0% found this document useful (0 votes)
25 viewsMLE Chp3 Notes
Uploaded by
Yash SolankiCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
25 viewsMLE Chp3 Notes
Uploaded by
Yash SolankiCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 40
. cu’ 3 |
Teaining Nedel gox Supewvised Leow in”d _ —
— Ded cut Nethodi- : —
{Tt _In Hus methed a past ef tee
——Input olata is held back (that 1s hew He __
name holdout oyiginares) ror evoluattey
af tue model. - ee oo
—= {This subset ay que input deto fs kp
; t used Qs te +est dara fos evoluattn9
Hae pergomance ay @ Model.
-—_— | In Qenera} FO - 80%. ap dus
= data dabetied ) is usec! fox model training
a. | mt is
is Ha ~mest—extveme case ay & fol)
—{CV. T+ uses _ ™m wecovsl ov ingtance ar @ a
Hime as Text data
~ | Tuis is don 4 maxim ire
to baine +. model.
{Le 4 COMP UIBA Da ly
very —PAPENSVe "rence —
Mot used 1 Proetice . _
fT
thet Count af Dede — ) BeotSteop Sampling . 7
_—-dt__Uses tut concept _a¢ SRSWR( simple __
fandom Sampling wrt. Replacement -
- Unlike +t _k-pelel CV, boetsteoppin'
“pices data instances from the Input de
carey ce bility — a+ so me data instance
fa cael mans that -fem Hua _input
pe
Date instences Lina wepeot
moudiple times ;
- This +ech nique Is pesticula ry uses) jn cae
of maj! Sfza, f-e having very ees number
= af Decta instances.
4
D lol 234567 @4~~4 a
soo eo eee oe
B34 s 30 Te
al
Here , Dotdset Dis Set from oleh
A_subseis ap Data instances are creates _
—Dy,D2 Da tot Repeating same Instance
-muttple times (her Training ser)
= for_testing set D-D, is farmed i
—_| D-Da,
— | D-Dy oe Vs Eages Leamexs —
4
a enn
> Fagey _ leasnexs3-
It fetiows tus typical Steps of machine |
—___eaxning ie —Olesfwoatton £ Jenerrolizadten
“as footie —ctlitsagex_teaxnex_is__
xeadu up ; » 2 + need
H2_"Sttes boce te tus Training Data,
= Lary Learners:.
So ‘t Comer up tow a reained mode)
ot tue end ap tue learning phase. |
——
“46, +a When +o +est dota Comes in |
igi. ti oO i
Eager leaxness +epe ME HMe iN learning
Pose +hon O24 _leamer:
Algosittams thot follows
Approach +
° Decision Tree.
© SVN
© Neurol iu,
FAQSK_Leoxning
-| 2 eS _SkIPS Hyg a
Nero);
| Peete
“Se, becouse a4 tus my \eaxnesa_goual Posomersic VS Non paxametsic cnocdets
Date
BF Bomamnerers Foginire: ae 4
- {tr uses featning Beta +O. elassiysing au
fa untae deta CIt just steser oven).
[T+ is OlSo Enouwn os tore leoning Cre
_PemMoyiratten technique based om ~weperarion) |
— | pue +n heavy dependenty on _iven Prat oing
—instance, it is _ojso nown OS instance _
—— eased Leorsming
t oo
7 |ta2y leaxn ve! AME jo rrdinin
because Not muh Oy ATsaini ng Qehuadly_ a
\nappe QD.
= Algosithm tat follows a2 __leansming t= =
© K- nearest Neighbor
# QuNMpRY! i |
> Eaqges Leamenrs:
= LONeY \+ weceives He dotaset, fr gtererd
Leorping |
=Then it doe net Loalt sex rest potra TO. , te Sl
learn
= | de tenes ang same in Learns SS & lest 4
in__Classipying potg - _ 4
i
= aie oreo st_s wio ve
=| BSS Stee te Teh
a
7 Stayt class: ive
- | Ltr teres Mona ne nc 0 aap __# __Moper REPRESNTATION 2 INTERPRETABILITY ————
= The goal ay supervised Leomring- is
¢ :
we Sx_dexive Toaget Puncton LOW Cla can oi
a-—_konget _vaniateie . = ae
“= Tp He __ Target function its well fas 1 in
by a Leaxnng_Aigo) , jt wil be ole +O
never _
a coms eoty classy c ser of Dota it has .
Seen
1
> Undewpi ttl ng 2
2 - ny
uoill not learn wer) prom tua training Dora, |
= A +ypical case of Unclerpiting moy Occ When
lwe try to wepresent Q non-linear Dato with L
dee linear finerten
Hebi
OF Suppicient sheai ni ng—Data
- So it will Resuit in poo’ perfomance with
coining Deka £ pecs pexpomance ov Terr dota.
tw Techniquet +o weduce @ Undew pitting
- Inesease mode) complexity
- Incsease +t numbes Of teatuxe, Pexjeming
Peaiuse engineering
- Remove Noise from pete
-_Tnewease _a_durodion Sn me
Page No. con
Under Fit
- Fox example fox Housing, prtce se Aycrionn
only Saft _axea {Ss Mot _enaigun +o preoiice
ate price we need more features lige No of
___pioot, Area where House is Intated ere,
| sagt =
- Tk we train model using Aseq only 2 by
simple Ve linea seyression ir will nor Give
___ocorrate ce sutts
|
7 Fee # best Fir Of Hoe mode} we need to Choose ;
—— Complex medel yaar Mere Number OF heatur C |
3
y 2 EN TN | |
= & [4 ! TP = Actual win g predicted wi :
__ FP FM petual Loss % precticted YIM a
_£N_HAP™ Actual win p predicted Tess
TN = Actual Loss 2 predicted ee
poets h roe uncity a ——
rs given by for No af covrseat
Classigicotroys (j-e TP 2 TN) pb divided
bY total numbers se class pico s don.
Model = TPtTN
Accuvaey All (T+ FP+ EN & TN)
Fox example,
Aetuol win Ketual cosy
predicted win ee 3 Te
_Presticret Tess z 3
_ g7 Te TOO | MPALIG howd Hume
eee sotee
7 inckicated
- Sensitivity FES US Wat Pewcenter
> Specipiaty tells us wr percemrage ab
12 Lo PReaee abee ea
posters wid
identigicd
.
nog ot
of _Misciassi pieatin
Uwewe comsestty
HOHE ignom> un:
Toe pezcenage
Using ewser sare
Fe +EN
6 "le
\oo
- Nodet
= Accursacy
yp Sensitivity Lp measuxel tus pe peoportien
Ot TP examples oy positive pimps 27
Qses uid were. Couxearty Classigied.
n
Sensitivity Te ss)
| TP+ EN ( Actial wirs@oass ”
es
| e7
Gt. tn
(How many a actual voins Ase Crassiyied ox
oN
Wins ) actual
( How many _o4 molignantr tumess Gre atKyery
_classigied as waey malignant)
AS tumcs per predicnien iS vewy Sensitne
pplicatfon, tus sen’ iWAy iy must be oO
hig as possible. Becouse malignant mut
be pxedicted as maliqnaar. Ty ir 18
predicted as Benign . hero Kam then ae
these example Moy Fate evexe problems
. Page No. —Prectsion fs td measure hen Wwe Want +0 WSeaseng
ge: e meas
focus on Linn i¥ing EP He
— Recall! dgaow tg i when We Ave poured a
. Liming EN Coord Spread example, an
Specipicity : ‘I+ measures tus _proportns ar '
Negative sampi@ wwoh have been Cov yeatt,
Specigicity =) TN = =a -
TN+ FP (Actual tess)
= 4 = 64-2‘ .
a +4
= Precision -
Itgives Ha proredttom ay Positive
Ped ictions
uduich ase Pauly positive
Precisions Th _ aS = 88 IOs
- TP+FP gs+4 eq
=| Ir _tndrose te Delia bility Bt a Meche)
- [in Predicting o Class
= OF Intere
LA model ipith ae
Tee pressions peceal—
_ to be me —retauie, FS
a
over “2p aakily ol pasty fate toe
_— Recall = TP - 85 _ as _T¥4y,
—____TP+FN © geya @7 "
Stn 87 ry ihehpwep erie becca toca ca 1 NTO Daren
CoxxeaHy.
- Recasl wesemblar phat pvoperten Of He Dale
tote) vwolns oer? predi “cocneatty.|
= Recall indicates jue peoportion af covveat j
sat Positive +o tua “reral_qumber
Naot} _nappe in, ley roexe
ecurre th Loos
model accuracy vdeally it should not happen
~ Kappa vabus oo model indicabes tus : |
od justed to “nodel accuray |
Ee fpalvalus Ge) ema pcadi=ar cite)
l= CPx) .
Pea) = proportion ay Obsexvesl agueement
between actual and predicred in
Overall! lata ser
Pea) = Ot TN =
Te+ EPt ENTTN
PCPs) = propasten af ex peated agreement
between actual & predicreal data
botw tn cose of Class OF mteret
3 well y 3 Kappa yalus can be Lat maxir4M , wie, Sl
ace ats perpeat agreement between date
's predicteon ¢ adic) volt ———~
Ge eP uaa EN - [ENT —
Al _ Atl ee,
— = \
= (21 18? Jah a eo
100° Jeo too © 100
k= Oy = 07886. 6.4162
L_ ©: 7226
- Femeasure:
- Tt is anotwr measure ae model perpomane
LA Oo mbines i
- precisim £ Recall -
- Here, we assume tat precision £& Recall -
have equal Lo elgut
_F-measure = 2x precision x Re cal!
— Precisim + Recall.
[example © Correat PredvcHon — © Nallqnant
| 45 benign
| Tncorvest precicHen T Nalignand
BS benign
Acwal orp
——<—
Tgnant | benign
ae ime S| 963 | <3 _Tt a geoph of sensitivity y
7 A ROC Cusvel helps to yjsualize perpomance
- = ‘ ae mete]
censitiNIH] @/ TeRC TL positive sore)
= TP
Te+ EN =
Falce postive Rote = FP
C1-specibicity ) EP+IN
fens} Psobguitirt |
da Yo ——_- PRO
1 to
1 [a _
a ~ _
[eee —_
o Jo]
AR FRED a
| a
In Roc , FP nate is plowed against TP
— at diferent gasification turesheils,
elias example, Here we measured 5@ tur resholds —
ee aye ee =
~__ Tp we assume aA lower Value of Classif¢-tcaton
| Heresbeld tte _mode) Classi pie mere item
as_post tive.
aa FOX turesheld 0-2, if Ppwedicted Valus [Prob score Yp
22> Classiyy ir as Positive
XK o2 3 Classipy it as Neqasive CI =
r D> In He ROC cue | He FP wate 2s Plotted
i Cin her2ente) axis) gaaings tue positfve
t sare Cin tut vestica) axis) at ditresent
classi ication ttrreshold. _ a AUC CAxea Unedex tur curve)
218 Fee 4 _
AUC is He area af +We dimensional coace_
under 44t curve extending Por (6,0) to ~
CLL, where each point on che curve
_ Ives a eet ap tous Ise positive
_ Que ata Specipic Classification +threshad
= The Curve gives an inedfcation Ap tue
—___ predictive quality of a model
—— Aue value an ~som OF! with an
= AEUC Hass tian O16: indicating shar tu.
Classipier has no predictive ability
Fig shows tts turves Oe Hae Classipens
-—_|Classigier and“ olassiyicns _
= {The Aue of Clascikter eo is
: tut AUC Of Clascitier 2
sal 80 _cClassigi yt
—Mere Haan
is Beten duan Classi piey 2. -__In Requessim task, Arey pitting a
ede), v9 ‘i ese Hat ve aa _
: by compartng mede) predictions +o Ach: _
: t
: Data.
__ Residuals CEswoy) = Ye= 9s
peqressi™
DBD tx Perfomance measure® following axe
Considered,
|
ce) MAE -Nean Absolute Exyey
i iSebtained by absolute cliggenence
—betincen_tut_model prechicins £ tee tue
Hat Cactual) values
|
~ MAE measures Average pen “mag nitude
ee dentfated by tu _wesuession modal’
ee
Saas ee nee nn ea eeaesceeeentOnSSITTEEE
ee e
nm
Syst -
Cal
macs 1
eS
NA ONLY Number Ore Samples
Jot f | f | |
= iy MAE 7S © means Here ts No exror
_ (2) mae: Mean Atosolute Cw
— It te squares a4 tue clip exrences ‘between
tte model predictims & wmwecRormriera
Cusaer Gdactuabr actued values is calc shores) |
MSE Values awe generally Lasger | Compares?
to +t MAE since wesiduals ave being
squared.
“In Case of doko outtiers, MSE will be
Much higher than tHe MAE
7 In mse ine He erryor IS being, squasea| \ ~—__Any predict
_ ; ly 4
ng eure is being heavily_penalined —
ee oe
a nt =
Root mean squares Ese
7 = ese can ne easitly intexpseted compased to
MSE because RMSE Unit -modthe Wty ——_
~ He output
RMSE = Rare
ate i
(4) _R Square (R*) - Coefficient ofp —
I we SST-SSE
eee Soa)
——|Hexé, Sst = Sum oy Squaxes_ tora)
7 2
} Sse ieyes
Us!
i
7 =
_t -
| SSE = sum of squased Exxas
} mal
fe SSE = (u-5)
— teu
es SST NS Show in Black Ink. CDigf ber” actual 2 meaniins)
SSE NS Shown in Blut ink Coith bet” astuchd bet tine)
_ Here, SST Is mean Line,
_ hen 8 I Good measure 46 evoluare Hus
moda)
-__The B- squaved valug tres Letween _o tO /
(0% ~ too.)
~ Larger val 4 R* cepresemts aun gb better
kit
Page No. rai Unsupervised Leasning — Clustext
— a
=| clustering alges iynms. Y
“Hut
______gveuping amongst
__ =| Howevev, it is
__| + pexzomance a4 0
= | these axe +o__ Chol
deta sets —
quire asic kY te __&*
clusteaing —
ae Ate
lang uonich lie 19 :
— Le. oO —clustexing-—— _ silhoutte Coetpicient:-
__=_Ir uses distance ( Euclidean Manbattan distances
uced) between lect elements —
mosr Commanly
as similarity measure.
Fa oO data ser cluereved INFO ¥ Clusters,
-, Sil haute — width = beety—- act)
max ¢aci)s buy
Here ke aie
ee
Sl
al igrance. between
ra,
Dl acd) is average alist
_ Lite clata instane &—
7 _ belonging te tue same_cluster
ST ba) i. ; jae distane ver”
_ bo) is lowest average clistance ©
—_it+ data instance 2 dara instances af ——
other olara_instances—__
all otney Clusters.
P | 7 - ee
> | Fer example, a
| age eT
k~ SN
1 [Xt wee Pi
| aes
eT]
x /
=F ] tusres 1
|
|
|
t a
T a * a \
7 a]
el
{
it axe 3 Clusters, aye QTven
— | Fist we i
[& Will selear [Fe dara pent prom
| Cluster L-
“i Mew We will cowate distance OF every
one Epsom ith di i
in F “fa_paiet 10 same
T
_Clusper 2 ee —w2ill_sake ave ap Se Cin 4u1s case Nizi0)
assume _Qci) we _ ger as | 0.@2
= Now, Wwe will Calculate, bi, (distance prom
Pte data fount te Other cluster's@each point #
ANG OF it
bin (0¥9) = ba) 4 ba) + baG+ —- Few)
M2 4
4
r Here no its total data peinis Th Clusres 2
a Here ae n= 11 ,
.
gz Same way we will Cakwate big ( distance .
Krom tHe arg point i othey clustex's (3)
2ach point £_ Av4 of tr) - :
big Cows big) + big (2) => --
Vs a
Heve Nas - -
- lets assume by Cava) = o-8 big cave)
{s__0-4 - zl
0, blz min Cbrtovg), bigtag) Yo of silhouette wo! attic
Tae rts ew
ined worm vores 15 V9 Of asl pe
Silhoutte width > 0- 8-0-2,
bor TH pone oR
“Tht Value ay cilheustte usite Rangss betoeen
=| 40 +1, tte Higu value Indicating
—HIgu intya - Clustey homogeneity g inter
clustex 2 Hebevoqeneity. a -
tema) evaluat - a
~~“ En ws opprach ; class lake) ie Enown por tu
Tr ts RNoton—that class labels axe not pasta, —
te data_used in clustentng ~~
— Clustew B has 4 ( whith has DB lave)
Se partys SeStt -
23 (A x¢s40)
=i
23
- A pusity Score of ft inelicares anor a!
dari points Qe in Covveat Cluster @020 00%
Cin 14s awn Cluster, )
“Se purtty Ts given by
pusity = | O_tmax Ce NC)
me
Here, lata set is havin ‘n' data insane €
\c! pneu Class labels vobiich aq enenares. ie
Clusters . Selesting a wode)
= Predictive medes _
= WOK Randorn Sampling
~_Stratiprea! Random gam
2- k- pid cv
= Basis for pod (Repeated Heider) _
=_K-,otd cv
= 16 fold av
= Loow
Tee lin
~lasy vs. Eagen Learners
— Node} Representatron £ Tnterprerability
Undlen pitto9 —
(2) Overpiti ng \ |
@ las ~ Vorvian ce
Bias ne Tradlcopy
= Evaluating Perfomance oF a Nodel
- ~_S4pesised Learning ~ Classipicat on
, } = Comission Nagtet x
-_ Roa £ AUC
= Supervised Lleaming- Regressing
—_NAE, HSE, ANSE, R
