Machine Jeaming Sadkin Chenay Unit—4 B53] evised leavening, Tapicg te -be diecussedt poi iat 2 ane ee att ot Sasi Aigna— p = Confusion atsix cake: o° Ave —_ K+ neowest Netgicbor Tae EZ - H Naive Boyerian. Classifier Kemels in SVra. = Neural snehwones: § | logistic” Reqnestrion Celnessetenion) Citect Gaited legit Regression) aod \dke} bain Classes ob | on © l eid Pluting ‘launch of examples ce __Sepesatiy oe fi ders Vines: Ay Feature Fe gutper | prredichive 10a Mee cole A seddosd in se) our toh je toger it | al? al? [ord T d predictive model 2 | af? \alP|s 0 oF) WWide gives « oubpul 3 Pa late ; : os of) | AL] 0 009 Go FOR EDUCATIONAL USE Pose 0} SsUnit -4 Machine leaning + Badin Chaney Pauloles._Suyppare We —intoduee Dei ob at} we fir a fimple Iineas ntearressio) __ zs line. 38 3 we Sos Bot Bm + Bex : irs Renavrale: of Wwelglr Ewe fick Wo. far enouah' point Ti ZzLo av v ‘ fh Nodue! of Comes ler} \ajghe __ Tre no i hoe Vi ana Thats Why We use a very Slnple ideal. "Squish on dala Anne. coanye fa ,\ + S&S Symod funetton. tr ig dePined oa 2 Cz) = Vite, AS 2, €(2)=) VA enya (22) lhe | eres AP eke 36 i c hs tv ro €(z)+0 y “ Led Abo 2s Cz) 20.5 fo z a dco ° a ewe predich 2 Cm Be + 1% + Ber)” FE fod We Can interpret as probalaiiity Het your o 0 i aes. f= P(yers«), ; We pant nialue 2 = Pot Prt +Pa ta Such pep LET} Wee in “Clase I en 2 has jo be realy skigh pie4 Lb has jo belong fe Classe Hren 2 has te oe ver| loi. FOR EDUCATIONAL USE. ‘ pege orUnit -4 Maange \eaming Badin Chonat Hoo , fuppose Wwe nok at a gpiap This ppink eb ii almost Cexrrain i H (29+) fo ite in Clase ome. Tiles OFM \ x8 owen on otrex Gile oF Nine We are . i mene Cortrain aloour tne Point tie to cleo __ A Clase tewp. Oo 220 we can Pwal oF 2 aS. 2 is Penpenticulas distance oom Hoe classifying lines So_'00 lescihyiney line z=0 4 j © 610222 bo.) omen wr el) ates \ Therefirse iy jo Now caer to. interpret 32 (2) PUMA. fore craammore. Sr framoid Nalue is Close’ bp ioe then & H y prt ib Means Snu_ame oo} freaW4 . Cerfrain aloouy _ t < how: Clone it jes Clase | \ i \ebs: Sony Sigmoid. 1g chose te 0-99 \ ten we know ar Nem New Garr auday —_———} A E4i ine. | ii jk - be lot Ne =| py Ger Suncken + Ja lagigric Reymeserion fenea/d: model iota > [Pe., es + = BolT Sea ePOu ; : at Ce ade ae CS FOR EDUCATIONAL USE peje 28 Gaes’Unit -4 \ Maohine leaming Sachin Chanan Ta ure usual earning, arcade we howe 1S fore mwen % > § V Cact (to tase LF _4rwund bry is Mibterent) Why for use least Square coat function. how’ Clasei fication. Tes abel y 59 a Tr twons out brig ie not good model: H > Nob a good Goer funchion. |_Tn_icinony Caesitication prvblen fos sen a] 9 {J || Suppose. we ase doing Cleseiticotivn oF Cancers ChT} Cop] Grd fo Cancer J oy Now Consides a Came where we foraty Mie Claas ity enamp\e A 9 ie 1 he Compe pak ta Incuséa foo fn eee, Hb very loud. Prak We are. J Dot able te Penaline Phe ae Lone minclass(Ficahion — a bec: dosh us Seuarre Coe Rinerinn Cow: ji Pie 9. We inttéad wee Ganebiing caved rea yk -entonpy Ices function -Tywng + a In 1-95} Which Was propetHes ag: @ fro ic yf OT Shalt toe very _IWgh foe mis¢laseitication ® Tyo. bur 6 ensntiendy (tae use js alwons Powirive), FOR EDUSATIONAL USE. Page olUnik -y Machine leaning Saohin Chanay Q|_potice J _alwoys Wes behween ett 0 bt | "Sy je either 0 oo | —? becanse Y 1S ground bith Yrewefowe in (PIs negative .. and the other term 1s atte Y=ae negative A we want He T°, Hobs Why theiy, Yip = qo in gas Geb FuncHen. ® Jyro 20) > Clete to cow Tx 0 2 5 sty > _clese fo “one To © @||_ iets Soy yto Yxl 2 Clese te one & 0.39 To-day} + (9) In 0-F)Y = 4o0'ln $o.93) + (- 5h th fiiicles2);) = 4.605) w -owreop is ley n |v Te ysl Geo > cce te tem cre alse (Sea ae ees ae ce eceed pene y ce sae eaters a a Je “Ju Ino + GS") In (1-8) 9 Lr ; © | = fovi9a +0}! 2 lon o 1-Cage OF Srvgclaaihication rem the Get function is Ney Wigh Tn Cage Gf Comméch Classification Hen J 22-0. 1 ae | a j Li eet FOR EDUCATIONAL USE, Pose OSUnir-4 : XI Machine leaminy Saghin Chaar Eracnple OR Gate ria! ClaneiticaHon: OR abe OR(% .%2) f ° ~ 0 = Per Bim Ba? x 3 Ficus A ° O90 ° a pe exe! ! 2.098. US yx 2D | eG) : 79S. ¢ lek x wepsesent O t 1 ) ros and 0 “Feppesent it a ' as = | - - Ov Lie Fpos OT Daw aca =OS, ie fens Zs Hx + P2xr* Fo, | Whab Weights Pe Ried Ba (BD win classify same ab of gabe | 2 2 ath | ye remembers Prat tre Segmoia Sanction Wad Seem O dos Te C2) KOS then 2 hes fos 25 Alanis jo foe Negahive, : 6% gimitaly foe remaining 2 2 hes fo be postrive . . So \oe__Know) Hat 2 = Pot BiX) + PoXr fos Pieres Case hen %\= X2 =O tren Po has to be. Megsrive - Alea police that Witowt Bo We Coutl not hove - Made Hug Possible at all. FOR EDUCATIONAL USE Page ofVoit-y Machine leaming Sadin Chavay Ve alge Know: Heh Pid Po hes. to be positive to make Hig Case fo Wok. in What seb oF Wwelgts Wi! WoT Bo = -1 Pis2 (Pole o Hote: © We Bo is essential (need foo Bias unit) @_Se# Po, By 4 Bai ase nor unique. Rre_creswh of! logistic weqession peed Nor be unique. (Ine can _mave Baake 2 boon) Ro + Bx + Pix. 272. Interpretation Hrat classify va line. Hie ze0 || Logistic Regression Some Re 0 < Te Phe estimated’ Proedbiliby’ 15° Greater Hae. Soh Hien Hae model laps. “1 o ise jb it dees nori(je. i+ belongs to Hre Negative Class, lebeled 0"), Gwis makes ita lsinany Class ne ie jut like. Iinear Resaestian tnodel, a _logiric Regression motel wression Mosel Cepia ton she op ne ii a) 0. és % B= hpls) = CRT) a FOR EDUCATIONAL USE Page oFUnit -4 4 Maaine lewmins, Saahin Crswan The_losistic - need CC) is a Symoid function poet - output A_oumber loetween 6 and | - This defined as é(2) © \ L + exp (-2) | CS eas sath = || ecco eee © — 2 once the _-logiatic Regression model We estimated | the Pasoalcinity 8 chp(x) Hakan jngtance % belongs Cagily Roieaau pol a geile Byp co Miss AABh chases pagan) temonan 3 | Notice Hab G(2) COs when y Zh0 and. : O(2) 26-5 When ZO, .$0.A logistic Requessio ‘Todel Predicts 1 jf 7% bis) phitive ahd 'Oo ff ik 1s: hegative . FOR EDUCATIONAL USE : | ot Pane -08 |Unit Pra Machine leaming Sachin Chanan Examples @__NoR ~ Not(ok) Sais Jor oR gate ae) 1 we Set Wel Wye? Wier o3) ° k } s\o ° Fimilamy fos’ NoR Gage 1 ’ ° Weel Wit -2 Wee -2 @|_Any Gote e Xo Xe oy We= 73 Wit 2 We t2 000 Oe Oat ° \ oO 0 l Veal \ | @|_wano Gate \ ‘i : xX XY «€ © ° hy Wit o-2 Wre-2 Oo ) | 1 oO 1 ‘ ) LO a © a) | EDAD oe bai otR Saas FOR EDUCATIONAL USE Page ogUnit -y s Madkine leaning Gaehin Chanay | (zs 4 | 1+ exp (-2) de 1 -1 det = > : dz fh (1+ expee))* dz { 2 1 ae (14 exp(-2)5* Aen) dled : = \ re] \ + exp (62) | enp(-2) SOM 6@)Ei-e@)] | red o0 | OR; AND, NAND..-. > Guessing for welghts Without using ayadient deecent. AS in _atsove CSmanples Wwe. bane ooly data points and We! hone te! dear | Clase yi ag. . inte . Tn_generca| it is ob pottiile to guess. Inerouse. We fail lea WWe_haue to See haw to Caltwlate Garadient of “egidriéeéaeseian.—ta_find_aut final. Wetglakss Fema FOR EDUCATIONAL USE fage 0Daita adhine leaming Baalin Chowan Gre logigtic _aregméstsion. ap eT i 0 ow Sea Gradient Descent. oT BW —e | a fener medal? Now the. a Haak we_donb Know) here Wo, Wi - z ae in Case of aa Resreatsion 2° OFT | GR TO ari SOT {OT 89 2 > dein eNT : WwW ey 22 2 Rule. e—| NODES Pie ts at dependency ~ {'T Changes leecouse Change 0 o 4G changes Wpecause Change io 2 and 2 changes due to Change iq W- = n \ oe (TG -2 yin GO} (ay) ty (1-9) fe \ ‘te Pow numoer of daha points. Otel : Gand FOR EDUCATIONAL USE i peqevnir-4 % Masking leaming Saolin Cranan Eramples x y 7 zt 1 HO) 9g RQ yor Ae Teo) 2 x2 x, 2 Xq> yd GO? ym 7 7 i T m ae xt xa” yom) ym 712 Teed JO garde au individual baea Gla neea 2 Ta & HIM jw in 2” : er as) apt ar oy QyeO- Bz Bw | ‘Je =f yin f+ t-yin O-FT aT. =f ye Gry ~ 7 y- 29 VF (HF) OT em a ee 22 LBZ ES 2 i = 70-3) Hz OT OY = > (=¥) | 9 (1-5) FOR EDUCATIONAL USE Page ieunis ~4 Maskine leaming sachin Chonan [oo --- wal tani Maga ha, OT 2T 2% 2-1-7? on Bz. 2B 4 Touva = — Eros * input: General Expression _ Llesietic Reyrestion vali ie i Bina] loves entropy | OT gin 4 y§ 09 ae QW. in re Soe eal Component of W We with toe faxing (Comegponding Eva “Hy: the Gare, eneprsesssion) a liness Reqvession wih. leat Squares. 4 \ ~ for Nogisric Regression He!) 1 loaigric, oF near weleHon. = for logistic Remrestion Hae lotr Gunction Tis nok, Convert . Ly The minimum in general ig Mor Unique - There are. many peeritole Uassifyinoy Vine - J logistic cegresmian dees not give unique. Goliubiogs — ___ FOR EDUCATIONAL USE Page - 13r : boik -4 ; Paaline leaming ‘ taabin Chonan OR Gabe = x y c ' 00 ° x => @- 5 Ee Sot 1 i 3 16 \ eat On. r OH p-— © L eas - O uete [ | F = 4(2) L * || Pemfosmance valuation oF Classification Agata okie predicts ay sloctiasate —Wlaicl Categarical : 4 1. Walt! Hae employ lean Hae efeganiza Hen, or s}oy 2 - | 2. Dees He Patent howe Cancers 07 ek? Does His Custromert Part Jobo ligh MSR, medium os ov jow Sex 2 bi! 0. deloatdtole in tems of hid m mance * most : cePlanait. Cespecia Loy aden See 4 ane cs 7 0 claseleigaHen Model owe aecumaory lade ee — 1 Speci cic! fo o a FOR EDUCATIONAL USE es | jea bore Mashing lewnting Sachin chaway Confusion mahi» A (ao Bien x | Contains the output of a binary Clagei tier “ha A tonfusion makin is a-palele with ws of the matix _ Indicates Hie obsened ov actual Class and the columns indicabes the predicted | @ Class.iit Jak ue ook iat the “Gn fusion Matix of einer clases Ries Haak predicts the loan defauth 0 Indicates Hat he Customers Will poy Hae laan and 1 todicates Haat Hie Customer wird € dofaut = The passitive clase for own furtnen diseuseton io | fogsrlab|s, Predicted NegaHve | posirive- actuet | Negative. (default 0) | True Negahve | fotse Positive - dlosenved Positive (demu =1), | Fatse Negative | “Toue positive A customers Who will L LA (usromes Who will elually not dePauit being _oPpedicted. for defaust being Predicted predicted Cowreatly i 1 ae one Who Wil dekawsF closenedd| eT] FRO. € Lape LTP wi ‘ e Is wo AL estemers Who wit A Customers whe wit. / ; defaurt boeing. predicted atualiy defaut being Predich Corpveatty , 0 ene who win nor defau: i \ ‘ to. ra emalr model , yemnance_ meabuse fyanslates to in berms of. Con Fusioi | takes FOR EDUCATIONAL USE Page Is~eo F Machine Veamnins Baokin Caney Accuracy AR a | Accumacy io He oumloen of Conmmeak prredichio Made_loy Hhe model Wy Hae total numbers oF crecevate Accursaay (Acc) = Tp +TN TP+TN + FP+EN The beet accuracy, $6:100% jadicating Hrat adi phe TS it Jood meahwre of performance 2 ‘Tr depends on tne dahaset. | ___Accunmey 6 good measure wahen Hae target Vamlalole | Classes in Hoe datanet dre ormpugldy lealanced. i€ we hove ‘a dataset with Ss% of birds & 4S Yo Sishes . Then We accusaey Nalue. OF Hre model, Wich + v 1s Oasrsit-ying, Pwis dataset, Makes Sense. ‘ OMrenwise , Aeturaey does Pot Mean muel, to an. meclanced- dabase} E Sus instance, TE we have 98 images of birds fos imoges -_ lof fishosh and_we predicted Haal a of tre images ave FOR EDUCATIONAL USE Page |6Unit -4 ‘i Machine leaming Saahin Chonan Songs ity or Recatl | coritivity (fect on itpoiauteebnte, QuB)aI6. Gteciated <2 | Pne_nunvoers. “on Cowreet postive. predictions divided . bey Hee yotal ounvoewr Of Positives . Th 16 alee Called aS Vecau (REC) os true posttive Kate CreR) | sensitivity = TP "TP + en. alge Negatives awe Jncluded hese because Hey ave aotually | the _(ovrect Pwedictions too. t ee Aunber of RN'S ows model predicts, He lawer | NaLue oF || 26 & good measure OF Pexformance 2, ¥ | she answer luighly. depends on the tage ak hand: i Hie oul abient ab Me increase bho false Negatives (En) envor. TF we SRTEIITES bn Navi fngrente. 0: EN iS wi Han the cost oF FP because enemy EN fst elie peng SS cane while phe? Ep Win nor. : Amos eulemane Who. aa datas Ne ; faise. Negatives. 4 speach noes mt FOR EDUCATIONAL USE Page 177 bees i ait \ Masking Jeaning Baglin Chana al Specificity : eh 4 fo aul those Haak aie Cormeobiy identified 6 -not having Hae Condition by Hae Model (rN+ fp) Specibicity also ‘Catted af “eu eo Negative Rafe. Specificity = TN tH N+ Ep Phe_hiqhes -numoes of Fp's ows model poedichs, He lowers ne Nalue of Speci Ficiky ' rai grease aslaagetem it lI b——— Bien ‘simplea; ik we Hine of epesigiclty. as a fyattinn ot negatives nak are. Geely _dentieied Hhen_a_pon fer! Fo_obnen_wdords,alaigh Specificity madel_wain_Cemeaty, diseasta__almnst@venyone {dha does not hawe Hae disease ff Loiti_st_qenemie many aise pasitives FOR EDUCATIONAL USE Page -18voir © Machine ) earning gaahin Chonan | Porecigsian. Drrecigtan ié Hae cats of Te! positives (Tpe) te all the. SCR LIN SPE RT TS EE Precision ( Prec) = STP TP+EP. The _hiqhes Nurvoes OF p's ows Model predicks the lower t tne value of precision. I [Je it a Good measure of Performance ? Tne answers again bagi depends on the task at hand - I } : : to identity An email o& Spam ox na épam. Te semi identify an _inpoftant email 48 pam , wie intrease. PN became! every: EP .Case Will make usi miss an important J lemail:, while We can deal with euersy paihel Hence vin His Cake , we Want to “ger eu ingen emi) Sn other wots, we need +o: minimize: Hae. diumiosae! ne innperfrank | emails identified as: gees 0 Qtoout the Pots ti iO nimi: 4 20 and, | Vae_aim for a PemFeot Ppuecision, it. Precision =1 (os 100%) Precieian V5. Recatt Maferstunalel} impsnuing dre atten reduce another, Toolset [Wonide,aredusing FN incmeates FP, aad vice versa... FOR EDUCATIONAL USE Page -19| __But, in_ovdem_to_popealy ou uate a models performance ; ___ _l|We have te enamine loots Precision and Recatt- Bub, What if we Wank to getian_oplmal sialue 0f loots lot trese. metrics 2 he way todo it ts 40 Comoine Hae —_ in the Go Caued fi > Stowe metric. 3 Fi Sco Fi Scoe is harmonic mean. oF the poesision and Reco. sey Fy Beamer (Pi) e1¢ ocuidesm i 2 | fill_ninps hu thad ¥ [Rec ~ prec} / a2 * (prec # Rec) (Prec + Rec) Fi Stowe pavvides a loalance between Recatl and Precision H Co" 20 Be igh Nu o. Stove 1S equal to high Nadues oF. loot Recau 4: Paecitsien meaning Hak We are minimizing ieki Fp). FOR EDUCATIONAL USE fage - 90Unit -4 Madine |esmin4 saakin choway Roc Cue»: bo ae ull An Roc Cume (Receiver. opewating Chovaskemietic Cume)) is a graph showing He petroomance of a Clagst Ficotion made} at abl ClasedFication Hvreshalds. Phe curve plots two | Parameters. ‘ f =. “Tie “Positive ate . = false -positive: Rabe i is therefore defined a follows: L | | sre positive rate (TPR) isa Synonym efor Tecal and € Tere TP TR +EN. Case Positive Rabe ( PPR) is defined a8 Follows: epg FP Fp+,TN i | An Roc Conve plots. TPR vs FPR ot wieterent Classiricalion ClaseiFies et [ tmesnolds.Inwening Hae claséieicabion Hrmeshold Claswieies | mere itens as positiven, Thus increasing. tooth false - Pagitives and_jqe - positives, i . ahypica) Roc Cusve. eats “arp ve EP Rate ry CR 7p ye Ep at one deaision F 7 “Wreshold- +H Rate 1 fake at anothey i’ Deaision Theeshela - ob o FP Rate. Mat AY f Compute. tae Solas in an Roc Conve we Could With different Clasicication twreshald y\but His would be ered ee FOR EDUCATIONAL USE Page - 2)i Packing Jeamniog | Baan Chore uni -4 ‘gan efficient , Sorting leased —___ algeria “Haat Con provide Haig infiamation fore us __ Called AUC || Aue: Area lndex the Roc Carve . Pllc Sands Fox ’ Area under the Roc (unie”. That 1S L bce 5. o)- to (hb! Ae 3 a syoilosontt CQ eee ones a ee eee ep iba'g ts abs INWANNNAN NAN NNW NPOUPNPRPN PPP PP pry | 5 t i o-¢ : eae @ Actual Negative io.) ‘m1 Actual Positive. ene L FOR EDUCATIONAL USE Page - 22.Flaeluine leom tiny Wnit-y Saglain Chonan. Alle_srepresents_Hhe_papbabil it}_thak a_vandam positive example. is possianed to the Might of Atandom Negative example. Alc wanges_in. Nalue from o to. A. mode) Whose predicHons Lore 100 % wrong hasan AUC OF 0-0; one Whose Prediction, ome 100%. Correct hasan Alico: ehiGha re siPii |ANc 6 desirable for following two reasons = Auc_is Scole-invariant. Tt measures how Weil predictions I > Aue i cat invariant. T+ Measumeb z ithe quatity GF models Prediction jmwespective af Udhat classirication | Hymechald i4 chesen: | However, holh Hnese cseasons Comes ith Caweats , Which mo limit Pre usefulness of BUC iO Cemain use Cases: : : | 7 Seale invamiance '5 not always desiralole- Gi Wel Cabilrroted Negat No Sal ‘ ih infmize one dupe of Clagedeicatinn enrver, er ny J fox enample When doing email gpam deteation , You ieeky jon: inieni jHves (even if Hak vesuths 1 ino Significant incmease OF false Negatives) Powe 22 ECuanveles | Gnfugion Mabixs CRypeot Gof) Pane Positive CTP) ¢ = The predicted Nalue matches actual Nate = The _adhual value ab positive and He mode) Prue NegahvesCTN) + i The peredichd Vale mabehes re aotue) value. iv predicted a Negative value. False Pogitive (EP) - Type-1 Exror: + t -The Predicled Nalue Was fnjscly Predicked = Pne_aa! ive. * Parodicled a gagitive value. =flso_known as He Type-1 Ertov. Faye negative CEN) Type -2 Exrore. +——_L—l =The psedicted value “wae lool Predicted. | = The athial slaiue Mis pogitive bub the mode) - 7 $uedicled Oo NegaHive Nae = Aleo known a& He. type +2 Exfor. Tn given Gnflusion enabix Caleutabe . Type-1 and. L Type. =2 Essors. TWpen\ Crcos.: Actua) was E | Aetiol Nalues “negative Prediched as positive.) Y= 560 Type-2 Cover 2 Actua) was [: 33 2. $e 3. fasitive Poedicred aS NegaHve . bE: Se Type-2 xpos = SO L f Page 24 23[Carntin[rase ~ | From given Confusion matrix Calculate aaeurracry Recalls Prregision.. specificity and f, Score - Tota) numberof observations = 000 | ows oolasel_inan example is imestanced dataset _ | __Prrerre-_aete- 9ha~cdoba—pot H % “ | qp2ao , Ty= 430, FP=30 FN=10 e | So__aetunasn| of ours __Model tw out fo be. l Accuraby = 30 +430 369.96 H 304+ 30 +930 +10 1000 Recoil = SUE ate SOs es 0-79 : TP + EN go+lo I I Prpecieion = al STPin Noe wl Suid nn 0S i TP+FP 30 +30 Specificity = “TN 930. 9-46 u TN+FP 930+30 Pi Scote = 2 * Precision » Recau Precision +Recott = 2¥ OS HOTS = OTS = 0-7 oS+ 07S hes Page -25~= [Camiin] Page ] Confusion Malsix fox muti class ClassiticaHon. leks dima a Confusion matrix fora mulbicless. Parolblem Where We awe to predict Whether a pentson loves facebook, infstageram_om tend Confusion matrix ould be 9 exe mabix: Actual Faoebook Tnisragram Snapehat Foselnas | pve ve. ~ve a 3 \ 2 3 3 tagtentam | Ne ane ~Ne 2 4 i gi Snapelrat Ne -Ne. we 7 3] bck The _toue positive. tme negative. fae Posstive and ive fax Cao olaas Would lee Calculated lj _adding the e Fatcloook Tnghasaram : Te > Ceti TPs cells Fp = Cele + Cells FP = Cotly+ Ceug TNs Ces t Celle +Cey + Cells TN 2 Cony + Ces telly telly FN = celly + Cetly FN = Cee t+ceng. Snapolrot | Tez Cells | Pps Ceu7 +Cely Tr = Cott Cette Tella t (ete EN Cella tee =a[Camiin| ase An Roc curve is a qraph showing Hie Performance SF a - Clossitication madel of att Classi ficabian Horedlatd _ Puis Curve plats two parameters = True positive Rate - - False positive Rate In Hig Cate. We hone loeen aiven With dotagel. & # Class PIbabityS Tp ep TPR FPR. \ Pp o-gor PL o 6-2 ° 2 ° 0:80 pao ° ou ° 3 nN 0-70 2 \ ou On 4 P. 0:66 Bisst bh 0-6... 2 s Pe o-ss 4 ! O-% of @ N 6-S4 q 2 O-8 O-b 7 a 0-s3 4 3 Os og 3 nN 0S) 4 4 0:3 o% €. 3 P 0-50 Ss 4 ho O@ 10 N 0-4 S$ ss M0 120 {rom _eiven deta we feed te Calculates TPR 4 FPR Tp: Tove Positive. fp = Solse positive - —pR = TP FeR= FP P n: page 27[Cain] rave J Logistic Regression Ertanple. logistic Regression applied ap Senhimedt analyois to pedPerm clasedeicatian honing two feabtwaes (% 4, %e) Obseniations are given in following “tatole IL l % (+ve) words Xz (-ve) words Y s 3 L | 4 ¢ ° ( ; 3 (x= 01) Sind phe plated | poswor : wo » a | Bozo Brz0 Rizo [xc ad eT Be ei Be | 5 3 \ os 0-698) 08ST ors ots 4 c O 0-877) 236943 09-0377)» -O-feo8 03762. 5 = 6(2) @ where, 2 = Ret Bre +BaXe 2 OFOXS+OXKS 20 72 €(2) = ) sisal 2 Leenp(-2) teeplo) = Tso + G9) 0-9) 4 s- Liinos + (1-1) In (1-0:5) P = - §-0-6931} Ts 0-693] ——| paqe zsal? = pO - of oT 2B = ot for) fyM- FO px 26 + Oo} KX OSI Bis 0-057 Boe at Be ot) ar a | = ote) {yY-JO) BS cr_| 2 OF OL KOS AE - Bl? = cas L A = ae OT | Re | eot ty fyO- FO V2 | 20 FOV KOS KS L BL) = ons | | FO = 6 Ca) = © (Be + BMT Bak) ce | 2 6 (hosp 1+ 0-9) = ©( 1-35) Ee oa ai 1 + emp (-198) Tr 014227 el * O-8771 rar Bez 0-0 + 01% (0-877) HI = — 6.63771 Biz ore +0-1 X-O-8771K 4 = —O-Jo0g: Bi = olS toy X-O'B77IXG = 793762 page 30classmate! a ; Exi] Roc Curve E Fuple # ass Probabinity TP FP. TPR EPR \ P o9s \ ° on ° j 2 n oes \ \ on on : 3 2 0-78 2 ) 0-4 il : 4 Pp 0-66 3 \ 0-6 ow : s nN 0-60 3 2 0-6 oO-u € Pp 0-55 4 2 og Oy Gare 7 N 0-53 4 3 0-8 6 l z N Ose 4 4 08 o8 3 N Os) 4y Ss OF" 1-0 yo Pp o-& s s a) 1-0 l i Emm given daln We need fo Calculate “TPR & FPR | p= Twe_ positive cp = faise Possiriv TeR.= TP FeR = FP n ae 2 6 Roe Cae ey | eed Ailes (0-2 Koz) + 54 ee (3) (0% * 0:2) +08 408) i 7 [ = 0:04 + 012 + 0-48 g a = 0-¢y y Fo AUC oF 0-64 3 ct | Tralee boslaverare (1- apeciticity )Camas Natidation < nntidabion ts a techaipue. Gr siatidating the model etn bing ian he_cuteo6Ipal dad and aig 60. preuiously ‘unseen Gubsel oF the input aga) __ We Canalo Gay Hhat uit is a techaique bo check how) a j2es fo tnd jeAt dataset . € I “Pn madi i i } the. Stability of model! Th. meand--laased. ef Hep only. oo. the troining dataset we Can't Git owe onadel on Hae for Hvia purpose Wwe “mesemve a pemmriculan Samele of Methods whed Sons Comes -Nawidabsn oleae woe ae) \. Nasidabion se} Approach t wot 2s leanel - one- ouk Cavse. Natidabians 8) Neonwes P- otk Cymes ivatidatidn a) ei f es 5. Hold -aut camss Natrida bon FOR EDUCATIONAL USENatidation Set Appsoaah. Woe deuide owe input datage into a feaining Ser and __ test om-Natidabin Geb in Vatidation Get approadb. =~ Both Hie Subsets ame given: 50% oF Hae dafaseb. | But it has one oF: He brig did just using a S0% dataset to bmain\ pum model, 60 He Model may. miss ouk to Caphwre important informotion of Ye dainsth alse Fede to hve the undead mel > 2. || Leawe-one- out Css. naticlation € ell tants! on livide the data ioto train f fest sets ‘Hour with iiialenle FL pasa) santas galt tia ty Rai na gubaels ne Selec la_Sinale dbosemniston)ab teskidata. and, eNOTY Haing | O1se iS Jalocled a6 Hrining data 4 the model is trained «| now the 24 obsemiation J selected a6 fest data and ||Hre model i6 krained onsithe semaining datas) > f Jeane= one-out Om9ss Natidation,(.beocy) .3$ ap, extveme (Case of | K-fold CV: ny pata : Tagine if k iS equal ibe 0 where! nis) ne numeer of Samples in tre dotasel. |!Guale tenfold case'S equivalent to Jeaue -one-out technique « boon dal 1 L The A\yorithm of Loocy -technique + Patna team fh the abr seteakde alle jew seb. rf pl 2. Qne oremaining -ArdiSamples, Will be He. braining sof 3. Frain Hre model ..on training, Seb: one Cael iteration, a Qe model must be trained. ae yo 4. Natidate an tear éebriiinye Aawns dune Liotf 80 S., Gone He, weputt OF tHe varidation. FOR EDUCATIONAL USE poge 24 la 2 vos . tl: Repeak Step I-s n Hmes asifos 9 Samples wehave 0 difterenr foaining and tesr eeks . 7. To get the Pinal Score oNerege the results Phok you got 09 Hep s. kK Toba, Samples I : GY | iteration tn A lyerotion - 21N € : 2 paint [x titeration. 3/n I YA iteration, nin [for oey Gtléom also has a: builtin MeHrod.. Er can le founa in the modeél- selection librory Codes _Smpoth RUMP. as.,0P. Lrsom _Skleoren- Model-SelecHion _impeyr ~ Leawe Onedu -€. teins ; : hide = np erro. (Ctr 29). t a = LeaneOne Out 0). for _twain-indem , temt indent. ia, leo: epi OD. paint ("Teasn:", train inder *TESTs:, best index) 3 Xtal Xteskoe x Phoin-inden],, XCtest inden} Awain. Yetest = Cine Ly Crestcindex] tian The 4 oocy 16 nah, it doesn't Waste | nuah data We_use only on \e Sanaple. | form -Hhe: Whole\daset 26 @_test Set, Whereang the cpeabiis..the shmaining Seb Gases} FOR EDUCATIONAL USE Page 25/42nN But tohen Compamed ii Ha k-ford ) voc wequised building —|0 models ingtead of k models. Where n ib Numtoer of Samples land_it ig | much higher Yran Kk. Tk_Means _Loocv | move Cmputationably expensive than K- Gold, Et Moy take plenty of “Hme +o C1pss Natidate the Mode) _Ubing Loocy .|| Leave -P=0Ut Cross Nabidetion. Leane-p-out Owes vatidation (Lpd0c) 16 Smilax to [pow a las it Orestes GU poséiole training 4 test dels ley _woing € P Samples as the teat Seb. i Su v 16 tue 4 for Lpoc. Brill, it is Woy menHoning Hrat unlike Loocu and Kfoid “fear Seks' wilt CNepap Fow LPOC TEP is Wigher Han! [4 : ique: foot Sop: or hte | 2. The wemasning 0-P Seizpien wilh: ee the _teaining Set 8. Train the model op fhe! Haining Get. on Caah itesation~ , G new model Must +6 +ratned) »: { | 4. Natidate” on’ test seh! S. Sane the -wesutt of Varidation 9 6 Rope: 2-s Cp Himes: y 9. Po ‘get! Wie! Pinel S. ‘ak 4 uit Hat Got on. Sep S. potion Leane-P-out ty iubing. Sklearm'- Sklearn- , FOR EDUCATIONAL USE fase 26/42|| Code iMpow pumpy ab np ‘i Geom: Skea, model _6eleation impor Leane Pout X= Mpranroy (Ci 21,08.49, (5.61 ,Ce,77) Ys npraray ([ 213,47) Ifo = Leave Poolt (2) Lars train index, test index in |po.split (XD: paint ("train:” , taain- index,” Testi”, test index) Ytwain , x test = X Ctoain_inden') x Trest_ index] | ye train, Y_ test 2 Y Lheain. index] Creek index] Leoc has a the disadvantages of He Loocy , but, Newettheless , ifs a6 Tobust ab Loocv- K- Fold Gomes Vatidation kK. Fold inhrmduces.a new iwoy .of Splitting He dahset Which helps to Overcome te” test only..once bottleneck.” . Tr is a technique Haat minimizes the ditaduantage fp the held -out method: feone Saas Phe algorsitimn oF the. k-fold Teh iaiies | Prete “ine number op foldg tk’. Subway 5K {$5 6 10 lout you tan, chnose any umber uihich: is Jess Haan. Hoe dalagete length | 9. Seli inte kK. Ide 9. Choose kel folds ‘a6 Hne. training Sek. The remaining Gola vail Le the test 'seb- at 4. Train he model: op.the training! eh. on each iteraton OF (786 Natidation, You must "rain a new model independenty. 5. Naridate on He teot Set - | £.. Bone Hre.rresut of, the vetidation . foe a FOR EDUCATIONAL USE, Page 27lqe~~ [ 7. Repeak speps 3-6 Kk times. Gashitime use He ‘wamaining fod as tre eet Sek. Tn the end you should hove — }—Nalidated fhe model on every fold Hrat you have. 8. To get the Final Serve average Hhe results rar you = Got on Step 6. L| raining YZ] Test ka 10 OWMWWAVT ET: OTT SOMiIny “To _perstasin K-fotd Ciyusé “Validation you can se. f Skleomn - Model_ Selection: Kfold impory oumpy as np ged 1 fe ! from skleamn. model-Seleation import KPold if = Op.amvoy'(£f5,21, £8.41. C5,6I £81) Yo s Dprarroy (Ct, 4) ‘ Kf s KPoid (0- Splits = 2) i a sil) As tir brain inden, teshindex sn kb. Splik (XD 1 Poor !(4 Trains" trainer. * Tess? test inden) , x train , X-reet = X [tooin_inder] |) X [tosh index] Jebrain i Y-test =" Y hvatn inden] 14 Ceeet inden] ee Ty is alweys beHer to use k- fold, schon iMs\ead of hold - out. s ‘ FOR EDUCATIONAL USE poae oft oneTn_heod Yo head Gmparition'\k-fold ives a move Statele 4 trushootthy xesulk Since training 4. testing is iperformed of several different Ports of the dataset. We can make the ovevat! Scove even move atoust if We increase the numbers of folds to teat Hie model on many different Sulo- datasets . 2 SH), Kefold Method has ai disadvantage. TIngveoting k wesutts in Hraining Move models & the braining process might be wealy expensive Gnd time Conguming . Hold -ouh Cass ValidaHon, Hold! out Owpss Validation ig the “Simplest and Most Common technique. You might nok know that jt isa hald-out method but you Comainly ibe it Cuery doy, ' t | The algorithm of hold-out techniques. \. Divide }re idotaset inte Hoa! parts: He training set 4 pre test Sek. usualy, 86% of He defaset Joes to the twalning set and 20% to the fest Set bubiyou may Choose any Splitting Haak suite Nou loetHem. 2. Train tre model on the training Set. 3, “Validate on the fest seh. 4, Sane re weautt of the validabion. Test I Train WH. /- ———— Tota} Datage} —>| We Usually Ube Yhe hold -ouf method 6 lawqe Dahagehs os it (Fequizes taining the Model ony once. J - FOR EDUCATIONAL USE Page 2g)yo_TToperttorm ihold-out CN xfourcae, use: Skteam - model- SeleaHon + train —test- Split Import oumpy ae Np from Skleam -model-Selection avon drain teat- Split. Ke yas. Np. anange (10). weshape (( $2), Tange (s) | me-teain., jest, ytoain, Y-testi< frainatest split (KY » ‘Heep ieize < 0-2. Tandem State =0) Hold-out has a Moor disadvantage & y s enen- lelicteeubinn Wise: TE 65 WE mey end up sn a wough Spor GLtenc: Hee -Splik. “The Hraining Sek Wilt, Nok. prepresenti He feat seh Both training & test sete moy dipvor. a lot,one of trem might be cation ov havder HAmrewent _Mireavent the: fact frat we fest ous model footy once Migut toe a botHeneck: for) Hris, mMeHrod- = Duets the peagons: Mentioned: before’, the. wesults \|[olohained ‘toy Hae: hold outstechniquemey be. Considered - inageursote. “i ae ‘ FOR EDUCATIONAL USE page 36/42classmate) ——Hetue Bees clas Si Flere pusing iw. UgoM Hac 1 a_Supeni ed _leorming, algo Ham WWidh is loased on Bayes Hneorein 4 used for satving Classiticalioo prolalem. | Tr ig mainly Used ig tery classification phat include: Naive loaves Classitiey is one of Hne gimole £ moxt Shope ond taste Haren ved, Gpemicol. and weet fruit Bones Ee the painciple of Bove Theorem.[Camiin|rese Boyes Theorem t Tr is atte Known as Boyes « hich I Used to P(aje) = PCBIA) PCA) pcb) Cale) i. 4 iliky liky_ of hypothesis H Aon the obSenved even B. p(By/A)_ is likelihood padloasoiliry : Pmbaboility oF He evidence Qiven tak tne Pedoalei tiny of a Vjpotesis is toue- PCA) is omiow paoloabaility + prvbdlaility of Wujpothesis. before Obsexving the evidence. £Ca) fo momginal probadliby : popbaloi lity of. evidence. S | OUTLOOK Temperatune Humidity Windy Play QolF [° Rainy Hot High Fase NO 1 Rainy Hot Righ Toue No 2 ovencassy Hot high fase Yes 3 Sueny mild Big $alse Yes 4 Sunny Cool Novma] Fase yes s Sunny Cool Nowa TPoue NO) 6 Ovesca$t Coo! Nome THe Yee Y Rainy mid High false. NO 8 Rainy Go] Novma| aise Yes 3 Sunny mild Normal Jase 712s page 37(Camtin| ase J OUTLOOK TEMPERATURE HIMTDITY WINDY PLAY GOLF Jol] Rainy mild Nowna] “rue yes u Overwcast mind High Toe yes 1 Sunny mild High THe No 1a ONewea a Hok Norma fase yes —Wetlesslty _iihether the day is guiteble fox ploying gall__iven tre feature or tne” day cles Phe Column_sepresent these feahures the ymas kfodividual entries “mepare sen: TE We tase fire ow OF the dataact We Can obéemve (sig Not _Suitalle fue ploying golf if the outlook is indy. We Consndernc rat nese pike Retain ae ‘That is if the fervpersafure 's hot , it does nob mewmitten a4: 3 e (yx) = PCS) PCa) pcx) 3 The Neriale y is the class Naviavle , Wich wepsesents SE ip is suitable to Ploy gol ow nor laiven Me Conditons. Nomatole XK wsepresent Me panamorens | Pedtumeg 1S) aby Page 38X is Siven as, Kos (ky Mp2 Xn) Verse X,_, % _ %Xn_wepresents the teatuses, i-e- Hrey Can be mapped to outlook, temperature, lumidity and ' - windy . by Subenituting for x and extpanding, wtring, Fhe coin cule We get : ie TPCS}m.- Xa) = POGTY) Poly) --- PUxnly) POD) | P(x) PURe) - -17 PCN) - | j ead looking ar the dataset and Gwostitute em into te equation - | Gr ail entries, in re dataset, the denominator dees nor Onange 1 it wemain Sraht Tw i Can be yémoved & a propostional: ky Can be intwpduced: a plyix.. Xn) & PO) Myer PCxi ly) me cape, the Claas variate | Pra baJailite. 4 qe Greg roo PG) Th, POAly) se, Hinawy we are ler with the task 6F caltatine ply) and pCxily) . please Note Hrat ply) is alsa called Class PyDvaloilihy and_p(xily) 16 caued Canditional Proloalallity page 34(Eamin|ese We need to tind p(xil¥s) he ean xi in X and Yj in} All Hrese Calculations have been shown in Hre taloles \oeloud. Proploalotlity of play QolF Guittnas p ( ploy ger = Yes) > Sliq = 0-64 P (ploy oir > no) = Sly 5 0-26 OUTLOOK Yeo No HumidiFy Yes NO Sunny 3)g Hs High 31a UIs a overcass 414 0 Nowma} 619 Vs woiny 219 3/5; Temperatuse yes NO Windy yes. No Hol us 2s Geer mild 4g 2s Tove a3 alo Cold 319 ss false. 619. 4s So in a 2 fs_eaoh xiin XandYsinY manually In Hae tables a3 80 So.d_we_axe done Wits oun pe Computations and the Wassities [6 weady. \or us fost it on a ned Set of feature todoy=( Sunny, Hor, Nowmat ,-Farse), So pyobaleility OF Playing @oiF ts diven: toy PC4ea| todo) = Plsunny Jourlook}es) PCHot temperature | yes) + PChowmal Humidity | yes) P( No wind l¥es) - plyes) Pfrodeyy Pase 4oand Prnloatel lity fo nok ploy get is given lo. Plnol fodoy) = PCEunny Outlook’) Ne) pC Het Femperaturre | Ne) _PLNowma}, Humidity Ine) P(No wind |e)» P(N) L P Choday ) Since plteday’) is Common in boHs pyploaeilities 1 We con Se ee ee ee J4nve_pltodaq) Gna Cind Prmportionel pyoatcilites os: PiGyes ee eee 2. ae sonst land. P(e] todo) 2 ee Gar _9.00887 S 2 Aes 4, Now —_Grnce _p(Yes.| bndaw) P(No|todoy) = | v ~ Into A _probateili Plyes [today > ve ooog 16. 82228 002116 + 0:00457 and __p (No! Fedey) = 0.004US 7 gil So. A716 O.0UUlb+ 0.00457 Since “plses| today) 2 p (No |tedon’) So, Prediction that olf (ould we. ployed is ” ea” pase 4[Camfin] Page. Beane le - Estimate Conditiona) paobaloitities of each attrioubes { telor, leas, height, smelly? or tne spesiad Classes mn] using the dala qiven in the fasoie Nesing these probaloilies essimate Hre probabilily Value Ser i = Free 622, Height stall Gmely_= No ) No Colo legs Height Sney— Speaies 1 Wuite RY Shot Ne ™ 2 Grreen 2 tau No ™m 3 Green 3 Slot yes "1 4 Ww te 3 Shot yes M $ Geren 2 shot No 4 6 White 2 Ta No H ai wwite ie NO 4 & Wire 2 yes H Page 42.(Camiin|rase Supp Neabinr -Maphineicinina no salen) Jfahen’ Syelia A en's eee eee are | machine; algerthm —ia_te_find hey pemplan [spate (atin pumas 06 fenute) Haat ltsncly_ laure, Hae daha points. Lounaing, 2. 6 a Page» > at o JX °° o , Support esate . OO veobr- a - Seppo Sti B ‘ ‘ Bs 2. BSC S=_ monimum ‘a \ a a morgin o 2G x H Te Sepectohe twa Claeses of dale paints theme are many I : lee : : 7 Hppepiont in is a plane page 43[Camiinfrose __| hpenplane ane_desieson boundames that help Claseity fhe data paints -. Data points f 7 Loe He _ Wy peaplane Can be_aHyiubed to ieferent classes. _—§ Aloo, He dimension _or hyperplane depends upon Hae Nunwleed — of featwres, TE Hie ourveer of input features is 2 Hen Wj pemmplane, i5 Justa Wine Te Hae nunmoes of input feature is 13then Hae hyperplane becomes a 2p plane, |__Sepewating_Hne_tlasses With a large tanga minimises __ A_‘minimum generalization enter’, means Hat When | @_loound_on the _€npeaked.genemali2atien erston, ned Ctamples drrive foe clastdfication He chance of |. making aa_ertten in Hae predictian leased on Hae learned __ OE OIOIMMM Since oR oe?[Camilin]rase Delebing the euppas vers Change he position pf the hypermiane + Longe —Tongin To fwition + Tn_logittic_ mreqwesstsi00_._W anae — Rinebion and deni es sie ETA ies reece i @ take Hre output of linear funcHon and jf Hat output is qaeatem Haan |, be identify at iit, one —__—_(\aas_and_ié - identify ir With another | clase. | ine : and -! —__ lin 8vM we abmin Me 21,17) Cf, tea) = f0 Te yw Fu) D1 l-Y"S(x) etse. e (uy, FOO) 2 (i ye feo), t tne Cat is 0 je the pmediched Natye and tre aotua) Value | —_|ame_of Gare _Gign. TP they axe nek, we tren Calewtate tue | os value. page 4s— J[Comiin] Page We alse _add_ 4 repulanization ponameter to Hie Cost ction. The Objestive Of the regulamizetion Pamamebex —__—_ 1s_to_toalance _Hae_mongin_mantimization and loss. AFter addi 9) G>_Hhe_smequlanizaton parameter, the (ose __ es Sunckions looks as_\ SeiicegaliaecaacGe: a Ge, ming A Iai + 2 C14 %} >), . i re fe Now Hot Wwe have the loss function, we take portal A ictunaatepeie aeateaeente. saneuelarmigl —| 2 Aha = O >We 1 AW naz yi ia 2 f ye Zaywy\2 a ( oie Lu wy > L Awe * d spa 7+ AW Ue, else = We W- & (2dw) pane 46sum mmonuimires the “momen Wey _learming @ Suitotole decision Yeoundory | decision suntace | Seperating ~ nypurpione. t Support ——, = eplimal Wyrenplane, wealer bin” tga 2 bleh 4 os ’o y “6 a ee ra oP A 7 0 > | aa tnoryin. L *T Xe a are rad a Weil Ve Se. thal c (weet to) Le oF Closed Xe (Wath) £-1, vx oF dese o Beeb 20 A. We ton Cleaty see Mat SYM tries bo mouimise. Ma the momins 2 Hwis cailed 'Mantimum Margin ClassiFien! b. Ghe Suppem veokars vit thaveeaaatly citer (et 1} Cc. The eneme negative Values ane forthe (Green chin Points _the letter it io Sor classification. cl Tho mone positive Hae values Ane Gor the Red data poittes Hho better jt is tor ClassificaHon rageKemer in SVM: Lingor \germe) » A sur Which is Used bo Classify data Which are Vineatly Seperatole is Called linear EVM. In omer Wors . a lineav EVM Seardres Lor H{pemplane With He _mantimum manyin. Tws is Why a \inear Svm is often termed a6 Maruimal margin Classities (mmc). Of - 18 amlboitory, let us Conssic ini wil Houtes Ay apd Ar a6 X = (kK, ,X2), Where x) and x, ore values: OF atyioutes Aid Ay vespectively fos x. SquaHon of a pase _n20 ate Gn bela Wo thy x1 +W2 x2 = Wheme Ws. wat 16% one-time fantante dafling Hae Jape_aod intesceph pf tne line Any—point Sing aleave Suh a hyperplane satisfies an HX) + mae Bo —Similatly any poi oe el ee epsilon ste Xa) An & Pl: i adie ion of aS: pan ee, Ss sian oe TE Can _loe Nii i it {kis nor easy i 1 i I Greater Haan 3.= Ty faok, Euclidean equation of a hyperplane io a” is : Vy ky Wo Xp. - +Wm%km 2b Where, 4j'5 dee the wea} numoess and b is a Teal Congstanh (cated Hne intercept dich Can be positive | negative) _ = “Tn_matsix_form_.a_hypemplane thus can ee nepresented af MWEX, +b =O Whee, \al= Didsyido Wm] nd X= [xy ) Xa ---%m] and bw is G@ tea) Constant. [to be evaluated given a training Set D- by ake FoF 4 a AY ler us Congncten that a hae dimensional taining Seb Consisting — hwo classes + and — a8 Shown in Fiquire. Guppose bi and be Gre Hwo decision bounderies above £ below a hyperplane , vespectively. fon gic 0 points X4 and X_ a% Shovon fo Figure. ons _x we the decisi doy the equator Ca Joe Written 06 _ Mix, te = Kk Where kK>O Biewlarty fore any point X— lotated below fhe decigion boundary J ne equation 16 wxitb =k! Where k! <0 Page 49iz Thus if. ue label au +’ are a& Class \atoe} + and JQ —'$ a6 dass jae! — , hen We Can predict he Class late} 4 for any te6h daka x as lors YS {2 “TE WX+b 90 7 TF WxX4+b 20 Su pose. Xi and Ay Gee poo points on He decision ; looundaries lo, and lor eopectively _ Thus My.) tle = 5 WW Xe tle = =) | 6° “Ag (CX) = X2+) = This Sepresentatise a dot(-) Produat of hoo Verlows Wand (x,-x2) Thus faking Magnitude of Hrese Vectons. the equation obbuined is L d= 2a Vw Where {wil = Jwit+Wit.-.4wit ia an m-dimentional apace, inain fais A a I eh a) as gi ener = eee ning cleirive “Audon ey C*, bt) mow 2 : i Mal) Wyi Congriden C no. Jane OF enfor We wit) not change hyperplane es an ge Qo* mi How many ene ors 2 ire of the enor, To nonhe 4 1 2uF ae nA Sn ® Re hg? oR RK Saal poms Sas e2 » | AEE a rabid Rea Sept maqyin 8 MinchseitleshinComlin|race ST * Siete linearly not Sepemade)e _____ foe re we dio a Webr fir line “4 of error ee Wiil_le_mene. > Sur _kemel perfoomarce transfosmaton from lowes dimensional Space to higher dimensional) Apace- e—_| KX Yo 32 FO) a Cal ix me] = (xP xe] fe we (ite +1) Xa [xix xt | Kee Kaci Yor Bio Reha Mr res 7" Ney \ x * * *% peat a ¥ Aue ¥ 2 a PF mR eS Naot ha ag. % RSH shows \ * I[Camiin| Pe | Gawseria p Radial Bassig Funckion (BE) || foo Knowledge aloout the. data. Zils qenema) puspose Kemnel, kised When these is no Equatioo_is; kext, Xs) zh (= xi GI) eap. arnold Kernel equation is. 2) KC a) = tooh (Katy +e) | fms and Cons associated with SYM. | poms S Th Works ‘Teally WE With a Cleas Mongin oF Seperahon — Th is effestive in Wgh dimensional Spaves > Th i¢ erfective in Cases Whewe Hre numoes Of dimensions is Qveater Haan He Numer of Samples. — Th Uses a Subset of braining Points in Hedecisteg Hiachen (canted Gruppo Neaborss) $01 it 16 atso memory efficient. Cong = Tt doesn'} pesFosm Well When we have lange data set because the sequitred training ‘Hme is higher. — Th also doesn't perform Wel, when Me data Set has move Noise ie. farjet lasses awe. overlapping = Sy4 ‘+ di: Pmvide pPevbatoiity Cxtimates, trese ae Calculated wting an _enpensive five-fold Crpss velidaHon. - Ep is included jn the avdleted Sve metved of [pytinen Scilib - leas) Ailernasyclassmate! oive Bojer Gemple, = A No Glot legs, Her guh SmedLy Species \ wWite 3 Shoth yee 14 2 id 3 m 2 ” a H 2 2 2 ig s data given ia Hae balol (sn. Hoeee pao) LIreY. i ‘ity Ne fic ned iagtance (color = qzeen , toy i Parolotdoi Lis, OF Species pC rm) = w/g > 0-S7 PCH) 4) Zoawe OMSE Clos m1 ie legs, rik 4 aie 2a ae 3 314 oly Geen Hy Va 2 Vay Wy Height Gmeuty om oH Shor 4 214 yee Vy Tau Yu 4a No Yu 3)yclassmate] Pe | ee ub beer ik po a New instance of featumes Coclot = Green legs = 2 Height = Tau , srmelly=No) So Probalail ih of 1) 16 given ly PCa) xP (Green tcolor /m) p( 2 1e4s/m) P (taut Height /19) | pCwo émetyt Loa) pcm) Pim) oy 2 tN ce 9 0084 - 4 4 4G 4 gM 2 @ PCH) A 4 4 2 3 4 & I gO OMEE | sow) fince PCR) + PCH) =) These numbesd Can be Convetted into a probability ls taking gun equal fo | (Nommotizahion) P(r) = 0-076 _ PCA)” = onan a PCH) PCP) 1S of H Cleas-