UNITV

MACH INE
DESIer N AND ANALYSIS OF
LEARING1 ExPERIMENTS

MACHINE EAANG UFE CYCLE:

wachine leasning L) Wwodal mavagmat

vd the dolieng of highy performin model is ag
inportaut as ste inbiall boild af tha madol by
Checsing igh datosc
The Conapts asond modol retaining,
vmodel vrslorumg, medol daployment avd modol
ae the basia fos Wacuino Jaanina
monitong
operabiong tta holos tta data scence team
deliver models
Jeihy perbmaibg
The VSe of wacive lkaming tas ireaveL
Subeiontaly in enterpise dakeu awalysHa Scenaos

bo ertvact valvcble nsighs fom business

data
Hena it veny mporbant bo bae au
is
moclel,
elosyslem to build bhe maclel, boild
chase bat
Cempute performamce mebics and
perfomig wodel
The model waintenane plauys a Cntical ole
Once ohe mdal is deployed inbo producbin
 waintonno

ncwdes Wudol op to dao wl

thorve
Cavose of tino

Machne
poCes that Covons nht forn Shnco data
Dha
ldertfttotitn to ocol doolepranb, odal bploant
cd wodol wainteyamte. A.b
Cutire activitos fall tdon bw0 brood oteros
Mdal dondoprnant and ML modal
3uch as ML
bperotions .
has
ilke Mache 'ne Joanng iecp
fulnaing pdaos.
Podrness gual TdenbirfKabivo
ML Paotlorm framg
Data colo cti ondala pepoasig
feabune evqimaariog)
Madel doolop mant (Tracnimy, honimg, evaloahon
Medol doplogmant (Tnfesonee pautietion)
Mudol monionng
Busimaze qual
hane
An organizatim (oneidenig ML Shoobd
 Valie to be gaiet by Solbiog poblamy.
be abe to woasne bucimogS Vahe

qinst Speetfic bucimas Checties rd

Citeia
ML poo bem famig:
In this phase, tie bocinass
is fscmad podom
wachine daanng probem:
what Ls bhsend ad ohat shold be predicted
Cenoon labo oX
Delesm n
tenet Vanabla)
predict an
perormance and metna most be
optiizd is a key step in ttuis phae
DATA PRcESS IN G:

Training
acurabe ML wodel Tauine
data pooLsi3 to Convert data into a Ugabla
fusmat
Data poosig steps indoda Colleefion
prepasiag data and tahne aginaiog
the procoss cf oreing
ovd
tenshg, ertactirg
selectig vahablas fpfrarn
DEpU MENT r data
MODEL DEVSopMENT:
After a model is tso'ne, ned
evalwvoated and Valiolated we cau oeplry the
made Dhto the prodcbior. we can mabo predich os
and infeseress aqimat to modol .
MbDEL DEECopME N1 :
ouolopnads Consids modal
boudng, traming, toning awmd enlovat tn
Meolal building inclodas seotirng a
Pipelne tat aulomatog Tho build ,aim
taqimg Qud prodoeh tov
envom manig

Bueimass
gal
wonitoihg
fanig

Data
procossir9

Machise aarnng i uyee pooas

 MON) TDRING:
Model moni toring Sen
tor modal is waiwtanng a dosred

detecti Dn
pefraname gh ealy
igitatiou
Gu1DELNES MACHINE LEARNING ExPERIMENTS
AlM OF THE SUDY :
glOhat are bie cbjecbives (eg
asessing bru
erpected enos of an algostm an a partiolad
poobiers, ect)
SELECTTON OF TtE DegpoNSE VAR\ABE:
Ohat should e e
MeasUne (eg esor, preasi on ard
e,
qaliy
CHOI CE OF PACTO2s AND Comdexiy, ec
LEJELS:

ain f tha ShayC tocors

Gre hyperpormoers (hon
-Be algoith is fix and ant o fird bast
hyperpararoBes,if we ane compmng algoit my
Joasming algoitom is a tack)
Choio of eypeomenta deaign.
*Use factonal dasigo unoss Le
tthat
ttüa fockss do not inferact
Rplicati en omber deperds dolset
Si20 t n be ba pt Small ohan Tha
dataset lorge
* Avoid using Small dabnseks hich loadi
Vaiarea and 1
and
dferencos oil not be sgnifant
eguts not be Conclosive
pertrming fhe epenimant :
*Doings a feo tral ans
Jamdom settimg to check at d is
as orpected, bejose doing the fechosal
expenmant
Al Tulbs shool be
epsdiocblo
Statisticol Qyalysis a tie data : Conclasi
be due to choco
jet shold vot
Conclosion and ecDromandations

trmyauty Conclosi is
hoed fors futkes expe mon tation Thone
tiat Dur Concosion
be orong espectalley i f tio cata is Srall
and noisy ohen DUr expectabibns ane not mat, it is mest
helpol bo invstigae ohy they ane hot
DIASET PPE AON:
Machine aaring it. abot Jaasrig
Some
popesti of data sot ard aplyng
nous dola.
(ommon 50ctico In rrachira
a
to esnlivato n algont ig to 3pt
dara at
raiming set (ohich ue Jaam data ooere
and Cne all a
tet Those
tætig
Casing
poperties
set, bn ohich

YIh tha troiniang data, data ane

gsian The Jabals. TTh tos data, data label
Unknoon but not iven The tranimg
dataset Con iss of 'ing eXAmpes
ylhe Yeal aim of Sopenv ise arnig
on tost dala Thab iis mot

bnon dosing loartning

valuog or the
po0rnotes at minini2e the las mcha
on ha
trauiming dola ie mot recussarily
the best poli
 The hainkg
the to°nèmg Sam ple
lho tat erospY ig tha expected predichtn
im clopendont tet Sanplo
is

mot
*pobem
a god
trcng
estimaoY for tha tat

* raimnng e Can be Tedeo by mabi

tho hypothesis mone ens itie
Seng itive to tscu'mumg dale
and poo
bot this
enerali2ation
oad tior
Fting
Set of Rxamples used
Tioriming set '
whore
A
the target Value is

Test Set:
to asegs

pertnmanes of a cagsifie
Mever used duing bhe troining
tost s e
pooase
Poruideg Unbaige estimae of tha
ererali atin eoret tha euooadge aboot
# ls
Traiming datach ue ue to Cnct cassitir
1he daba Sovce V ohi
 dotaset a haimng set is inplo rno nted la
boild p a mudol Lohile a test
les (or validotie) o
is Vaidate ta modol buil
x Data pons in tha rauming seto} o
exolocad fum tha tast ( Val'datim) set
a

set, a
runng Set, Valdatory Set (Sonme pocpla bse
ingtead) in each derotibn, 0r dividd into
in.
Set ,a Val'daiy Set and a tegt st
traung
ach iteration.
Macina daasne'
baialy iy
to prediet tie test data
to create mode)
So wa Tha traunubg duta fit The moda!
avnd teatng ata to Gest it .
The modes generate e to preict
the reullo Unknon oeh s amed as Tho
test cet
CROsS VALIDAT1ON CCU) AND RESAMPLI NGt
gValidation
Uged
bechqe in wachine zamng
o yet the eonr rate of the M ML
model Which Can be onsideved as Qose to the
 yate popoati cn
*I he cale Voloma Jore enn'ga
be epagentive of tha populaion yu nay
hot hoa the Validation Aechnia
Machine loaming, wodal hlidatio
olemad Jbo bhe protDS ohore
traimed wodal
dabcusot
Ls Qvalovoded Ubh a
tatig
*The tas tig data seb a separate
portion of the data Set frm aiCh
the Set is desved
the
lhe vain porpose of osig
data et to tet dha narali zatin
iy of a toaulmed
abily odel.

Data set
Traming ostig Holdot mtfkod
Cooss Vadation

Data pesmitng

Teatiro
 Cross Valudati on a techuqe. for eosvoir
ML
Sobeeks
mads
hrainng
data
Sereal ML

Use css - ldotien to doteck Ontthrg

gnovalie patterm.
In general ML Drol2a dasvong vwodols fronn
daba, wrth the cçm of acheivtng some 2ird of
dasned bohauian ea,pAicton or coscifiaton
But <bhis uerie tosk s broken dacn into
omeyr
of peeal cases. when trraining is dene
that data was nambe an be sed bo. tast the
performamce ol tha laanad mdal on
This tie basic jdaa fo a oole coas
of model ealivatieu athods callad cos
Validatbien

*Typs of CoDss Validatioy mbhods ana hold t.

k-fold and loave- Dvo -oot .
The holdot mathod is the sëmplt kind af
no wo
CoOSs Vaidabion . Tha data cet iS Sepamted
ses, agd tte troining set and he testing dako
Set
*The
fncbieu appoimate tits a facbisu Usiny
the' haindng set only
 K-ld way bo impsne
over dhe haldat waticd.
dabaseb is divldad into k Sobsets
and the holdot mattuod is epeated k timno,
* Ech time of tie k Subcets is us ag tha
test selt and tthe othar k-l Suhseks a put togttar

to fosm
t Than he
trials is Computeol.
Leae - one -out C0es Validaion is k-fold
Validation abon to i4s ogil ex feseme
oitn k eqel te nombor
o N ta ombor of
of data poins
Set
WWoans thak NSeparate imes
data
the mction apox wak ie tsouned n he
execpt tor oe foint ad predichiY is wacd or
avd a pedi'cibny
Ttat point
k. fold Chos Validation:
k- fold cv is where
given cola set
Split ito a k womber of
Sectiors folds
ohere each old is used as testing et at Som
poBnt
,Lets toba the Scanansics o s-fold Choss
lalidabior (K=s) Hene tha daloset 1s Splat Into
5 o ds
tha first inesa cbion the firs+ fold is
Used to teet tie modol and bta Test osed Jo trau
the modal
*Dh ta Seeond
the
iteabhion- 2ud d is used
tattng set Ohde me Yost Sesve as ta
trauming Set This proas As
each fold of tte s fulds Dapeate
as been ased
unbil
the testmg dta sat
k -fold \hlidation, is petormed
per tho follaoing stepe
partition the ogimal huning dota et into
K equal Sobsets,Fach Sobgzt is alled aa fold
plds be naned as A,fe.. k.

gkoep tthe fold fi os Validatior Sot and

Keap al remaining k-i folds in bhe (oS
Validabion trou'mng set.
#Estiate the acuay af tor machihe
Jeaming modd by wragig the qcera cles
deied n l tha kases of CousSS Valldatio
ohe k- fold Ooss Valdabion mebhod,all
origionad taunig daka set
Dsd fors bottu trantng ay oel as alidai on.
g Also each enhy is used fo Valiaion vst
Once

Fod 4
lods
od s
t Tiau'ming sel
he acvontage of
watrs Jass aw the this-mathod,
daa qet
is aab the
dividad .
ve data polng gete to be
onto. and gek to be in a in tegt a

Sek
eracty
k - times
iaming
|he Varionce of tne
js educa
The disodvautage of 1e wotnod is ttat
tte raummg wothod
mcthed hasdro be cm Surabch
9ich maans it talas k timos ag mwoh
Compojebien o .
4

A vardaut b t s wthod is to
divide te
h daln nto aa tes) and traing
Set K
dueeek
hrag times
Tndependanty checse hew
Jang eoch
set is and do wany tri als Yat oorca
OVeY

Boolshapping:
It is a nethod off ample sese tat
is wch mae
heral. "an cos aidatio
ha idaa is bo use 1o obseed
Sanple to astimate tfa populat un dishbuhon
amplas Cau be drauon fono tha
*Then
csti mated popolatior ad Tha Sampling
dishibotioy o ay ype o estimato an itsel be
stimoded.
boctstrap is ,faxibe and
* The
paur Statistial tool. fat an be used to quautty
ayeroaked wlth afiven estimato
or Satisial lsarng wmetfod
 *Por example iH ca an ostt mate
of t e Stendad e r of a Coeffiaent oY
Cofidene inleal or tcat Coelhcien+ .

ppose at olsh to Invag t a fixd

Sum of menay
into JOo mara'a oests
thal
yeled ehne of x and y raepecthi valy
ohene Xend y ave uaetitieg
we wil ovest a fnetion, a afracton
mon inin X and oi invegt t
remaining
wish bo oose to minçmize the
Jtotal isk, o Varian, o! Ünes tnenb
¥ n obier ooTds, ue wwan Jo rinimize
Var (ax + (l -a) Y)
Ohe Can show thab te Valwe thab
winimizos is bay,

where
Vor),oyVar (y) and oy= Crv)
 But tho Valcs a o,0y ad dxy ao onkwon
Car Cormpole. osnatQs )

Aata set
*y vsmg daba
ant thak Contaim,
measUomonts br X and
Can hen
minimbo ur Inenont using
dhe aonta of ouY

To esttmate Oho etndard daviati on o oo ue

epaatad the pO@AS of Simolatio (oo pairad cbseovationg
of xand y esthi matimg d looo fines

DbBainod ipoo esimae for d

ohich
We
thasebyCall di, d2,... o00
FoT the simolatins the
pamles wc

he Vale of d is o6 .
tne
The wane Over all ,o0O egti natog fos d is

rel

Coge to
doviatio of the esti mateg
 lopo

|000-)

oO83

Vay good idaa of ta

gies
Cumacy f : SEC&)0-o83
o a rondn
Sample fom th Spoabing
pöpütabin, USOold er pect
d approsimatd 0-08

There trso fns ofaf bodstvoping

ohich dife primanly in dhs the population
estimated

NONPARAMETRe B0otS TRAP :

Jhe won parramehic booshrap a Sample

of tha
the. Sana ci2 the data
as s takon Bor
fom
the datba wa Teplocemant
measune 10 Samples, we
Create ha mple by replicatig
Some o tietee Samplas ttat We haue

Sesn and omittig othos alrady

 SEMIPARAMETRIC BootTRAP: (9

the
The
Jem
resaunpling bolatmp Can
that
y The Seniipaametic
the bcotstap agsumas ta
population imcdg otfer temg
Similas ko oha
a Smooted
dsenad Sampla by Soupling fom
version of tte omole h'shogo
Senple
be done.
Vany siapy by frct tabig a Soumpla toith
eplacement fro the Obseed ample ad than
a ddimg noise .

PARAMETPIC Root ST2AD:

paTametic bocsapping assUmes thatte
data Cones fro buouon distribufion rt
UuenUOn paTametes
estimate the parameterg fm
tha daba Jae aud tan'
tthe es tiwated
'shibution to mlae the
Soumplas
MEASURIN
CLASSIFIERPERAURMANCE
*A binaa assifiahion a
ttat agsins ls aa mtad
dgs do a bho bas iS
is
dosursphion
 The pevbosmana o a bina clasitiey an
be cssessed
by abuating is
tost Set uoith knoon
coit mOLn abols in
br Confsj sn mattx torth ochal Cas93. i Conbingeny
Tows and precte classes in Colomra
MeasoTOs pestornanca hacd to etishy
Severa Ctena!
must
Cohavertiy ophao The
aspect of performarta of interest
most be intotie enogh
becomo toidaly sed,
ome ae. tonickenty reproted by
wide
Vogearchors, Quablimg Cowwoinibg
nabling
Concosù ons o be doauon:
moct be Computabievally
bractoe, bo mabch he apid qrouth in Scal
wodom dala Coocisu.
fol
hay.massbeSumple to epot as
womber
Single homber kor each mobthd - daaet

Combuabiou
perfoswemco metics fo binay Classifi cati oy
dsignad to (astnad tadoefts bohaan fou
fdametal populabieu quauttie True
false positives, toe vagpti a oud fabe posi tie
*The evalatbi en
vaqpkives
measunes in docifiation
potems dainad Bom a matix witt tha
a
wmbey
Cosecty
for each class, vaed Couhsi
aud inceNoctty Cbsifed
wotnx
The Confvgion watix fr a binay
Closcifialior oldom is Shoon belos
Trve ass prediced cbac
egatie postive
False napbive Tre positive
Negative False positiue

A Cosien matix conbains abaut cal

poadited lassifications dtne by a

classificati System
pesomamea f och yste is
data Matx
eon fusion watx iss also calod ag
 (
false posiive
Ohich ame
Examples prediced as positive,
dbho class
Fase
vagtie
vagatie :
Exomple8 pedice as hagabie,olose
toe Cos is positi ve
Tieje positves:
Exoms Cassectty psedlced as pertaing
to the positive Cass

Tre hagaties:
Exonples ase coseety.pirediced as
belonings bo ta hegptie class -
measne wat
wost Usedin
tn racHco.
he eralvabi on
is the acunag rate
Acunaey ate Tne nogabie tt Te
Te positive

false vasatis +. Fale pocitive +Taue

\hegabie + e
ACweAcy AND ROc CURVES : psitie

Bimay classifi atbion acura

y
mem'e qpanty tthe two ype f. Comect
predietiong Cud Hpes of eoos.
 23
mees are acwtacy AcC), preesi ohon
Aypial
Tecal feee postive ate, f- wmeane:
* bach mene a diffesont
The predicatie mode!.
Ospect

(Meagnas the fractin Df

Comect predicHons
stua Psaction af
acha positieg among these examglas Ttot ase
predietocd ás positie
Recall moagUnes
positiveg psedited as positie
f-meagUTe haUnic
wean af precision ana /teca.

Operating haractenstia (Roc)

Rocaier
Gapha hae dong baon Used in Sigmal delector
tobo dapic ttha hoadecfe botwean
dit rate aurd fale aym fate8 DVer
chanal
pacent yeas haue eau Cun lhreele
in the
boc qraphe n tte wachie
 An ppc
plots dve pusitive rate
yaxis false positive ate oh X-axis

Simyle Cou tingeng table Correspords

Simge point 0m an poc plot
The porfromance o a Rano an be
Cussessed
|n a drauoig
plot, e a n as
a plecauoise imear Cme
an

Tthe Cune Stark in (o, ) finishas in

C,) and is wonitoriatly dorraasi
bota ares

Roc Curve the e positive

rate Ds
plo tied
in fmcbieu o
felsoPositive
rate Cl00 Speeitiiy for dfferent tot-off
poins of a parameter

Fach point Cuve

eprasents a
Senaitivihy Speuhciny paiy cenegendy
a particular deu'silon threshdd .
The are

a weasne of ho uel a paramaer

dishngulsh belucn thcse
Segmants
*An ROC Cue s ConveX if the lopas Cnd

mont
olmgtmualy hon - ushen

A
wo o Concauiadjacant
wona y Sogmens oith
Cvave
increasim
than
Slopes, tnd'ates Loally worse
andau raneing
90ald bete
raubing pertoace by Jonlng
Seaments imvoLued in oha Con Cawihy
Tus
Creabtng a loarsey elassifeN
PRE ceION AND RECALL:
Relevamce Subjecti e hotion
Diffenent way dtfer about bhe
TeleraMee non- Yolavenca of porttealaf
docomeuk dho qen quasti imns
Yesponse o
IR Sygem Searchos is doesment Colechien and
ems Odorod esponses
Is Called Tei Set of
 26
A better eanch geilds a better raukod
and better ranbed sts hep tha gerr
imhomatbBoy raod

This Set oP eords jm tha

database. oiich holevant bhe Search
topic
GsSUmed to be eithet
Recorcs
olevat or imelevant
ay not
achal Yetnval get
sel o volovan
match dote
perkety
reords.
Set

Set o
Yelevant
ifemsio
the dateba

\nelonl
items
reiewed

-s rolaOm ems- reved

Relewamt Tlems- ot vehve
 Recall
(21
is bha ratio
nomer of
velevont Yeeords Yatived to tohe total
womber elevam the
dalabase
s vsoaly
percenta
A A = wmber of
o ralevo
Teds eied
B Nmber of reluvan
Recai A.
A
* (00. reords not
A tB Teeved
Preesin Yabioof dota nmber
of rolewat YeLords heeved tthe total
wmber I'elsvent and elevan reords
rehived . expressed

pestenta A: wmber of
relevat
Yeods ratneved
C- Nmbey of i
YO LoYds roeevElyawe
 (

preslon = A
A+c

As Yecall pracision deoma,

and daes tRa peclsi n
F- Measme

ts a measna d et
aturay and
i's daj'med eighbad hamone
mean Yeaal
tta test

The P- aasne F- Soe

mest Singa hombe"

mease in Tofonation ehieval, Naal argage
poocssimg Qnd Maawie laamworg
4-meagune Comes rom Bnformabi on
Ratival CiR) Lohare Pecale is

ohich relavant dooments ara ehiee

recalled
by bot i+ S
by Tn
posi tie
eleeleve
Rade CTpe)
Senaitiity
 is tha
Y Precision
etieved docomens fraaueny
porecdichiong ase
Whi ch
ralsvant
Comeet and s a fon of
also kuown positie predicti ve
Vase Cppv) oT pasithe Aoraey (Tea)
0s lmtended Combine these i nto a
singe meousne af
of Search epecsvenes
2 pNeeisjon x ocall
prousion + Peealu
MOUTICASS CLASSIFICATION
is Machie laonming
tosk ttat
casifiakon.
Consists of more than huoo
Classes otputs
Each
of N deffenen
hraiming poimt belongs
belongs to Una
Closses
The goal is
Laich, given a
fnchion
data point, oil Conecty
pradict class Jho tha
point
kalanygs
botMean
gtned Averagi
aNe Yoe poe LisiDo (MAp) ts also
alled
peusim at Seen ralevent docomank
Dt daermne preu'sion at each paih wben
neoo relevemds doomant yt
retrieed.
Avera of preusion Valo cbtalnay
docoment each bino a relwent
cscyment is ehieued.
MAP |
N

ohee
6 NUmber of elavont dolomOnt
for
NeOmber aquenes
P(aea
doe mans
MOutCLASS UASSIFICATION TECHNIQUES:
truig
aiffenent clos point balogs
 The gol Cenatvet a fachi
which guen a he dala polnt oi

Cosecty prodicjed ta class o coh'ch

whichh
tto neo point balorgs
The oli -clogs clousifiaim pvoblan
Yefers t assign'mg ch f tee obsevations

0nto ona o k classes.

A Connmon combine pair
otse is
Compoisim Jy vohmg
Conshek a rule
beheon nan

Gaechmg
oo -
closs uoit te most
loss.deeisions r

\oing otedime
Degutnas Just peiy sise daesiong, it
predes class dabel.
A xXX

Binary casah clasciioin

 One is Al (OvA)

*For ttu's appoach wwe Taqina

cassifiers, where k th
N2 k
bimaytraloned oito posihve examples
Caseifer ls
and vagatia

ekomple belngig to te offer k-| clases

Ohen an Unknun Qxampe

the classiAes prodsng te awim
naxio
Consldeed ta Loinner and
Dotput s
tthis class (abd is assinad to tat

example
Eror Comechg Doptpat Cedomg
yEoros Coechng code appro aches
o Conmbime bimamy cossi Aers tm
way that You exploit de
Corelation and Commeet
E- Test

When Smalt Sample si zo 130) 1s

Considelod the tests a Ye InappliCa ble
because the aSS umptipnS We Nacle Por
Sample tests, oo ot hold goocd Por
lavg
Small Samples.
Small Samples It ao not Possi bla

ko aBre,

i) hat tie random Sampling dsti bubicn

Statistics normal
a

Sample Values are syfielenty case

r) The
bo Popation values bo Caleula be Te
estimale.

Thus entisely
to dea! with pob lemo of Small Sanples. Bat
Ohe Ghold note that tae methods and teorg
q Sma ll Samples applicoble bo lange
Barnples but it Coveise

when
Sample &ine Sma |), as to ften
ease n paetie, tu Conttal linit
(leorem loes not
not appy ne must ton
èmpuse 8\icter aMunptcons on teu poplatioy
to quwe latstial valdik to te kestprocacha
asunpbio
Fáom bwich Tlu Sample b balken has
normal Aobabiiby dostibution bo bogis wits
* Dogee fheodom (di): ey degue freedom
we mean numbr classes bo whih
he value Can be
nignod asb taily Dr
or af
withot Asthidios Or
Ainiations plasd.
For ercanple ashe to
whose todal 5o. cleorly
we at feelonn
e
Mubes eSoy lo, 23, g but tu fourt5
rmter, fireol ine ttu botal
6o To- C1o +23+ )
7)i o . Thus'we e quen
a estkiction, hare bue qreedom asclaylcon

sf paeesbnm (dy) 1s denotad by

dagrea Nn-ki
~Cnu) r of and it jo given by
where mbe 4 lases and k:nmber
G indpenolent Constsains.
Mean
t- Test por gingle lome tfiom a
when te Gample Values eart distubetios
nor mal distwbtion
was worled ut by W. Gose tt
called JE a
One t- ohibutan
Lngortcnotely ee es not
litaibutions gor lauh
Ttere are feent t- tere a cwtain
I n:
diyunt Value the t- dshibution
t- distbuticn but it n: 43
Say trat te Vavable
Va
aliHle dit fuent we
t- s hibuton wt n
* Sappose
Siple random anpla
n rs drawn FAOm Population. if te epulaton
Paom which
tnple ks takan Pollow8
normal elsibution
random Vaiable

Folows tudeot's t. Disthibudion with n- ( deges

Pheclom
The Sample ean and the Sanple ghandard
dovicutien &.
umber Ree
choices lyt efter a ample Rtatutiu Suh as
is calcalated when you wse a b- lsiibasion so
Cstmabe a pppulation een,
Rqual bo onu los than
df. = n

I::Population s normal attougo thà alumpton

Can be relae
a. Random
St int erest.
Sample wo drauon Faom tuu Poplation
Based Lompari Son ealeulabecd Value
tuito tine thohe tial ' Vaue Prom he table,
we conclucle:
 &hape Steclen t's t stibudion

Prepertis Students t- Distaibudiorn lffent

difeent Por
4. vhe t- olistibudion s
deqtees fheeotom.
rte t- utibuton Cented at o and
2
dyrnmetnis about O.
3. The tolal lnler te
te leyt onel tie ea bo tae rogh t

magnibede ineeases t e gtaph

A3 te
Cpproae but hever

area tue
de taiks a tre t- lisibution is
tn tee tal[ te norma)
largor fan te area

utribution.
tte t- dstibuteon N olkpendent
6 qte 3hape
lnereales tiee cshibton
AS Sampe Siza
becomes Opprorimately nomal.
Tee Stand ard lewiaton } reater
than .
8
tre t- drstribufion
tails
Je
a hf greator than the area in the 6ails Of
Tuo Standayol norha listibution, betaue we

}ndsodueng further
an
estimate ot
Vaniabilety
q. e men, medlian , and mode fue t- ostaibution
eual to
 populor is ho) p
o-efhcient Comelabin The
lation! Coe foy Test
Woma)
960 |:
|:962
O42 2
213|
251)
o25 totreodg lagnee
af
n
wonmal) te
ed dishobuon
Qre fofrocdan
r
to t a
dogsecs
f ous Vard too Vales Critical y
Valtag Citca) T-
numbeY&.
lore Lawg tue by claser
bo
get
values increOog
Sixe
n Sampe tte as
urve. olensity ma
l no
te toloser get luve
y teireasos In Sample tu'
 33)
Shonat h of tto volahnship
desenbi
beean wo Vanad
CoYrolation
The Co-efdent tho clopo of
behemn wo Vana Ha
Yeopeslon
Vamabos aue bean andardiod
when bote
t a r moang aud aividing
Ssbhihgstandasod moang Or doation
\oy bolwaen
Comolation
rovg
mias one
plus
Casinptive
tohenSpecial dishhbutional assUmptiong
Shatisic, ho aboot t e Voaniates
wada
to be
whi ch
it is Calculated

Comelatien oefaens
fest fo
Rrmola
t r

with dagness
State moll and alemate hypotesis
Ho
Ha p +0
 Hee p Is tta popuation Comelatior
Lo-oftiedent
Stote tho Significanto davel.
the fest eatisttes of Corobhn
to-efHelent oith
aboe - dahod
above
brma
do's ion Use
Vale approach or tha p- Vale approach
Fnalty Stete tfe Conlgion
Me emar's Test:

It is a hon- paDamene tost

paired nomial data
Used for fmdimg a

chang ppportov for paired data

Compare tha
losciflers Dn N ideny om Sirgle tes
Sel

This testto Comp On

the porfomanca of bo caacihers
Lame test set
 This test
4
WomeY ohich A and B
wabe reehons

Me NOmarlg test applid to

taldes wtth ratchad pairs
conbigeny
Subjecs tro
datcm'a Lo haastta
warqial traayenu
tha and cdom
are oal.
malin ascUmptiong o test
The tfnoe

hominal aato
most
Juoo Oud one indopendent
ategones aud
toitt oo ConnoCted oops
Variabio

Tha grups im the dapendonb

Vatiabla be wohaly dosie

be sando Semple
k-flod Cv pained t Test:
we Use k- fold Validation to get
K
Tratnng Validoion set pais
 To traln the to classifi Cahsw

algoni ralming sots Ti;

K and test

Vali'dation Sek V;

Tha percantagu of
Vali'dation Gets
clossiers (On

oordod as p and p?
buo obgcifiahisn alguitims
daue Some eor rate, Hhon we expect
Jhoue Same
ttthem o
eruvalently,te differenca of tfair maansis
Tates
in
,The difperera
ld i as Ri =P-P tu's a paied
test
that is or each both aloithns
bhe Sone
Validaim Ses

k
(Pi-m)
K -)
 *Undar coha hyoóllais ttok o,
have a Shatist t tat is 4- di'sibute
oith

Thos tte K-fold

veyecls
hyoteate ttat hwo Coasihcaton
lgonthns hoe ttha Sona emoy rate at
Biniicance toel
tta interval C-t K'ja K-)
we want o ast whaatter te
first aleosth Jhas dos ttan tha
Se cond maed a sidad hipofkasis
test :
Ho
the tegt
rejace, bur ca'
first
tal Ono bas
signhiant
Less is Sogprted
 ttat eah tegtstS
Advamteg
independon+
But bha
traeng set
oerlap
Tha
obtaimig a tood shat of to
moont of aiati on thab cooUld be
bbsene each tsang Set weno,

Completey imdoperelons f pnawia trau'mirg

tete

Vanana in
the t statisti ray
Somattas Uhalosastimaded, te maan he
eshimated
regu in t Valos
large