Scribd
Upload
2 views61 pages

Classification

This document discusses the concept of Classification in Machine Learning, highlighting its importance in various fields such as Healthcare and Banking. It contrasts Classification with Regression, explaining how Classification predicts categories rather than numerical values, and introduces Logistic Regression as a method for making classification decisions. The document also covers the Sigmoid Function, decision thresholds, and the practical applications of classification models.

Uploaded by

mahmoudhaplas
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
2 views61 pages

Classification

This document discusses the concept of Classification in Machine Learning, highlighting its importance in various fields such as Healthcare and Banking. It contrasts Classification with Regression, explaining how Classification predicts categories rather than numerical values, and introduces Logistic Regression as a method for making classification decisions. The document also covers the Sigmoid Function, decision thresholds, and the practical applications of classification models.

Uploaded by

mahmoudhaplas
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
| Day 2 - Classification: How Machines Make Decisions - © welcome SARC aa on Sea Inthe previous lesson, we learned about Regression, where Machine Learning predicts a numerical value. ‘eg2_ Classification is one of the most widely used @%6 Machine Learning techniques. (@ Tod, we wil study another important typeof Machine © ttis used in; Healthcare, Banking, E-commerce, Learning called Cee cation: © Cybersecurity, Social Media, Self-driving Cars, © _Inseed of predicting numbers, Gasification predits categories or labels. isn Alster tne oss erry Seamed XY ; @ Craseifcaton models. Peal-worid Data |g, Byte endef tod estan youl understand how | mechinesdecde between diferent cles, €B_Moctin Learning Mode Examples of Machine Decisions KEYNOTE |s this email Spam or Not Spam? p ij 's this transaction Fraud or Normal? Today's lesson Wil the customer Leave or Stay? focuses on how Spam / Not Spam |s this image a Cat or a Dog? machines make ——+ | Fraud / Normal Sie neia i sous nema maces ae sia Cat beg hobs or il Cloneiction probloyel bocauee tarpeds nano] Decision Postve / Negative | | ¢ the answer belongs to predefined category. Learning Objectives : ZZ S By the end of this lesson, you should be able to: = nif Thede Objaciives matter Explain the difference between Regression and i session: Classification: S Cassteaton cotSetieine |) cosine ine &__ Identity whether a problem is a Regression or ooe% Classification task © Explain why Linear Regression cannot solve Classification problems. 5 Describe how Logistic Regression works. Regression: Classification: Coninious Preston: categorie Presiione ai ferstand the purpose of the Sigmoid Function. = a Und 1d the f the Sigmoid Fi uf Interpret prediction probabilities. Fenn Geen iond anor Clsieation precics Understanding Classiteation @® Een the idea ofa Decision Boundary. cep ay tontparrsepae - aii a fang» colton bulking meligert ystems. % Pesce strengths and limitations of Logistic eee: a tons ln Imoces are uiton te same Basie you lism oy Recognize common Al applications that use os Classification. a tocisnie sicwow @ 201 @ Lomecron OD reseesson BD Finecon G coosng ne conect lor cri for accurate resus | Agenda ‘Today's lecture is organized into two main sections. Learning Journey @ Part 1 — Regression vs Classification s a Pe : vt i i = & Whatis Machine Leaming Prediction? Rae eet Regression Pees ae = sh Two Types of Prediction Classification Problems * \ai Regression vs Classification = & Real-World Classification Examples oC = @ When Should We Use Classification? Logistic Probability Neil Regression Decision @ Part 2 — Logistic Regression Casifcation jiagram = ® Why Linear Regression Cannot Solve Classification = JC Logistic Regression = of3 Advantages and Key Note + J/ Sigmoid Function Limitations De net memnoriseicorttvmnet . ii OA i % Probability | Applications ee eens ee «= |X Decision Threshold = # Worked Example algorithm exists and when to use it. ® ch Decision Boundary = Q Review | What is Machine Learning Prediction? @ Mmenmesttemi @ Setornncseen os 2 =o GSS i =e e- eee | Regression vs Classification Regreston or Clarifeston, > Regression and Classfeation are both @ servic Examete Superted Machin ening goth. > Theman dferece ithe type of cuput o as they rode. How much none wil acon pendent moth? Tse Ragrenion baci th serine @ Models predicts: © on = ae ‘Will the customer leave the bank? oral ie ‘hs Cleon bcrn tn somerbone |\@ ‘to one of two categories, 2 cot eg | @ te | tmmorin | tromam | | Qe —— > On hint nk Dia Sc rng cone Tempe Dhan Dntecton ster txts mera rete ee Th can trie he Mah Lig set Qn © © WovALnEsD | crenioae snk oe val Reareion Catenion | ote ip fees | et en 2 Ee | “The target tells you whather you need = = ( Pree number enna Pres a etegory (ees) Classification is one of the most common tasks in Artificial Intelligence. 22, Many systems around us make decisions using Classification models. Application Possible Classes + BM Email Filtering: Spam / Not Spam * EB Disease Detection: Positive / Negative © @q Fraud Detection: Fraud / Normal * > Customer Churn: Leave / Stay * (©) Face Recognition: Person A / Person B © Wh Image Recognition: Cat / Dog * 1% Loan Approval: Approve / Reject eel So Popular? Examples of Al systems using Classification: M1 Gmail detects Spam emais Many real-world problems require ‘making a decision rather than predicting a numerical value. Computers use Classification to | | ff Banks detect raudlent make these decisions transactions. automaticaly Hospitals classty medical images. Social media pltforms classy A. irpprorate content. KEY NOTE ithe answer blogs to 3 ites ruber of eategeres, Cis ition iste he conection Gy Setting cars recone traffic signs and pedestrian, AcTIVITY “Think of ne aplication on your phone that probably uses Classifica Explain what categories it precicts eeu ™ | When Should We Use Classification? choosing the correct Machine Learning aig stats with understancing the target vera. [Big Ue Cissteation when the opt belongs to 7a one of several predetined clases. ‘Ask yursel these questions belore bul @ model © Ami preditng abl tea ot rambo? © oes tte aut btn oad categorie? © smi eya wo moke adection? ite ans is Yes, losiston shay the comreet approach. Al Connection Profesional ita Scents aye ety 8 ° KEY NOTE The tage orale determines te alert Pret hose pce Pearson Pret monty ses Rewresion acon hadnt cnaeaon Pred empoye sary eaesin _—___ “tees er YY =e aa : ACTIVITY / DISCUSSION QUESTION For ech rb bow dee wert is Regrton or Csinton © Pesannarows unpace 0 onenteseater scan yaa © Date reneancee on © resets satin rte rreteok | Cth einen Cor Dos | Why Linear Regression Cannot Solve Classification LUnearRagresion es a = ae | Why Linear Regression Cannot Solve Classification Inthe previous lesson wo lsrned tht Linear Regression predicts numerical values. een Attest, we may thee we can use it for every prdltion i cases ee corse eee problem inding Clasfication. longi scan > However, this snot the right chee. > Uneor Regression hes no upper or lower se canna etry produce vai proba @ We rotten ner Regretson press The bis ral 2enue @® Aiconnection ara oaae hen podeting en ee aaa For example, it may predict GY House prices, > 82% chance of disease. © sows 1 Son chnen eee cna set 32] (08) (27 | [154 ‘ Temperatures Unear Regression cannot provide these ut they do nt make sree fr Clasfiaton ian imogine wewantte desty (FA spam || (2) Not Spam © Kernore near Reresion predicts continuous numbers © A rection f 3.8 12 dows nt ponent ther dan > Clnsenton needs probebities end categorie © vsurtnesoe0 ray __ nner ee ee a = ae cn predenny lmnte stat ind ten 9 and I The Sigmoid Function ‘The Sigmold Function she heart of Logistic Regression, > taj eto convert any numerical value Into probably between O and 1 > This makes the output suitable for Clatiication, Intuition Tn te Smal Faces aa cone: the probity ange arbeead Sayer nea ee acpotamny nomeroeat ree = eee © rernore mics, | come | SON. eae | Understanding Probability aac a G _rrecspsortoete assent poe, GP tourney ene rr 005 tay — 020 nay GB) > wesenm aso 7 ieee 0.80 Likely 0 Example 098 Very Bel Soper nad ped: ‘This means thre la 92% probably ‘atthe sample belongs othe postive less Al Connection any Al pplationsdplaycontdence scores Impossible Uncertain Certain + Spam Detection: 96% Spam @ KEY NOTE + Medial Biagnosis: 87% Peitve ‘A proba not he fil decison, These values ae probable. tis only the modets condence | From Probability to Decision > After predicting « probability, the model must decide which class to choose, > Tomake this decision, Logistic Regression Uses @ Decision Threshold. > The most common thresholds: 0.5 @ ovcsion tute = ) (Can We Change the Threshold? Depending onthe business problem, we may use 03 0.6 0.8 Al Connection Medical systems often lower the threshold 10 avoid missing dangerous diseases. KEY NOTE = The threshold converts probebilties into decisions © Hressity205 > ramen) aati © Hrs 05 > ence oan toe oy | tet © VVALNEEDED | onion tet orem oO 05 Negative Class Probebity <5 Positive Class Probebity 20 | Decision Boundary The Decision Boundary the ne or surfs) ‘tha saparatescforen esos >> Everything on one side balongs to one cas vetting onthe other side belongs tothe othr clas ‘Simple Example 1 nage acing dre: a al In Two Dimensions + The boundary uy ee. + thigh mons becomes ne cramer comin uae Ce —— Soe ‘Al Connection + Dirt est ern hp, OQ ror SEs ‘Tram pred Cee fo points above of he boundary. | Advantages and Limitations of Logistic Regression @ se cor Meche Leng sui Lote Rees ba sents and weakness DD) crete con Qo Sinpletoundesans ye) Sansa asuite Om ar Oo==. features = Se Q never sure” Senate toleont oe Produces prtatities = ® complerimage recognition o eosnunnenesn (GS) Nitsa tetas @ Hah oninear rblens seeuncy ® Largeseae deep eaning tasks @© 1 H0Te | seme tts ten | Al Applications of Logistic Regression i coparecsseeeaanes Lagasse tel bey © ewes Bare ‘clot plein atone eperemeeet ‘ster conor ili preset, ne Wor | [@nare | ‘heed pte ah Fabio and Small Worked Example @ ‘Suppose a Logistic Regression mode! predicts the probabil that an email is Spam. Email | Predicted Probebity | Final Deckion Eiaila os seam % Email 076 spam Enaile as nor spam Emaild oz nor spam @ ere + The probabilty comes fist. + The decison comes second, e — Model calculates P(Spam) = 0.76 Step 2 0.76 > 05 Compare with Predicted ‘Threshold Step 3 | | Common Mistakes Many beginners make te following mistakes ‘when larring Logistic Regression, ‘Thinking Legis Regression i 1 Regression algcithn Belavng the mode! predicts numbers instead of probabities. Acuing every Cartiton prea «an be solved perfect. Forgetting hat changing the teshold change the predictions. Thinking higher probity aloaye ‘mean higher scursy Tips “Yap probing ses Y Chee etl bd on te ba pom © wore |e @00000 “The Model Outputs ‘Probably Ply = 11x) Output (Pretty) 7 | Changing Tveseld Changes Decaions sre pnts ihr rca) ‘sorption Lopate Regression a Clasication lori | Mini Activity - Predict the Output G _rsume the Deion Trethld i 05. Naaeks Probability <0.5 Predict Positive Predict Negative ost 2 Discussion 0.62 2 © Woicr precctons belong tothe 049 ? Positive class? 028 2 © Wien belong to the Negative class? What would happenif we changed 076 2 e es the threshold to 0.7? ‘This activity helps you understand how probabilities become decisions. @ LEARNING GOAL | Logistic Regression Review > Let's review the complete workflow of Logistic Regression. © GB) cxecircicne © F(X) catedatea near equation. il © LL mere || a) | ED) retusa OCD corinaprovanaty + E]R decision Treshotd © x © © mmm | sincaan 5 © @ a Logistic Regression is 8 Classification algorithm. It predicts probabiitios between 0 and 1. ‘The Sigmold Function ‘makes this possible. ‘Athreshald converts probabilities into final decisions. Knowledge Check Answer the folowing questions. eee @ nse mi tence batman Reena tn? —_magine you are building an Al syst Teeter crest car © vey cri tgeionahCanen palm etci? oan © waist ose he sgn Fanci @ vous rorcooe © taco prt 087 rot Regreson or Csiestot © vous not comnon Dente @ vrtss dean sense es © sae wtf agin Q crner moe | | | © creer open tat gee psn | What is K-Nearest Neighbors (KNN)? ee Ll = {rnuthemateal modal during tating. > netond store th {prediction only I New Sample © cosa co 008) 0 seca nestor | SP) ibn sonte en mot a appoors most often among ts haares eighbors I. ocittreeticrece-ronl | eprom th rer feet ght = Ri mca tp nh he sit ten {Nita the tang onl oe sone Seen sme How KNN Maes 2 Prediction = 5) feconmole Wc th lm ak th 5 dat ing ‘Slr and hsnr my dee Cintas see eamapey nee eS ms con ‘A Connection [freee concn eae crn in ey in | Sept sitters npn np ein man ele Nairbon = | a ‘Sir dts alan blige to ae ater etetetint a jy Ce Ai ‘lve et hed 4 © revnore testes ee how 2a sree pnd | How KNN Works | Ce © heme TE Q came a mee © 4° ©0000 Sco Blu appears mor oon ‘hore sample lsd as Bie, e Arce ERR Pear nearest Neighbors 1 ] set vite t steps aoc Kev Nore pale Scere te ey — step o 8 0 Pe é Os mates prec by ann ‘otoymerng antec ton > KNN decides hich neighbor ar “loset” by messing the dstance between dota points cdaon Dita, > Iecaleates the strightine distance between ‘to points. © ten ov Form Fe two pits: (4, ad CY) d= 0q— m4? +2? @ rnen » magne chou ons mp. > Tohru sng a tr han enh en ten. > Terr tn has etal e e ance Measurement in KNN © svat cxamoie + Polat A= 2,3) + Polat B= (5,7) 247 7 d= Fee = 25-5 ‘wo pas ae om exch athe ‘x dort nmdto cain mana Macnee rats prom Choosing the Value of K © rrvescnonsoenotnmet — SEIN csonpi exanole Q mm ao cea aripatein the decton. crac crater ‘choosing he wrong ve an sNersmanatonne ee ‘reduce model perfomance Myron saitingsss = sesame © oro em ‘There sno single best aie, ‘ate Scientists usualy ty several values and compare validation data, @ ror npn aipion = Ma per ip @ sencnttve | @ Neon Very sl Higher Hyperparemeter tuning ko overiting snimpotnt pot of Macie Laing Very large K ~ Higher ri of undertting eter ‘ot nyperparameter tuning. ‘The valu of K affects both accuracy and generalization. ‘Small K Sine Nise Balanced Best Performance | Advantages _ Limitations of KNN DP Giterane | Mini Activity - Predict the Class Using KNN G vores GED Tec” ® avesrions 1 @ bie © wr 2 © tue © wirerwouthappenit: | K=3 3 © es and te ners neighbors wre A © be © ee) (© med) (@ mes 5 @ hea Would the prediction change? s 5 Kes ce * Sines THM ran-stane somo fehl om Dict with your parte mone) «Wy ight charging te valet change the en? > Pred Be ag + Tesng X= aay oe ie pin yor anne LEARNING GoaL _TS®ttvty demonstrates thatthe choice of K deci affects the mode's prediction, ‘making it one ofthe mest important hyperparaetersin the KNN algorithm. | What is a Decision Tree? es © errno [@ steer tot Node eee os in kes lab wr ry gin G eee > scr O ume ae eae els row do he poi ase ov} [eons [Wer] [omer] Lea Nodes: Teel psconer se acannon dan, (AP ervinints) oe AR omni > ete mn Machine Laaring oat, ADecion Tow worn wey te sve wy can dal folow very deen Irae by he med How a Decision Tree Makes Decisions A) [A Decision Tree classifies data by moving from detopctowvenroenectownctnces” — @) sept sunatte ten noce eo ese ts (Oa © ses O se04 © 205 kt ttn | (eam t ‘Question 2 teroa pet f CE enor oe (ere ete) 1 Fina Class (Precieton) | Information Gain - How the Tree Chooses the Best Question ‘Sip ation © vor Sect emp a= 97 2. Rent: St Med hgh Cone seam | tee + eeoirnee ht 2. Rete Wel Seprted (ow Coto, {tah fortiori “Tha tre choot the quason with th highest Information Gain separation ad ler conto, | Building a Decision Tree 6 See == peace ere ee ener I - on Find Best Spit sani te nr cats all caste best st sing hiemstion ain ee Drennan tren nw © 0000 © 12097 | eto eit an ctr @ wcomccnon | Berra leninunca tones 0 | Tree Depth and Overfitting > > Tee Depthisthe numba of vlna Onno Q=-_ > deptatecr towns he me + sate @© wor ovrttngt ee ee seat eee Racer lesenees ee eee Q vvr me — Mayle + Les very dete pan + Conmamerse te taning + May ado Overting. Advantages and Limitations of Decision Trees ©} vwsion tes pop oc hy tet pt a oy ret They are @ good choice when: eo Like all Machine Learning algorithms, they have strengths and weaknesses. © Expiainabitny is important Cs ee ecicios Qe ® 9 Cea nr Qn a —— S35 Cs Many explainable A OKA) solos vse Declon Tees beaut humans can easily ‘understand thelr eseoning, ©o Aj od Decson Tes pro one fhe bet sans @ own | eee HH | Al and Business Applications of Decision Trees | What is Naive Bayes? > isobar Hein List Senate cance > Sides nora ay pin deo bay (i Saeee inp ptt @ riers % ran sors + gy romeo | Probability Intuition and Conditional Probability Naive Bayes is ult on he ea of probabity. eos en | mw ‘What is Probability? - ee ee T Q\ cose hae ce |. an. r 4 > EP aren a ea toe lsoon sen Oe al @ werescentoal Why ls This Useful? “whet the potty foe eet happen. Nive aye contins te rete ence ive tec ever ha ead happened {Seta the mow tat es ca ‘Wats th prbsbty tht an era Spam A Connection {ent te word “hee? Many Aleystame update te predictions that changes oe predton Contonal Probabity maser ha pose 6 eee | Why Does Naive Bayes Work So Well for Text? > Naive Bayes perormsexpacily wall hon working wih txt dat Examples include: a ngindinors BZA vo ct new docomerss eet | Example: Spam Email Detection © ceenecinwe negra pnb ‘Al Connection Were nrg meso ines | apes l | Advantages and Limitations of Naive Bayes ‘When Should We Use Naive Bayes? Sentiment nae G7 voyranong oe eee 5 ? A eee ee | serene % Peseeance depends nda quality Cant es ccust tan reat ite Oen | ——oo.. SS | Comparing Classification Algorithms | @ crorire neni abort > ct cteon sonar om eg an whe frat perce Taste nathan Teno algtten tata tebe, hen sling smote one © Typectaan © Dee © Neder eintity © Speed requ, © Aecrey earns CD) metem GD cont == G:) IN Connection > Pre Sit aly (2) 1» ® seein Smile nese nd conpue 5 pues bet Seng Orme om, EP) worth @ ood Ta crater ee a oes I Knowledge Check and Discussion ogF o=xx=x==» I Why Isn't One Decision Tree Enpagti? ai af ahh AAD 4 og (oser e More Rte Deion ° J mie see KEY NOTE © © Neny nei won toute ohn mae beter acon. » | What is Ensemble Learning? > Seinen nsiannine AA) ever gerertzion See ee a, 6-0-0. 0-8) How Does Random Forest Work? > Random Frets oe ofthe most popula Enombie Learning algorithms. > Iie clad a Forest because it contane LI ee eali| migra °o > Each oes using a iohty ferent subset ofthe taing data > Because every vee lars diferent patterns the oes do nt all make the sre mists, Why la Better? Instead depen on er oon, eee oer) Al Connection Canines Ioan pelos et without rina ata! KEY NOTE | Render Forests spy calecton ( ecion Tons wong ae I Voting Mechanism in Random Forest GBA, Bier il Deciion Teor make their precictons, Random Forest must chaoee one fl newer Fer Cation ti doe wing May Vein °e Sippoe fe ves aka th folovng predic, Po ein | " = mea = ta tpn ts om mes = @ Count the Vote @ 5 Not Spam = ote > eee f Bi Spam = A votes, Effective? Techenlh feist abe fe (Spam } © Hone vee makes @= bcos he rl pion. A GB norspam = 1 vot5 ‘sae, te other | tomcat pce cS ros it} tal Tees = 5 Ovum = == |=" fas rain | Spam | Advantages and Limitations of Random Forest sports because carbine the stents of mary Decision Tees ® i rescence) Sens Random Forest orm || Xd [ARAL eee eaad (meena | ‘ aceionstpeeoe Cee, gum of aa Ore cere camping eces f eee a Cpe 6 oa SS Ss 5 © veoicae —Domnrrescion J Aalainee Cop Caton aa be || = @ crore coterie lhl | Random Forest Review (ieee eines == ae eB ‘oe = t moe AAA [S| || enn mor + 1 a aT SEED inet tee tears fromm ‘Mutiple Predictions nee 6e~s8 Ne) =) + Se competes AIL) € ae =” Fe © ty vg ste oe ’ —— ee EBD vee jority Veting Final Prediction Ors eer ieee eae fa as. © sone oF most day anes (cass acrs Lsming lettre | Why Is Accuracy Not Enough? > cian a irc any een, [aaron] jena meagre eiegeresets pc aael gegen 3 {missed all Fraud Cases Detected Most Froud Cases @ Dermat 00 Tee Dane (1000 ransctow) © ter 0 (© tera 9 © rout 0 © rat 0 > memodel conc pe ett 00 wart, al retin al rion (© corey pend 5 © Conny pends? 18 vente (© Facet mal dct Sieh eee oe eel I Confusion Matrix Praicted Cle © rouse © Neptne Fae epstne ON Wyk ea OARS ee ees Tate netics prove much apr derstanding of roa perorens © severe | etcemmscerens |B ee I Accuracy 6: rere pecetage often tat ne coc Visual Summary he pee “How often the model correct?” ce my TOT om | — cea a eeara ae 92% oe terse ea eet ed ete ony pee Wen i eciracy Use? ‘eorng om ss i rumber of ops 0 So = cp eae nas = oo =e On Se = US [ Precision om ©" = Proicted Positive © Seepen th motel rn (6 Went model pds jonah Span Poste how oon coc + + Tea) Spm Peco fois onthe quaity of PEE + 2 wenomal emai. Seer CEP LEI recone 18 +20 = 90% === © spemDetecion Why is Precision Important? Gorarcaia) —[morsarestra «High Precon mere the model ™ = toe Approal eos 3 B) cents + Tek poet on a Sa aie cone OUD @ Aeomecion % > eee cetera t ean ba wy ote G = © oh recon = ever Fal Poster = Mee lable Poin Prete | Recall “Of ath cial Pov aan, how may hh sal = > ‘ca ocne on detecting ot many posto en posi. “Why ls Recall Important? igh Rca onthe mod nee sey ten pater nes Tsar ppesions where Isang patie ce dangers © rrr toca Ineong tact perc KEY NOTE ecl asnues Row many ee postive eases the mode sucessful dato .? 1+ The model erect detects 47 f ham Recall = el TP+FN 47 + 50 = 94% ‘Actual Positive Cares (ptions ao ty he Ow Ss a LLL ————] moto © vactncis Se, ast Min) qa Tome nm aeeeentieeieeeey +48 ty Hee ' 4 Trove Poot: Ac poste eaves tht wre oe etc re rae Negus Al patente tet were mid th made Recall Medical Dignosis © Focseson Focuses on igh Rca preiced postves ‘cua postves + Dancinotemsecpint | SESS + Sona ety pent may ze oe cidrae Neaeo ez fdices = Fale Postives False Negatives a= Moses prediction quality detection sity @ wiicr one tse? @ Meomecion PAC @ Foun Depend ont ot fe terntennin | aera ft Cn mica SST ||| et Fens enPrecion | Foon cat Ce ae rena) | F1 Score EE Precision and Recall. Precision x Recall + Tin mee cal th Score x Precision x Recall_ { mombines both mats noone va. Fl = 2% precision + Recall Why Use F1 Score? How F1 Score Works {© Amode shoud no hav a f © Very high Precision but very low Recall ltr aa Recall Q. | © Very high Recall but very low Precision ‘rommer U eee | © _F1score rewards models that balance both IN == ae Eee °o Preion 95% Precision 8B% Noda ay Recatl 40% Recall «87% Secs, GQ) Mcometin ea) | | Gemma | | ee Ones | ROC Curve ‘The ROC Curve helps us evaluste. f ‘Cait del eros ROC Curve Example ‘leven chin thresh, Instead of measrng performance st only oe threstoi, shows how the model behves sthe i + Tune ate ea Siva + Fn ane “True Positive Rate (Recall) a en r oo | Oz oa 06 * Seeaawenmeetee Fb Ponte Rte « puoeeacarsigetee (Spey) neroan pote © 20 carve Summarys Tout et gin anon, omg We @ vere rvesar 1 The ROC Curve helps compare multe Cleafeaton models @ Acomncten ‘ata Scintit often compare ROC Curves before seocting the final moda. [better model produces {ROC Curve thet tay over tothe toprleft corner. | ROC-AUC (Overview) > rican terrence (CO > Istadof ooking tthe whele ROG graph AUC sears pefemanes sing one number: © revues KEY NOTE I Business Interpretation of Classification Metrics Different business problems require different evaluation metrics ‘The best metric depends on the cost of making mistakes. e Important Question Pee © Which type of mistake ie more > Maing pain expensive forthe business? QF concerdetecion Recall erp leaacor a Before choosing a metric, ack De Be somoewcion © Precision aes. @ ‘Al Connection > Dena aemry Succes lpr we By, Foonesion —G@_ Recall frauen wnsnctons voted no only by Spe fecal acre beso by thot busier impact. Prat ‘ion "td Recmnanin GH Precision Farr potac 6 2 Bun rg char be sel narra BEY customercturn GB) F1Score ——_wnlemocdng coma, Cres Soon ‘business objective. {Remembers anlage irr bane fe iret cots or dent pe of ere. ‘th the real business eject. Oa + + I Choosing the Right Evaluation Metric + Thoin eal auton mae tsb ety Caton prob. + Each ore tle aeston @ quick cue a @ omomy seimrmmemmpses GS | Complete Classification Algorithm Comparison Se [ Lesson Summary EE “ny pron mt prt ee ow i ese pry 0 Padme tar ay oleae Sermon | pao Patra ong L Exit Ticket & Al Learning Roadmap __ QOk=x=r =» “S rst Ws tee ef ve: ‘Topropre forthe pnt ase Yeshua compte contr * os | — i © phsteranatemnatmen gf | @ tient Ragrnton an Cleator? Cee . Coe ft © eis cetenon de © const ee Catenion sorts ih Cin gt wip © mw © eicipe bardin cent ae GD ows Matin tearing journey fr SEEN @-@-a-@ eo ° °o Remember: Kop prating, ep bling and never stp eaing “The best Al models are bul by curious an consistent eamers he you G I Looking Ahead - Next Lesson

You might also like