Scribd
Upload
13 views

ML - Linear Regression

ML

Uploaded by

Amrita Roy
Copyright
© © 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
13 views

ML - Linear Regression

ML

Uploaded by

Amrita Roy
Copyright
© © 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
You are on page 1/ 21
Bock awk > i Supervised Learning Machine ed Onsupervised leaxnin P Uc Supervised —> Reaéession Slusieeid A Classification a Cocctpect is known) = = —] —] = —] ; ; Regeession hs Lineoe Seqeesston = 2. Polunomior Reggession a SVR = 4. Decision Téee 5 5. Random Fosest = = = 3 a 3 a a é. aves 2 TKN Classi cation: —> 1. Logistic Re : 2. svV™M 3. Decision Tee &. Random FPosest Ss Naive Bay es 6. Kn asession 1 bbeie Gnepecetsed \eovening: Coutpet ts unknown) Sigetaeb stinDBs can ¢-KMeang') i my dose, Hierarchical, SilhoueHe Sessing yt Saipeten ised ssatining : t Fa: tx ulpdependent Fealuxe dependent : Oegcec Eoxperierce alaxep) features SiNeGbp me fee ete Ts Sok Poy a Tok ( lp = continuous —> Bc Ase sto0 Psoblem Ea Z Independent Feat. 7? dependent No:of play Wo-of Study Fass! Foil Feat. hes | hés Olp = cartegosicos = classificakon peoblem Plight Psice Péedichion —> Reqecssiee paoblem Peo Fiee Podest) -> classification PPP PP eee RK KAKA}. paoblem » Peedict Ate Qunitty Ident —> MEE Seaee ‘ Peobtem Rain Péediction — classification Reoblem + Buying dayo +! “classi ¢ eg ay fa patton classificakon eo : péoblens i Wetes Independent featuces is basically inpat peateuces : hes neat Ke Tee ae Cone indepen dent vine and one oo featuxe) eq aim: +o céeate A model, aban dakes inpat ; as iets and peedict wetgnt dataset: helgbt. aa Eq: Aim: Basedon the no- of 6 datapoint Cosigi Pans n arta poin Coats inal) = peedicted daka point "difference between cea) points and P&edicted poinis is called éesiduols OS Brod. © Based on the teaining dataset, it finds the be it Wine tn such a way that *he un ©} between eal Points md” péeclicte J sheoc be mintmeam lest, ase need to undexstand oe ase cacating a Siraight line 2 Be. best €i+ line ie at Best best -ft line ts hothin Slope but a equakon of straight line 4 Y= mate me he Bot Bix allovce intexcept. eq”: of hola) = Go + 61x = line hola) = Go + Gre ie slope Intexcept : [inteccept)} when 2-0, thé line meeting the eynaads), that axrHoulore point ts sivas “known as intercept. _ with the unit movement to the x-aads what is the movement in the ta ao a Go and 41 best kit line win be 9 By change V_ change fF ie } 4s 00,01: changing the values, | ee # win hem AKS line be” inersldn Fe ~ a alter, some erations , we will eta best kit line which is AL. known as taining Of the model. + +. \ we need to minimize Abts + Rest eécoes usin equaton called as VY cost function: \Cost Per Eemalion > foé eas ey colewtations Cderivatives) Geet < Com C «> \20/ fal eres ee —~——"pkedlicted actual: mean SQUaHe eo &- (one ofthe eost Lamction + we need to minimige cost functton +o act the best fit oe. dividing oH m 40 att she vee ag © Cue Ee ChotorS sue ‘ 2m je ‘Aegan the values ok 60 and 61. Letus consider Go=0 Ne Card = 1H lets assume t=! holed = OCH) = 1 polo): Oot Or eo Teda\ E Goto, ri) wo i) = La+o +o] HOS Sf os- “+ C2994 (US- i A GtEadient descent cuxve anis point Objective: owe main aim is to come newe ae minima: ae cannot change © value manually: Ahexe shoatda be some mechanism +o elt @~vatue pols ane lobal minima. To overcome wis, we Use Comvexcqence AN a Convexgence al wares ~- int 4H \ COCs oo ames a of Os ee ue: F step se! as small steps Wik? a Sadia cee — Moweane sa heir : 80H COD ED Repeat anki! gonvergence - > lesening Sate R= BNE OS Ve slope: Beeeecigbt side of the line kacin downwards d sc oe NC? Slope estabt side of the tine focing apwoxds Sf 4Vve’ slope ve! slope ! 6 = 6) ~xCve) = aj+e eans We axe incxeasing 6}. => —$—y — = - — | 3 a 3 z s 2 & ebeaxning, Rate It decides the speed oF the ota Ik |e ts vexy sma h VA widl dake moée time iia Fo keach te minima. KE ts oe lovee | Ae oul yarn breve and thee, And won't each to ce minima- aX, shouldbe atcund o:co) fos Smaller, eee ae \ sie Rs Df —Senverge nee atgosi thro 2 a 360,61) = Sa ce =) -cost function cotta INS actual o& a Weal point m= eee data point difference between cost funclion and loss fanction- 6084 function: we Lind exog fosall the points es and take eee OL ihe Loss function: we find ewroe fo& obsexved polots and if we find exnvoé fos atl the points then tt is loss fanction- all or} loss funchHion:= Chaco” -¢ ! i a péedicled ~ " Nalue. actual VANE ea Terr 5° isi sin point we need to find ite “los 10 _achievc qleeas mintnas dexivative w.64 40 Oo je9 ‘S a (6) y\~ 3(60,6:) = 2 rot hota’ a4) 4 . Be ares Nae Wola) = 00+ O10 x are 2.9 3S fonenlgl’S Se (Co0+6.03'-y') x| t= F es a C+a0-45 aA a 2 ) ’ = 2fUeod-gh) = ceorern) -¢') —— iz) m Zil|- denivotive w.64. 1, \=1 2, 3(60.9) 22. ie gto 5 ie e 2. 5 (004 ora)-y a Te ene ot ce G81. F Creend-y')" jay Or a 61 = ot S Cheta)' - i) td m ial } } \ ) ) ) ) Mi et TT ELS Types of cost function! @ mse: mean “squaxe Sic cae. nt = Meee MAE: Mean absolate exno® @ RMSE: Soot Mean squaxe dependent catuxe Encpenlence ~ Nivea tay Nath Neate (y-§)? Ctaxns)® Coy anit mor —7 hme complerities pe x ___iwoxeaned: 7 be : xe we don't do Beg joa dependent peats 9. we dep y-9y Squared —> exioe > penotiaed Imonrea Mean! Absolute Bus PEAS L no MAG as. DP) icy ‘ rm a | ad i Adv Za @ Robust +0 outers. @ s+ will also be inthe. same unit: ___disadv: a Sonvexa erins tisually takes moee dime. opttimigation ts a comptenc Aas: fe) time segsuming. ive cannot be feet Reot mean Square Ewroe CAmse) mses |S CRON | = ee eee yf Ags disadv: = Se Creceiabie > ~ Not &obust to - Unit Semain same. outers - Habe Less Punction The Hubes, toss Of hers the best of both wodlds by balanelog the MSE and MAE togeted , My 7 2 > Cy- Ftv) ‘anki Ugly, cov) = S]y-Feo|- 45> — otherwise = It Says! fee !oss values less thandelta, use +he a MSE, fog toss values ee t+han'delta, J use +he mAe- ysis the MAE tes longer 10SS values mitigates the weight that we put on outliers so that we sHil a a well- £oundeol model. At the same time, We use the MSE & Ahe smaller Joss Values +0 maintain o dail furnckon encore the centre et has +h Sth foot ok Maagnityuing *ne ‘oss atues AS cc as oo One q lates Ahan. loss fos dhose data points dies below) , data Epa OTS 1 os Be hub Mabon. Noss any Piste You jee! tnat jou need a balance aes ivin oudiexs some scl but noFtco “much. Mow can we etea\s ik a model ts good 68& eee quadkatic fer notion down-ewciqhts them ocus the woe on the oo ersco& ee pecfosmance mebacs + cd} measuxes Ane sspemancs ° BS tnesmodel: f : Best fibline Rt ( eR squaked = wees 7s xy + Z ¢ y ~~ Ss \ Na ee. SS awccagey PKA # SSres = Sumo Squaxe - oh Residuals ie f An SStotol = Sum of Square average a AD _low value R- squaxed See te CED) a ~\2 =U 2 C4) ~ oat valuc GF onexage hs EE = = 9 Ssmalivalue j Biqqes value J-s smal : Naluc " R- squaxed = 0:85 = 89%. accurate = oS = 184% accuxaie If R squared Is (NE? Aen the model ts not ashing ts en | R? 3 ~ve | a = tH exe Cyy-G) > Cyy~J) — R? will be Deqo NG) Is * | |R2et | - > | a ae 1 ____s djusted R-squaxed eat No digect cones avon pecerk | fo [ \10-65 . [\sataxy forencter } __ House Jocation' — bedxcoms ia e ] Peice | - Befose, when ae sige Of house was p&esent as an inde pendent feature to get the psice and has some R- SU Naluc. But with addition of city location = ot inckease- ayrex that with addition % ss path bcdxoom , Pisce, Me Value aes inckease , bus there is wet wh die eo! Connela ce, the Value gender and px aN: eae = & featuxes added adj. R?_ Independent Yeates Se. Size of house 63% Pay 734, location _ Baie P=2 € SQ. 06 df bedrooms a6, P23 Sor. Grended easy. P24 K a frie Do Aisect comaelaltors wie -+ psice Fe Neer Pah tncxease, but this shouldn't happen: Yo solve this pot lem we use Ac sc eo , Adjusted R2 = 1~ C1-R2)Cr0-1) zi ce rs yen besenyaiails ns M2 No-op data poinss Pe Ne-of Independent features x Adjusted Ris thebeS+-meteics to evaluate the model: | os (By ao wna 0,0 ae ioe ee reef age arid ee er J RAINS, { woo del Veain- dadaseh - ceca ae |!e0e datapoints laa Testing | O30) moded best iting C Bias and Variance) gven|iHing and ndentsing, 9 } C ‘ 100 ; [Teaining dataset \ ain yh een Ss A\ Fein | | Vatidete J / Sco 200 V eS tyain the model bu pekpaxomelee tupning +ne model. [iene | Gener alige Model Téain Sata — NE geod Accuracy ens Test Data —5 Vere i Gebd Accuracy Cast.) + rer SI O4X aim is to get qoud dest and drain Neooey Téain data — very qed Accuracy Soy. Liew mr + est Data —s5 Mere Laced Accnacr usy. D low Voueiance] Veditng eo TRAin —> & ; » —> Very good mecuNacr C40¥-) L lew Bias] BES is Bad Bec eeeaee Cae eg High Vaxiance | Qndex}ittin TRAIN —> mode accwrac is low. Chih bias] Test —> modet accuracy is leva / nian: pa a CJ ( ook hi 7) Varian ocd * WE Can solve this issue b perfeémin hy pete ter tunni ’ ‘ Bienen unning os by ineReas "4 0-4

You might also like