Data mining: a database perspective.

M. S. Sousa, M. L. Q. Mattoso & N. F. F. Ebecken

COPPE, Federal University of Rio de Janeiro
P.O. Box 68511, Rio de Janeiro, RJ, Brazil, 21945-970
EMail: mauros, marta @cos.ufrj.br, & [email protected]

Abstract
Data mining on large databases has been a major concern in research com-
munity, due to the diculty of analyzing huge volumes of data using only
traditional OLAP tools. This sort of process implies a lot of computa-
tional power, memory and disk I/O, which can only be provided by parallel
computers. We present a discussion of how database technology can be
integrated to data mining techniques. Finally, we also point out several ad-
vantages of addressing data consuming activities through a tight integration
of a parallel database server and data mining techniques.

1 Introduction
Data mining techniques have increasingly been studied7;9;21, espe-
cially in their application in real-world databases. One typical prob-
lem is that databases tend to be very large, and these techniques
often repeatedly scan the entire set. Sampling has been used for a
long time, but subtle dierences among sets of objects become less
evident.
This work provides an overview of some important data mining
techniques and their applicability on large databases. We also spot
several advantages of using a database management system (DBMS)
to manage and process information instead of conventional at les.
This approach has been a major concern of several researches, be-
cause it represents a very natural solution since DBMSs have been
successfully used in business management and currently may store
valuable hidden knowledge.
 One requirement of data mining is eciency and scalability of
mining algorithms. It makes the use of parallelism even more relevant
to provide a way of processing long running tasks in a timely manner.
In this context, parallel database systems come to play an important
role, because they can oer, among other advantages, transparent
and painless implementation of parallelism to process large data sets.
It is important to notice that, when we mention the use of large
amounts of information in data mining, we are not referring to usual
large DBMSs, which can reach more than one terabyte of data. As
data mining methods often repeatedly scan the data set, mining in
such a large database is not cited in the literature yet.
The remainder of this work is organized as follows. Section 2
presents some important mining techniques currently implemented
in data mining systems. Section 3 describes how these techniques
can be applied to both at les and DBMSs, enforcing advantages of
the latter approach, and Section 4 presents important advantages to
be analyzed when considering the use of parallel databases to mine
knowledge. Section 5 describes a case study and an implementation
using a well-known classier algorithm with a tightly-coupled inte-
gration with a database system. Section 6 points out some related
data mining systems, whereas Section 7 presents our conclusions and
nal observations.

2 Data Mining Techniques

Data mining is a step in knowledge discovery in databases (KDD)
that searches for a series of hidden patterns in data, often involving
a repeated iterative application of particular data mining methods.
The goal of the whole KDD process is to make patterns understand-
able to humans in order to facilitate a better interpretation of the
underlying data11.
We present four classes of data mining techniques typically used
in a variety of well-known applications and researches currently cited
in the database mining community. They certainly do not represent
all mining methods, but are a considerable portion of them when a
large amount of data is considered.
2.1 Classication
Classication is a well-known data mining operation and it has been
studied in machine learning community for a long time8;18;19;24. Its
aim is to classify cases into dierent classes, based on common prop-
erties (attributes) among a set of objects in a database. After the
construction of the classication model, it is used to predict classes
of new cases that are going to be inserted in the database. Adequate
applications for classication include medical diagnosis, credit risk
assessment, fraud detection and target marketing.
Among classication algorithms, there are two methods that are
widely used for data mining purposes: decision trees and neural net-
works.

2.1.1 Decision Trees

Decision tree methods are a kind of machine learning algorithm that

uses a divide-and-conquer approach to classify cases using a tree-
based representation8;24. They usually use a greedy algorithm that
recursively subdivides the training set until reaching a partition that
represents cases totally belonging to the same class or until a criteria
is reached (pre-pruning). When deciding what attribute is going to be
used by each subdivision, a statistical test is adopted as the splitting
criteria.
After the growing phase, we may have an overspecialized tree
that overts data, providing more structure than necessary. Pruning
comes to play an important role in producing smaller trees with better
accuracy when considering new cases24.
Even after pruning, trees can represent a complex, dicult to
understand structure. Rule extraction is the nal phase used to get
smaller and less complex rules with similar accuracy. A greedy cov-
ering algorithm is often used to select a minimal subset of rules that
covers the examples24 .
The main problem with trees is that they need extensive data
to uncover complex structures. On the other hand, they can be con-
structed considerably faster than other machine learning algorithms
producing results with similar accuracy21 .
There are many implementations using decision tree learning al-
gorithms, such as CART8 , ID3 and its successor C4.524, and SPRINT
(SLIQ's parallel implementation)25;19 .
2.1.2 Neural Networks

Methods based on articial neural networks7 provide a general and

practical method for learning functions, which are represented by
continuous attributes, discrete or vectors. One important character-
istic of the algorithm is its robustness when dealing with errors in
the training set. Basically, neural networks have been used to in-
terpret visual scenes, voice recognition, and they are not only used
for classication (e.g. neural networks are widely used for prediction
purposes).
Neural networks typically require more time to nish than, say,
decision trees algorithms21. Training times vary depending on the
number of training cases, the number of weights in the network, and
the settings of many learning algorithm parameters. Finally the abil-
ity to understand learned target function is not important. Learned
neural networks are less easily communicated to humans than learned
rules.

2.2 Association Rules

Mining association rules2;6 is particularly important when trying to
nd relevant associations among items in a given customer transac-
tion. An example of the output of such mining is the statement that
80% of transactions that purchase diapers and milk also purchase
milk bottles. The number 80% is the rule condence factor.
A formal depiction of the problem is presented in Agrawal et al2 .
Let I = fi1; i2; i3; ::; img be a set of binary attributes called items and
let T be a database of transactions. Each transaction t is represented
by a binary vector, with t[k] = 1 if t bought the item ik , and t[k] = 0
otherwise. It is interesting to notice that we are not, at this moment,
worried about quantities, that is, how many items were bought in
each transaction.
Now let X be a set of items, such that X is a subset of I . We may
say that a transaction t satises X if, for all items in X , t[k] = 1.
An association rule is an implication of the form T X ) Y , where

X is a subset of I , Y is a subset of I , and X Y = ;. If c% of

transactions in D that contain X also contain Y , thus c% represents

the rule condence factor. The rule X ) Y has a support s% if s%
of transactions in D contain X ) Y .
The problem of mining association rules can be decomposed into
two parts:
 Generate all combinations of items that a have a transaction
support above an informed minimum threshold (large item sets).
This phase is considered the most time consuming task of the
algorithm.
 For all large item sets, generate association rules for the trans-
action database, which is done in a straightforward manner.
2.3 Clustering
Clustering algorithms3, also called unsupervised classication, is the
process of grouping physical or abstract objects into classes of similar
objects. Clustering analysis helps to construct meaningful partition-
ings of a large set of objects based on a divide and conquer methodol-
ogy, which decomposes a large-scale system into smaller components
to simplify design and implementation.
2.4 Sequential Patterns
The discovery of sequential patterns28 has been motivated by appli-
cations in retailing industry, including attached mailing and add-on
sales, and in the medical domain, for example.
The input data is typically a list of sequential transactions and
there is often a transaction-time associated with each of them. The
primary goal is to nd all sequential patterns with a pre-dened mini-
mum support, that is, the percentage of data sequences that contains
the pattern is greater than a given threshold.
Thus, sequential pattern algorithms are useful for discovering
trends in data, such as:
The number of sports magazines sold to customers with credit
between 20,000 and 30,000 living in city 2 is increasing.

3 Using Large Databases

One important characteristic of the data mining process is that it can
use data that has been gathered during many years and transform
it in valuable knowledge. The larger the database, the better and
the more accurate the knowledge. An important observation is that,
when we refer to the use of large amounts of information in data
mining, we are not referring to usual large DBMSs, which can reach
terabytes. As data mining methods often repeatedly scan the data
set, mining in such a large database is far from being a reality.
Data mining applications often work with two kinds of data
sources, and each one has its own advantages: at les or a database
management system.

3.1 Flat Files

Flat les have been largely used as the data source of information used
by data mining algorithms. Even when operational data is stored in
a DBMS, many systems extract a portion of them to be used during
the mining process, working only with data that ts in main memory,
which limits the total amount of information to be used. Implemen-
tations using at les oer some advantages related to performance
results.

3.2 Data Mining and DBMSs

Database technology has been successfully used in traditional busi-
ness data processing. Companies have been gathering a large amount
of data, using a DBMS system to manage it. Therefore, it is desirable
that we have an easy and painless use of database technology within
other areas, such as data mining.
DBMS technology oers many features that make it valuable
when implementing data mining applications. For example, it is pos-
sible to work with data sets that are considerably larger than main
memory, since the database itself is responsible for handling informa-
tion, paging and swapping when necessary. Besides, a simplied data
management and a closer integration to other systems are available
(e.g. data may be updated or managed as a part of a larger oper-
ational process). Moreover, as emerging object-relational databases
are providing the ability to handle image, video and voice, there is
a potential area to exploit mining of complex data types. Finally,
after rules are discovered, we can use ad-hoc and OLAP queries to
validate discovered patterns in an easy way. We must not forget
that information used during mining processing is often condential.
Thus, DBMSs can also be used as a means of providing data security,
which is widely implemented in commercial databases, avoiding the
need of using encryption algorithms to process information.
3.2.1 Loosely-coupled Integration

Even when using databases, most of the current data mining ap-
plications have a loose connection with them. They treat database
simply as a container from which data is extracted directly to the
main memory of the computer responsible for running the data min-
ing algorithm, just before the main execution begins. This approach
limits the amount of data that can be used, forcing applications to
lter information, and use only a part of it to discover patterns.
Alternatively, some applications dynamically perform queries to the
database, but work in a client / server architecture. It means that,
depending on the amount of data being transferred, unnecessary
network trac is generated. Moreover, they are often written in
programming languages that do not have any integration with the
database system.

3.2.2 Tightly-coupled

The idea of executing user-dened computation within databases,

thus avoiding unnecessary network trac, has been a major concern
in many DBMS systems. One example of this sort of implementation
is the stored procedure in commercial databases. For instance, Oracle
provides the possibility of implementing procedural code in PL/SQL
and storing it within the data dictionary, in a pre-compiled form
(PL/SQL is an interpreted language).
Some authors have developed a methodology for tightly-coupled
integration of data mining applications with database systems5 , se-
lectively pushing parts of the application program into them. They
have performed some practical experiments with a relational database
system and propose that, whenever possible, computation should be
included within SQL statements.

4 Data Mining and Parallel DBMSs

Given that it is desirable that a data mining process handles a large
volume of data, parallel algorithms are needed to provide scalability,
in order to end the process in a timely manner. Classication and
discovery of association rules are two data mining techniques that
have been extensively studied for parallelization purposes4;6;17;18;26.
It can be explained by several reasons, such as the importance of
these methods for commercial use, and their scalability.
There are a variety of data mining algorithms constructed to run
in parallel, taking advantage of parallel architectures using specic
programming routines. Alternatively, parallel database systems can
provide parallelization in a transparent, painless manner to the ap-
plication. There are many advantages of using parallel DBMSs:
 The implementation becomes simpler. There is no need to use
parallel routines such as MPI libraries. The parallel DBMS is
responsible itself for parallelizing queries that are issued against
it. We have to structure our queries so that they can fully
exploit parallelism oered by a particular DBMS.
 DBMS can choose between serial and parallel execution plans
transparently to application. Depending on the data mining al-
gorithm and technique used, it is clear that parallelism is more
useful in the beginning and intermediate phases of the pro-
cess. In the construction of decision trees, for example, the
training set is recursively partitioned and processed. Thus, at
some point of the processing, parallelism may not be useful
because the communication overhead between coordinator and
slave processes dominates the processing time, because there are
not many training examples being searched. Based on statis-
tics about data, the DBMS can evaluate the cost of parallel
and serial queries and decide which one is more advantageous,
depending on available machine resources.
 Opportunity for database fragmentation.
Database fragmentation20;22 provides many advantages related
to the reduced number of page accesses necessary for reading
data. Irrelevant data could be just ignored depending on the
lter specied by SQL statements. Furthermore, some DBMSs
can process each partition in parallel, using dierent execution
plans when applicable23 . In general, DBMS systems provide
automatic management of such functionality.
However, some drawbacks have to be considered, such as:
 Less control of how parallelization will occur. Although there
is the possibility of some customizations, such as setting the
 nodes that will participate in the process, the DBMS itself is in
charge for parallelization.
 The algorithm may become dependent on some characteristic of
the parallel DBMS used. In a parallel database system, we have
some chance to direct the optimizer to use a specic execution
plan, but our algorithm may be dependent on some character-
istic of this DBMS.
 Overhead of the database system kernel. A kernel of a database
system is designed to handle a large set of operations, such as
OLTP transactions, OLAP queries, etc. Although it is possible
to minimize the overhead and customize the environment to
take the best available advantages of the corresponding archi-
tecture, there will always exist some functionality implemented
that is not applicable to the data mining algorithm, which can
degrade performance when compared to access to at les.

5 A Classication Case Study

To study and identify problems related to database mining on a large
amount of data, we have implemented a classication algorithm (de-
cision tree) with a tightly-coupled integration to a parallel database
server. We have constructed a set of stored procedures that are stored
within Oracle Server data dictionary in a pre-compiled form. We have
been using a multiprocessor IBM RS/6000 SP2 system with Oracle
Parallel Server 7.3 installed14;23 .
We have decided to base our implementation on decision tree
algorithms because they are well-suited classiers for data mining
problems, producing similar accuracy when compared to other meth-
ods for classication. Besides, they represent one of the most widely
used and practical methods for inductive inference, and they can
be built considerably faster than other machine learning methods.
It is an important characteristic when focusing on handling a large
amount of data. Moreover, decision trees can be easily constructed
from and transformed into SQL statements1, which can be used to
query a parallel database system.
Another important consideration about our implementation is
that it does not have any memory dependency, that is, it will wok
no matter how large the data set is. If data does not t in memory,
the DBMS itself is responsible for the paging and swapping when
necessary. We do not propose any new data mining algorithm, but
a dierent approach of implementing a data mining system. Our
work is based on C4.524 and SPRINT25 , adapting them to use a
database system instead of at les. A more detailed discussion of
database technology features appears in [26], while the description
of the implemented decision tree algorithm in Oracle Parallel Server
may be found in [27].

5.1 Problem Statement

The data set used during the classication process is based on a
life insurance system, which controls information about customers,
insurance contracts and components of insurance taris. People in
management positions use this database in a data-warehousing en-
vironment in order to help decision making. Information about cus-
tomers and their dependents includes year of birth, job, sex, marital
status, and household income data. Besides, the system holds data
about contracts, such as the year when contract begins and ends,
modus of payment, name of the agent responsible for the contract,
and type of insurance. Finally, the database keeps information about
price of each tari that is associated to an insurance contract.
Although a series of normalized tables were available, we have
constructed a single relation because of performance problems re-
lated to join operations. Oracle oers a variety of methods to per-
form joins in parallel, namely nested loops, hash joins and star joins.
However, even when performed in parallel, joins represent a time con-
suming task. This single relation is a relatively small one that ts
in memory, with about 50Mb, 200.000 tuples and 70 attributes. For
data mining purposes, this relation represents a considerable large
one when compared to others that appear in current literature.
We handle attributes with dierent characteristics, namely: dis-
crete attributes with very few distinct values, discrete attributes with
a reasonable number of distinct values, continuous attributes with
only some distinct attributes and continuous attributes with a lot of
distinct values.
5.2 Typical Queries in Rule Induction Algorithms
Most of queries issued in decision tree algorithms, or in rule induc-
tion ones in general, deals with two kinds of operations: grouping
(GROUP BY) and sorts (ORDER BY), when working with continu-
ous attributes12. These kinds of queries represent the time consuming
tasks obtained during the implemented algorithm.
Select marital status, class, count(*)
From mine relation
Group By marital status, class
Parallel database servers can easily process this sort of query,
and it is possible to take advantage of parallel machine resources. In
Massively Parallel Processing architectures (MPP), for example, each
node can independently compute counts on the records stored locally.
In the next phase, a coordinator process can consolidate these counts,
and then return the result back to the user20 .
5.3 Tightly-coupled Implementation
Due to the importance of database technology, we have focused on
implementing most of classication algorithm within database. This
approach oers a tightly-coupled integration with the database sys-
tem and its query language and also avoids network trac between
a client and the server. All processing is done in the server and in-
termediate and nal results are also stored within it in the form of
tables. The client only issues a command (a stored procedure call)
and the DBMS is responsible for extracting results. Moreover, it
decreases the number of calls to the database, makes our code con-
siderably simpler, and pushes memory consideration problems into
the database system.
Our implementation involves the use of embedded routines writ-
ten PL/SQL (Oracle native procedural language), which are, when-
ever possible, called within SQL statements.
5.3.1 Tree Growing Phase

The following query presents an example of SQL statement used dur-

ing the growing phase of a decision tree, and it is used to calculate
the best splitting point of an internal node using gini index statistical
test8;25. The inner SQL statement groups the training set based on
attribute values. The GROUP BY clause is used instead of a simple
ORDER BY to count the number of occurrences for each pair at-
tribute/class, since we cannot split cases having the same attribute
value. Hence, the splitting criteria is computed faster than before,
using only one database call per attribute in each phase. When the
attribute has too many distinct values, a discretization would be ad-
visable.
Select min(giniSplit(salary, class, num))
From (Select salary, class, count(*) num
From mine relation
Group By salary, class)

5.3.2 Tree Pruning Phase

During the pruning step, a user-dened function is created based on

the original tree. It is used to classify test cases in parallel.
Select class, Classify(marital status, year contract end, salary)
From mine relation test
We have randomly divided the data set into two parts: one for
the tree growing phase, and the other for tree pruning. This task
was carefully performed to generate relations with the same class
distribution. Test data is represented by mine relation test, which is
used to predict error rates associated with each tree node.
In the above SQL statement, the rst column returns actual case
classication, while Classify returns the classication generated by
the algorithm. Function Classify is used to nd out the number of
hits and misses during classication, which will guide tree pruning.
5.3.3 Rule Extraction Phase

Once the classication model is constructed, a set of rules can be

derived reading the tree, from top to bottom, until reaching each leaf
node. The n rules originally dened (where n represents the number
of leaf nodes) are generalized, and some of them may be simplied
or even eliminated. Rule generalization process is a time consuming
task, since we examine many possibilities in order to eliminate irrel-
evant expressions in rules, producing a simpler subset. This process
issues many SQL statements, in which WHERE conditions repre-
sent rules discovered in previous phases. These statements, each one
processed in parallel, return the number of false positives and false
negatives examples that are going to be used in conjunction with a
well-known method called Minimum Description Length (MDL), as
it is done in C4.524.
As in the pruning case, this rule set is transformed into a PL/SQL
function, which can be called from within a SQL statement, enabling
the classication of new cases in parallel. This function receives the
attributes used during the data mining algorithm and returns the
predicted classication.

5.4 Performance Considerations

Depending on the database size and the complexity of applications,
the data mining process may represent a long running task. In our
case study, a tree with over 38.000 nodes was generated, what shows
that it is not a trivial problem to be solved. Most of parallelization
occurs when handling decision nodes that are on the top of tree, so
that more data is going to be analyzed. For these phases, we noticed a
signicant performance speedup when additional processors are used.
Although ad-hoc le processing algorithms often outperform the
ones based on DBMSs, there are many advantages of implementing
the latter approach. In our study, code simplicity proved to be one
of the most valuable advantages to be considered.
Performance improvements could be achieved through data frag-
mentation techniques. One characteristic of SQL that can degrade
the performance of a Rule Induction algorithm is that its structure re-
quires a separate query to construct the histogram for each attribute.
Thus, it is not possible to perform a single SQL statement that can
bring us the result reading data just once. Fragmentation techniques
tend to eliminate irrelevant information during query execution, min-
imizing such problem. Vertical fragmentation can be used to break a
big relation into smaller pieces, whereas horizontal partitioning func-
tions can be carefully dened in order to diminish the amount of
information that needs to be scanned in each decision node.
All implemented modules do not rely on any specic Oracle
Server feature. They use general functionality, which is currently
available in most of commercial parallel database systems. More-
over, this implementation can be easily used by another application,
since the constructed stored procedures / functions are just ready to
be used in a parameterized, exible form.
6 Overview of Data Mining Systems
In this section we present some currently implemented data mining
systems and some other interesting works in the area. They all have
in common the fact that they were implemented taking large amounts
of data into account.
6.1 Quest
The Quest data mining system4 has been addressing the development
of fast, scalable algorithms. Implemented algorithms include associ-
ation rules, generalization, sequential patterns, time-series clustering
and classication. Quest algorithms have been parallelized to run
on IBM's shared-nothing multiprocessor SP2. They have been work-
ing with parallel implementation of the mining of association rules
(APRIORI6 algorithm) and the SPRINT25 classication algorithm -
an evolution of SLIQ19 - where multiple processors are able to work
together to construct a classication model, namely a decision tree.
These algorithms run against data in at les as well as DB2 family
of database products, but databases are accessed in a loosely-couple
mode using dynamic SQL.
6.2 DBMiner
DBMiner15;16 integrates data mining with relational database sys-
tems and provides a SQL-like knowledge discovery interface for re-
lational operation-based data generalization, and thus facilitates its
integration with the existing commercial relational database systems.
It has been designed with emphasis on simplicity and extensibility.
Its modules include multiple-level characterization and association,
discovery of discriminant rules, classication, and prediction.
6.3 SKICAT
The SKICAT system10 (Sky Image Cataloging Analysis Tool) was
implemented to process images resulting from the Second Palomar
Observatory Sky Survey (POSS-II). It uses a variety of classica-
tion supervised learning techniques to automating the reduction and
analysis of a large astronomical data set, integrating methods for im-
age processing, data classication and database management. Thus,
the purpose of SKICAT is to enable and maximize the extraction
of meaningful information from such a large database in an ecient
and timely manner. Some learning algorithms (GID3*, O-BTree, and
RULER) are used to produce decision trees and classication rules
from training data consisting of astronomer-classied sky objects.
6.4 Other Works
Freitas and Lavington12;13 have presented a series of primitives for
rule induction (RI), after analyzing many KDD algorithms of the RI
paradigm. Core operations of their candidate rule evaluation proce-
dure are depicted considering performance characteristics and paral-
lelization. Primitives can be translated to usual GROUP BY / OR-
DER BY clauses in SQL statements, which often provide a potential
degree of parallelism in parallel DBMSs.
Holshemeir et al17 implemented a two-level architecture for con-
structing decision trees. A DBMS server (Monet Database Server)
was used in the experiments, and Quinlan's ID3 algorithm was used
as a baseline. Results are presented using a case study with 100k
cases (2.6Mb compressed). In parallel execution, a 25k relation was
used with 4 nodes.

7 Conclusions
Data mining and its application on large databases have been ex-
tensively studied due to the increasing diculty of analyzing large
volumes of data using only OLAP tools. This diculty pointed out
the need of an automated process to discover interesting and hidden
patterns in real-world data sets. The ability to handle large amounts
of information has been a major concern in many recent data mining
applications. Parallel processing comes to play an important role in
this context, once only parallel machines can provide sucient com-
putational power, memory and disk I/O.
We described some important data mining techniques, present-
ing brief descriptions about them and showing how each one can con-
tribute to the pattern discovery process. Furthermore, we presented
several advantages of implementing a data mining method using a
DBMS instead of conventional at les. Our practical work exploited
many specic characteristics of DBMSs, providing a tightly-coupled
integration of a data mining technique with a parallel database server
using a complex application. We have exercised adverse situations
such as large number of attributes, discrete and continuous attributes
with many distinct values, observing problems and solutions during
the whole process.
Experimental results have shown performance bottlenecks when
using a DBMS when compared to at les, due to the nature of
current SQL oered by most commercial databases. However, if we
assume that data mining of large databases is going to become in
the future a routine task and DBMS vendors are going to implement
new features that could help the data mining process, then database
servers will play an important role in this context. It is possible
that changes to current DBMSs and SQL language could enable data
mining operations to be performed more eciently. Besides, through
emerging object-relational technology, a potential area must be ex-
ploited. Other interfaces, such as those for integrating with indexing
and optimization mechanisms will be available in a near future, which
can oer a means of interfering in the parallel optimization process.
References
[1] Agrawal, R., Ghosh, S., Imelinski, T., Iyer, B., & Swami, A.,
An Interval classier for database mining applications, Proc. of
VLDB Conference, Vancouver, Canada, pp. 560-573, 1992.
[2] Agrawal, R., Imielinski, T., & Swami, A., Mining association rules
between sets of items in large databases, Proc. of Int. Conf. ACM
SIGMOD, Washington D. C. pp. 207-216, 1993.
[3] Agrawal, R., Gehrke, J., Gunopulos, & D., Raghavan, P., Auto-
matic Subspace Clustering of High Dimensional Data for Data
Mining Applications, Proc. of the ACM SIGMOD Int. Conf. on
Management of Data, Seattle, Washington, 1998.
[4] Agrawal, R., Metha, M., Shafer, J., & Srikant, R., The Quest
Data Mining System, Proc. of the 2nd Int. Conf. on Knowledge
Discovery in Databases and Data Mining, Portland, Oregon, 1996.
[5] Agrawal, R., & Shim, K., Developing Tightly-Coupled Data Min-
ing Applications on a Relational Database System, Proc. of 2nd
Int. Conf. on Knowledge Discovery in Databases and Data Min-
ing, Portland, Oregon, 1996.
[6] Agrawal, R., & Srikant, R., Fast Algorithms for mining associa-
tion rules, Proc. of the 20th VLDB Int. Conf., Santiago, Chile,
1994.
[7] Bigus, J. P., Data Mining with NeuralNetworks, McGraw-Hill,
1996.
[8] Breiman, L., Friedman, J., Olshen, R., & Stone, C., Classication
and Regression Trees, Wadsworth International Group, 1984.
[9] Chen, M. S., Han, J., & Yu, P. S., Data Mining: An Overview
from Database Perspective, IEEE Trans. on Knowledge and Data
Engineering, Vol. 8, No. 6, pp. 866-883, 1996
[10] Fayyad, U, Djorgovski, S., & Weir, N.. Automating the Anal-
ysis and Cataloging of Sky Surveys. In Advances in Knowledge
Discovery and Data Mining, pp. 471-493, AAAI Press, 1996.
[11] Fayyad, U., Piatesky-Shapriro, G., & Smyth, P., From Data
Mining to Knowledge Discovery: An Overview, In Advances in
Knowledge Discovery and Data Mining, pp. 1-34, AAAI Press,
1996.
[12] Freitas, A., Generic, Set-oriented Primitives to Support Data-
parallel Knowledge Discovery in Relational Database Systems,
Phd diss., 1997,
http://cswww.essex.ac.uk/SystemsArchitecture/DataMining
/alex/thesis.html.
[13] Freitas, A., & Lavington, S. H., Mining Very Large Databases
With Parallel Processing, Kluwer Academic Publishers, 1998.
[14] Hallmark, G., Oracle Parallel Warehouse Server, Proc. of ICDE,
pp. 314-320, 1997.
[15] Han, J., Fu, Y., Koperski, K., Melli, G., Wang, W., & Zaane,
O., Knowledge Mining in Databases: An Integration of Machine
Learning Methodologies with Database Technologies, Canadian
AI Magazine, 1995.
[16] Han, J., Fu, Y., Wang, W., Chiang, J., Gong, W., Koperski, K.,
Li, D., Lu, Y., Rajan, A., Stefanovic, N., Xia, B., & Zaiane, O.,
DBMiner: A System for Mining Knowledge in Large Relational
Databases. Proc. Int. Conf. on KDD, Portland, Oregon, 1996.
[17] Holsheimer, M., Kersten, & M. L., Siebes, A., Data Surveyor:
Searching the Nuggets in Parallel. In Advances in Knowledge Dis-
covery and Data Mining, pp. 447-467, AAAI Press, 1996.
[18] Kufrin, R., Decision Trees on Parallel Processors, Proc. of the
IJCAI Workshop on Parallel Processing for Articial Intelligence,
pp. 87-95, 1995.
[19] Metha, M., Agrawal, R., & Rissanen, J., SLIQ: A Fast Scalable
Classier for Data Mining, Proc. of the 5th Int'l Conference on
Extending Database Technology (EDBT), Avignon, France, 1996.
[20] Metha, M., & DeWitt, D.J., Data Placement in shared-nothing
parallel database systems. VLDB Journal, Springer-Verlag 1997.
[21] Mitchell, T. M., Machine Learning. McGraw-Hill, 1997.
[22] Navathe, S. B., & Ra, M., Vertical Partitioning for Database
Design: A Graphical Algorithm. Proc. of SIGMOD Int. Conf.,
pp. 440-450, 1989.
[23] Oracle Corporation. Oracle Parallel Server Concepts & Admin-
istration Rel. 7.3. Oracle Technical Manual.
[24] Quinlan, J., C4.5: Programs for Machine Learning. Morgan
Kaufman, 1993.
[25] Shafer, J., Agrawal, R., & Metha, M., SPRINT: A Scalable Par-
allel Classier for Data Mining, Proc. of the 22th Int. Conf. on
VLDB, Mumbi, India, 1996.
[26] Sousa, M. S., Mattoso, M.L.Q., Ebecken, & N.F.F., Data Min-
ing: A Tightly-Coupled Implementation using a Parallel Database
Server. Proc. Int. Conf. on DEXA Workshop Parallel Databases:
innovative applications and new architecture, IEEE CS, Viena,
Austria, 1998.
[27] Sousa, M. S., Mattoso, M.L.Q., Ebecken, & N.F.F., Data Mining
on Parallel Database Systems. Proc. Int. Conf. on PDPTA: Spe-
cial Session on Parallel Data Warehousing, CSREA Press, Las
Vegas, 1998.
[28] Srikant, R., & Agrawal, R., Mining Sequential Patterns: Gen-
eralizations and Performance Improvements, Proc. of the 5th
Int. Conf. on Extending Database Technology (EDBT), Avignon,
France, 1996.