Table Of Content

PersonalizedMachineLearning Every day we interact with machine learning systems offering individualized predic- tions for our entertainment, social connections, purchases, or health. These involve severalmodalitiesofdata,fromsequencesofclickstotext,images,andsocialinter- actions.Thisbookintroducescommonprinciplesandmethodsthatunderpinthedesign ofpersonalizedpredictivemodelsforavarietyofsettingsandmodalities. The book begins by revising ‘traditional’ machine learning models, focusing on howtoadaptthemtosettingsinvolvinguserdata;thenpresentstechniquesbasedon advancedprinciplessuchasmatrixfactorization,deeplearning,andgenerativemod- eling;andconcludeswithadetailedstudyoftheconsequencesandrisksofdeploying personalizedpredictivesystems. Aseriesofcasestudiesindomainsrangingfrome-commercetohealthplushands- on projects and code examples will give readers understanding and experience with large-scalereal-worlddatasetsandtheabilitytodesignmodelsandsystemsforawide rangeofapplications. julian mcauley has been a Professor at the University of California San Diego since 2014. Personalized Machine Learning is the main research area of his lab, with applications ranging from personalized recommendation to dialog, health care, andfashiondesign.Heregularlycollaborateswithindustryonthesetopics,including Amazon, Facebook, Microsoft, Salesforce, and Etsy. His work has been selected for severalawardsincludinganNSFCAREERaward,andfacultyawardsfromAmazon, Salesforce,Facebook,andQualcomm,amongothers. Personalized Machine Learning JULIAN MCAULEY UniversityofCaliforniaSanDiego UniversityPrintingHouse,CambridgeCB28BS,UnitedKingdom OneLibertyPlaza,20thFloor,NewYork,NY10006,USA 477WilliamstownRoad,PortMelbourne,VIC3207,Australia 314–321,3rdFloor,Plot3,SplendorForum,JasolaDistrictCentre,New Delhi–110025,India 103PenangRoad,#05–06/07,VisioncrestCommercial,Singapore23846 CambridgeUniversityPressispartoftheUniversityofCambridge. ItfurtherstheUniversity’smissionbydisseminatingknowledgeinthepursuitof education,learning,andresearchatthehighestinternationallevelsofexcellence. www.cambridge.org Informationonthistitle:www.cambridge.org/9781316518908 DOI:10.1017/9781009003971 ©JulianMcAuley2022 Thispublicationisincopyright.Subjecttostatutoryexception andtotheprovisionsofrelevantcollectivelicensingagreements, noreproductionofanypartmaytakeplacewithoutthewritten permissionofCambridgeUniversityPress. Firstpublished2022 AcataloguerecordforthispublicationisavailablefromtheBritishLibrary. ISBN978-1-316-51890-8Hardback CambridgeUniversityPresshasnoresponsibilityforthepersistenceoraccuracyof URLsforexternalorthird-partyinternetwebsitesreferredtointhispublication anddoesnotguaranteethatanycontentonsuchwebsitesis,orwillremain, accurateorappropriate. Contents Notation pageix 1 Introduction 1 1.1 PurposeofThisBook 2 1.2 ForLearners:WhatIsCovered,andWhatIsNot 3 1.3 ForInstructors:CourseandContentOutline 5 1.4 OnlineResources 7 1.5 AbouttheAuthor 8 1.6 PersonalizationinEverydayLife 9 1.7 TechniquesforPersonalization 12 1.8 TheEthicsandConsequencesofPersonalization 14 PARTONE MACHINELEARNINGPRIMER 17 2 RegressionandFeatureEngineering 19 2.1 LinearRegression 21 2.2 EvaluatingRegressionModels 26 2.3 FeatureEngineering 32 2.4 InterpretingtheParametersofLinearModels 40 2.5 FittingModelswithGradientDescent 42 2.6 NonlinearRegression 43 Exercises 46 Project1:TaxicabTipPrediction(Part1) 47 3 ClassificationandtheLearningPipeline 49 3.1 LogisticRegression 50 3.2 OtherClassificationTechniques 52 3.3 EvaluatingClassificationModels 54 3.4 TheLearningPipeline 62 v vi Contents 3.5 ImplementingtheLearningPipeline 72 Exercises 76 Project2:TaxicabTipPrediction(Part2) 77 PART TWO FUNDAMENTALS OF PERSONALIZED MACHINELEARNING 79 4 IntroductiontoRecommenderSystems 81 4.1 BasicSetupandProblemDefinition 82 4.2 RepresentationsforInteractionData 84 4.3 Memory-BasedApproachestoRecommendation 86 4.4 RandomWalkMethods 97 4.5 CaseStudy:Amazon.comRecommendations 100 Exercises 100 Project3:ARecommenderSystemforBooks(Part1) 102 5 Model-BasedApproachestoRecommendation 104 5.1 MatrixFactorization 106 5.2 ImplicitFeedbackandRankingModels 112 5.3 ‘User-free’Model-BasedApproaches 117 5.4 EvaluatingRecommenderSystems 121 5.5 DeepLearningforRecommendation 125 5.6 Retrieval 131 5.7 OnlineUpdates 133 5.8 Recommender Systems in Python with Surprise and Implicit 134 5.9 Beyonda‘Black-Box’ViewofRecommendation 138 5.10 HistoryandEmergingDirections 139 Exercises 141 Project4:ARecommenderSystemforBooks(Part2) 142 6 ContentandStructureinRecommenderSystems 144 6.1 TheFactorizationMachine 145 6.2 Cold-StartRecommendation 149 6.3 MultisidedRecommendation 152 6.4 Group-andSociallyAwareRecommendation 155 6.5 PriceDynamicsinRecommenderSystems 162 6.6 OtherContextualFeaturesinRecommendation 167 6.7 OnlineAdvertising 170 Exercises 174 Project5:Cold-StartRecommendationonAmazon 175 Contents vii 7 TemporalandSequentialModels 177 7.1 IntroductiontoRegressionwithTimeSeries 178 7.2 TemporalDynamicsinRecommenderSystems 180 7.3 OtherApproachestoTemporalDynamics 188 7.4 PersonalizedMarkovChains 191 7.5 CaseStudies:Markov-ChainModels forRecommendation 193 7.6 RecurrentNetworks 201 7.7 NeuralNetwork-BasedSequentialRecommenders 204 7.8 CaseStudy:PersonalizedHeart-RateModeling 210 7.9 HistoryofPersonalizedTemporalModels 211 Exercises 212 Project 6: Temporal and Sequential Dynamics in Business Reviews 213 PART THREE EMERGING DIRECTIONS IN PERSON- ALIZEDMACHINELEARNING 217 8 PersonalizedModelsofText 219 8.1 BasicsofTextModeling:TheBag-of-WordsModel 220 8.2 DistributedWordandItemRepresentations 230 8.3 PersonalizedSentimentandRecommendation 233 8.4 PersonalizedTextGeneration 237 8.5 CaseStudy:Google’sSmartReply 247 Exercises 249 Project7:PersonalizedDocumentRetrieval 250 9 PersonalizedModelsofVisualData 252 9.1 PersonalizedImageSearchandRetrieval 253 9.2 Visually Aware Recommendation and Personalized Ranking 254 9.3 CaseStudies:VisualandFashionCompatibility 257 9.4 PersonalizedGenerativeModelsofImages 267 Exercises 269 Project8:GeneratingCompatibleOutfits 271 10 TheConsequencesofPersonalizedMachineLearning 273 10.1 MeasuringDiversity 275 10.2 FilterBubbles,Diversity,andExtremification 277 10.3 DiversificationTechniques 279 10.4 ImplementingaDiverseRecommender 284 viii Contents 10.5 Case Studies on Recommendation and Consumption Diversity 286 10.6 OtherMetricsBeyondAccuracy 291 10.7 Fairness 295 10.8 CaseStudiesonGenderBiasinRecommendation 300 Exercises 303 Project9:DiverseandFairRecommendations 305 References 306 Index 322 Notation CommonMathematicalSymbols MachineLearning y vectoroflabels X matrixoffeatures x featurevectorfortheithsample i f(x) modelpredictionfortheithsample i r residual(error)associatedwiththeithprediction,r =(y −f(x)) i i i i θ vectorofmodelparameters σ sigmoidfunctionσ(x)= 1 (cid:2) 1+e−x (cid:2)x(cid:2) p-norm,(cid:2)x(cid:2) =( |x|p)1/p p p i i (cid:4) ;(cid:4) regularizers(cid:2)θ(cid:2) and(cid:2)θ(cid:2) 1 2 1 2 λ regularizationhyperparameter L;(cid:4) likelihoodandlog-likelihood UsersandItems u∈U useruinusersetU i∈ I itemiinitemsetI I setofitemsrated(orinteractedwith)byuseru u U setofuserswhohaverated(orinteractedwith)itemi i |U|;|I| numberofusersandnumberofitems Ru,i measurement (e.g., a rating) associated with an interaction betweenuseruanditemi xu,i modelestimateofthecompatibilitybetweenuseruanditemi RecommenderSystems β biastermassociatedwithuseru u β biastermassociatedwithitemi i ix