, 221-225 Operational data flow, 80
of data values, 217-219 Operational impacts, 86, 87-90
of information policy, 225-228 Operational rules, 195-196
static versus dynamic, 229 Operations, 146-148
Medical diagnosis, 410 on domains, 152-153
Menus, metadata, 256 on mappings, 158-159
Merge/purge, 367 Opportunities, costs of lost, 91
Merging, enhancement through, 407-412 Overloaded attributes, 314-316, 340,
Metadata 343-344 .
browsing, 275-276 Overloading, 105, 107
cleansing, 348-349
defined,253 Packaging and delivery of data, 32
domains, managing, 268-269 Pareto Principle, 128-131
domains and data types, 256-259 Parser generators, 347
domains as, 163 Parsing, record, 344-348
as a driver of policy, 276 Pattern analysis identification, 310-312
generic elements, 254-256 Pattern matching, 346-348
information policy, 122, 226 p chart, 133
management of, 33 Percentage nonconforming, control chart, 133
mappings, managing, 269-270 Performance, 247-248
mappings as, 163 Philip and Alex's Guide to Web Publishing
parsing, 346 (Greenspun),415
rules, managing, 270-275 Phonetic similarity, 359-362
schema, 260-264 Policy manager, 41
usage statistics, keeping, 265-267 Portability, data presentation, 120, 222-223
Meta Data Coalition (MDC), 254 POSTNET, 378
Metaphone, 361-362 Precision, data model, 107,213
Micromarketing, 16 Preemption costs, 91
Migration projects, 14 Prescriptive assertions, 170
Misalignment costs, 92 Prevention costs, 88
Misfielded data, 340, 343 Primary key assertions, 190-191,442-443
Mistrust, costs of organizational, 91-92 Privacy
Model generation, 246 data ownership and, 29
information policy, 122, 227
Namespace, 257-258 reference data, 166
National Change of Address (NCOA), Probability assignation, 348
379 Processing chains, 192-193
Naturalness, data model, 108,214 Productivity drop, low data quality and, 74
Navigation rules, 197 Proscriptive assertions, 170
Network analysis, 371-372 Provenance, enhancement through, 405-406
N-gramming, 362-364
Non-null value rules, 174,431 Quality issues, 247
np chart, 133-134 Queries, 262
Null value rules, 171-174,430-431
Null values Rebalancing, 139
data presentation, 121,224 Record linkage, 412-413
data values, 114,217-218 Record parsing, 344-348
specification, 172-173,428-429 Redman, Thomas, 101
Numbering nonconforming (np chart), control Reduction costs, 89
chart, 133-134 Redundancy, information policy, 123, 228
492 INDEX
Refactoring data, 139 engines, 284,288-289,293-294,445

Reference data exemption relation rule, 187,439-440
agreement of usage, 165 firing/triggering, 287-288
domains and mappings as, 161 functional dependency, 189-190,441-442
privaqr, 166 importing, 300
publication of, 165 managing, 270-275
publish/subscribe model of, 163-167 languages, 283
stewardship and, 163-164 translation, 297
subscription-based distribution, 166 vahdation, 297,299
Referential integrity, 192
Relational inputs, 246 Scalars, 258
Relation rules, 186-188 Schema metadata, 260-264
Relations. See Mappings Security, 31
Relevance, data model, 110,215 information policy, 123,227
Report generation and queries, 247,266 Security administrator; 42
Return on investment (ROI), 14-15, 97 Semantic consistency, data model, 110-111,216
Rework costs, 88 Semistructured data, 414-415
Robustness, data model, 104-105,211 Semistructured formats, 340, 341
Rollback costs, 88 Sentinel rules, 208-209
Root cause analysis Service interruptions, low data quality and, 74
debugging the problem, 389-391 Sets
debugging the process, 385-389 analysis identification, 305-307
correction report, 391-393 data types versus, 148-149
defined, 382-384 descriptive power of, 149-150
determining what needs to be done to make Shewhart, Walter; 125,126,127
correaions, 391-394 Similarity, 321-322, 354-364
role of, 381 defined, 354-355
Rule-based systems phonetic, 359-362
advantages of, 284-286 Simplicity, data model, 110,215-216
backward chaining, 290-291 Soft impacts, 84
business rules defined, 281-282 Soundex, 360-361
data quality, 295-300 Spin costs, 89
data quality rules as business rules and vice Split attributes, 314-316
versa, 282-283 Stakeholders, identifying, 40
deduction versus goal-orientation, 289-291 Standards/standardization , 33
defined, 283-284 address, 373-379
evaluation of, 291-294 benefits of, 336-338
examples of, 280 cleansing and, 351-352
execution of, 287-289 defined, 335-336,338
forward chaining, 290 enhancement through, 404-405
granularity, 287,292 testing standard form, 338-339
integrating, 286-287,292 transforming standard form, 339-340
Hmitations of, 294-295 Statistical Process Control (SPC)
rule versus knowledge base, 286-287 control chart, building a, 131-132
scalability, 287 control chart, example of, 135-136
seleaion, 295-296 control chart, interpreting, 138-139
Rule discovery. See Clustering control chart, role of, 128
Rules control chart, types of, 132-135
See also Business rules; Data quality rules defined, 125-126
approximate matching, 418-419 development of, 125
attribute value restriction, 175-176 goal of, 137
association, 330-331 identifying special causes, 140
completeness relation rule, 186,439 maintaining control, 140
consistency relation rule, 187-188,440 Pareto Principle, 128-131
defined, 280 statistical control, defined, 127-128
derivation relation rule, 188,420,441 variation, causes of, 126-127
INDEX 493
Stewardship, reference data and, 163-164 Uniqueness rule, 443-444

Storage, data presentation, 121,224-225 Unit costs, 123,228
Stored procedures, 264 Unnatural patterns, control chart, 138
Straight-through processing (STP), 65 Updates, 267,450
Strategic data consumers, 38 rules, 196-197
Strategic dataflow,79 Updating missingfields,372-373
Strategic impacts, 86, 90-93 Usage statistics, 265-267
Static measurement, 229 Use case analysis
String operations, 147 actors, 240-241
String types, 145 defined, 241
Structural consistency, data model, 111-112, extensions, 244-245
216-217 identifying, 242
Structured Query Language (SQL), 425 triggers, 242-244
selea statement, 426-428 variants, 244
Subclassing, 108 User community, 31-32
Subdomains, 312-313 User interface (UI)
Superdomains, 312-313 creating, 460
Supplier management, 33 dependence-based data entry, 458-459
role of, 381, 394-396 dependence graph, building the, 459
Sveiby, Karl Erik, 8 selection criteria, 457-458
Tables, 260-261 Validation

Tactical data consumers, 38 data, 54-55
Tactical impacts, 86, 90-93 data quality rule, 198-199
Thresholding, 357 rule, 297,299
Time constraints, 265 Value
Timeliness data ownership and, 28-29
data values, 115-116,219 lookup, 348
discovered domains, 313-314 matching, 308-309
Time operations, 148 ranges using clustering, 324-325
Timestamps, 146 restrictions, 431-432
Token parser^ 346 rules, 174-176
Training Variable control chart, 132
of staff, 45-46 Variants, 244
topics covered by, 51 Variation, causes of, 126-127
Transactional data, 161 Views, 262
Transaction factory, 451-454 von Halle, Barbara, 281
Transcription errors, 340, 342
Transformationflubs,340, 342 Wang, Richard, 101
Transformation rules, 195-196,449 Warranty costs, 88-89
Transformations, schema metadata, 263 Whole number types, 145
Trigger directives, 195,449 World Wide Web. See Internet
Triggers, 242-244,264
Turf, data ownership and, 29-30 XML (Extensible Markup Language), 65-67,
456-457
chart, 134
Union, 258 Zone tests, 139
BIBLIOGRAPHY
Berry, Michael J. A. and Gordon Linoff. Data Mining Techniques. Wiley. 1997.
Berson, Alex and Stephen J. Smith. Data Warehousing, Data Mining, & OLAP. McGraw-Hill.
1997.
Bowman, Judith S., Sandra L. Emerson, and Marcy Darnovsky. The Practical SQL Handbook, 3d
ed. Addison Wesley. 1996.
Celko, Joe. Joe Celko's Data & Databases: Concepts in Practice. Morgan Kaufmann. 1999.
Fayyad, Usama M. et al., eds. Advances in Knowledge Discovery & Data Mining. MIT Press. 1996.
Frakes, William B. and Ricardo Baeza-Yates, eds. Information Retrieval Data Structures and Algo-
rithms. Prentice Hall. 1992.
Elmasri, Ramez and Shamkant B. Navathe. Fundamentals of Database Systems. Addison Wesley.
1994.
Enghsh, Larry P. Improving Data Warehouse and Business Information Quality. Wiley. 1999.
Huang, Kuan-Tsae, Yang W. Lee, and Richard Y. Wang. Quality Information and Knowledge.
Prentice HaU. 1999.
Juran, Joseph M. and A. Blantos Godfrey, eds. Juran's Quality Handbook, 5th ed. McGraw-Hill.
1999.
Redman, Thomas C. Data Quality for the Information Age. Artech House. 1996.
Sunderraman, Rajshekhar. Oracle Programming: A Primer. Addison Wesley. 1999.

Loshin Enterprise Knowledge Management

Uploaded by

Copyright:

Available Formats

Loshin Enterprise Knowledge Management

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Loshin Enterprise Knowledge Management

Uploaded by

Copyright:

Available Formats

What were some of the issues with the voting systems used in the 2000 US Presidential election?

What were some of the issues with the voting systems used in the 2000 US Presidential election?

What are some examples of how data presentation and validation issues led to voter confusion?

What are some examples of how data presentation and validation issues led to voter confusion?

PREFACE

While data quality problems are widespread, it is rare for an event to

POOR DATA REPRESENTATION

A poor decision with respect to data presentation resulted in voter con-

INVALID ANALYTICAL MODELS

For many people on election eve, it is customary to sit in front of their

CONFLICTING DATA SOURCES

By 2:00AM the VNS (and consequently, the reporting organizations)

Timeliness is an important aspect of information quality. We have come

DATA QUALITY ON THE PERSONAL LEVEL

My personal connection to data quality began at a very early age,

I have been collecting what I call "Data Quality Horror Stories"

Without even realizing it, everyone is affected by poor data quality.

Data domains and mappings between those domains

1.1 DATA QUALITY HORROR STORIES

1.1.1 Bank Deposit?

In November of 1998, it was reported by the Associated Press that a

1.1.2 CD Mail Fraud

David Russo, 33, of Sayerville, NJ, admitted yesterday that

1.1.3 Mars Orb/Cer

1.1.4 Credit Card Woes

1.1.5 Open or Closed Account?

Three months after canceling my cellular telephone service, I continue

1.1.6 Business Credit Card

A friend of mine is the president of a small home-based business. He

1.1.7 Direct Marketing

li':;"-,fW.^W^ :>' *'

Ytmr '^ QMI i gBWiiilllM

> Yon iiii# 4ift icitfl CQIIM^ lil 904

<%ifi. "-.-.v sr-jtNmcx*

Figure 1.1 Mysterious bill

1.1.8 Tracking Backward

I recently ordered some computer equipment, and I was given a track-

Date Time Location Activity

June 28, 2000 5:25 P.M. NASSAU-HICKSVILLE,

1.2 KNOWLEDGE MANAGEMENT AND DATA QUALITY

1.2.1 What Is Knowledge Management?

An organization's data collection is a valuable business resource that

overlords, the ability to share and benefit from enterprise knowledge

1.2.2 What Are Business Rules?

A business rule is an assertion about the state of a business process. All

1. http.7/cestec 1 .mty.itesm.mx-laava/sdsites/cursos/pqg_base/clefinicion 1 .htm

Business rules are likely to be expressed in a natural language. An

The opportunity to take advantage of the data and information re-

management. According to the Gartner Group, "It is critical for enter-

1.3 REASONS FOR CARING ABOUT DATA QUALITY

The data quality problem is pervasive in organizations across all indus-

1.3.1 Low Data Quality Leads to Operational Inefficiency

2. Explored later in this book.

record is found to be incorrect, the record needs to be deleted or fixed

1.3.2 Low-Quality Data Constrains Decision Making

Information production can be used for either operational processing

1.3.3 Good Data Enhances Data Warehouse Utility

According to an article presented by the Data Warehousing Institute,"^ a

1.3.4 Bad Data Leads to Incorrect Conclusions

Just as a house built on a weak foundation cannot stand, conclusions