Stern School of Business 44 West 4th Street Suite 8185 New York NY 10012 narchaksternnyuedu Vahab S Mirrokni Google Research 76 9th Ave New York NY 10011 mirroknigooglecom S Muthukrishnan Google Research 76 9th Ave New York NY 10011 muthugooglecom A ID: 34752
Download Pdf The PPT/PDF document "Mining Advertiserspecic User Behavior Us..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
theonlinesalesaftertheinitialexposuretoanonlinead,aswellasliftinotherimportantonlinebehaviors,suchasthebrandsitevisitationandthetrademarksearches.Theadvertisersseektounderstandtheimpactoftheiradnotjustontheimmediateclickorconversion,butthelikelihoodoftheeventualconversioninthelongtermandotherlongtermeects.Userstakespecictrajectoriesintermsofthesearchqueriestheyposeandthewebsitestheybrowse,andthisaectsthesequenceofadstheysee;con-versely,thesequenceofadstheyseeaectstheirsearchandbrowsingbehavior.Thisinterdependenceresultsinstruc-turalpatternsinusers'behavior;advertisersneednewtoolsandconceptsbeyondsimpleaggregates(liketheCTRandtheCR)tounderstandthem.Whatarethesystematicwaystohelptheadvertisersrea-sonaboutstructuralcorrelationsinthedata?Inthispaper,wetakeanadvertiser-centricdataminingapproach.Westartwithdatathatisdirectlypertinenttotheadvertiser'scampaign,thatis,usertrajectoriesthatinvolvedadsfromthecampaignofthatadvertiser,includingadimpressions,clicksandconversions.Suchdatacanusuallybereportedtotheadvertiser,provideditisaggregatedandanonymizedappropriately.Next,webuildadataminingframeworkthatcanhelpadvertisersidentifystructuralpatternsinthisdata.Ourcontributionsareasfollows.Weproposeagraphicalmodelbasedapproach.Weformulategraphsfromthedatacalledadgraphstocap-tureco-occurrencesofeventsadjacenttoeachotherinusers'trajectories.Then,weintroduceavarietyofad-factors,whereeveryadfactorisascoringrulefornodesintheadgraph:thesearedesignedtocaptureimpactofadnodesoneventualconversion.Forexample,weintroduceadfactorsbasedonrandomwalks,that,foreveryevent,calculatethelongtermprobabilitythatacertainrandomwalkinvolvingthateventwouldeven-tuallyleadtoconversion.Ourpaperpresentshighlyef-cientalgorithmsforconstructingadgraphsandcom-putingalladfactorsweintroduce.AllalgorithmswereimplementedusingMapReduce[11]parallelprogram-mingmodel.Usingdatafromthesponsoredsearchcampaignsofeightdierentadvertisers,westudyvariousadgraphsandadfactors.Wevalidatetheadgraphmodelsbyshowingtheirstatisticalttothedata.Also,weshowinterestingempiricalpropertiesoftheadfactorsthatprovideinsightsintouserbehaviorwithrespecttobrandvsnon-brandads.Moreover,weshowvariousnaturaldataminingqueriesonadgraphsandadfactorsthatmaybeofindependentinteresttoadvertisers.Finally,usingadfactors,weshowhowtoecientlyprunetheadgraphtolocalizeanddepictthein uenceofanypar-ticularadbyitssmallneighborhoodintheadgraph.Ourapproachworksbytransformingthedatasetofusers'trajectoriesintographs.Thisisachievedbypoolingdatafromdierentusers.Asaresult,adgraphsloseinformationaboutspecicusersortheirtrajectories,andonlyencodeaggregateinformation.Also,becauseofthewayadgraphspooldataonlybasedonadjacentevents,theyencodecertainindependenceassumptionaboutuserbehavioroverpathsofmultipleedges.Wecarefullystudystatisticaltofadgraphmodelstothedatatoensurethatthisassumptionisreason-able.Poolingthiswayalsoresultsinsignicantcompression.Forlargeadvertisersinsponsoredsearch,thenumberofdif-ferentqueriestheadvertisercanbematchedwithisoftenintensofthousands,andthenumberofviewinguserscanbeinmillions.Incontrast,adgraphshavemoremanageablesizes.Ourapproachgeneralizestodierentsettings.Whileinthispaperweapplyittostudyuserconversionsinthespon-soredsearchdata,itcanequallybeusedtostudytrajectoriesinsponsoredsearchthatleadtoclicksonly,orconversionswhenbothsponsoredsearchandcontent-baseddisplayaddataisavailable,etc.2.RELATEDWORKThereareempiricalstudiesshowingthatthesponsoredsearchadvertising,aswellasthedisplayadvertising,canhaveasignicantnumberoftheindirecteectssuchasbuildingthebrandawarenessandliftinthecross-channelconversions[18,20].Thepriorresearchattemptedtomea-sureimpactofanadontheuserconversionbythenumberofindependentpathsfromtheadtotheconversionevent[5].Ourworkhereisasubstantialgeneralizationofthepriorresearch,asweapplymoreadvancedPageRank-basedmea-suresforanalyzingpathwaystotheuserconversionand,moreimportantly,introduceagenericadgraph/adfactorbasedframeworkforreasoningaboutthestructuralpropertiesoftheusers'behaviors.Miningpatternsintheuserbehaviorisnotanovelidea.Marketbasketanalysisusingassociationrulemining[2]isapopulartoolusedbyretailerstodiscoveractionablebusinessintelligencefromuserleveltransactiondata.Userbehaviordatahasnumerousapplications,including,butnotlimitedto,improvingthewebsearchranking[1],frauddetection[12]andpersonalizationoftheusersearchexperience[25].Inthispaper,weminetheuserbehaviorfromtheperspectiveofanadvertiserrunningasponsoredsearchcampaign.Thereareseveralstandardminingtechniquesthatcanbeappliedtotheuser-leveladvertisingdatasuchasminingforthefrequentitemsets[2]orthefrequentepisodesinse-quencesofuseractions[22].However,thesedonotcapturethestructuralcorrelationsinthedatasuchastheimpactofmultiplepathsbetweeneventsthatourworkcaptures.Adif-ferentapproachthatattemptstocapturestructuralcorrela-tionsisminingthefrequentsubstructuresingraphdata[16].Duetodatapoolinginourgraphconstruction,patternsthatweidentifymayormaynothavehighsupport,becausesomepatternsmaycombinebehaviorofmultipleusers.Also,wearenotinterestedsimplyinfrequentpatternsbutinpatternsthatindicateahighlikelihoodofthedesignatedaction:userconversion.Further,ourapproachofusingthesteadystateprobabilitiesofdierentrandomwalkstocapturestructuralcorrelationsdiersfromthepriorgraphminingtechniques.Thisapproachiswidelyusedinotherareasincludingwebsearch[24,4],personalizedvideorecommendations[6]andusersignaturesincommunicationgraphs[9].Ourresearchiscloselyrelatedtotheproblemofmodelingusersearchsessions,forwhichawidenumberofsolutionshavebeenproposedintheliterature,includingadvancedla-tentstatemodelssuchasvl-HMM[7]andMarkovmodelsthattakeintoaccounttransitiontimebetweenuser'sac-tions[14].Weintentionallyrefrainedfromusingcomplexgenerativemodelsfortheunderlyingdataduetoseveralreasons.Atrst,generativemodelsoftenrequireindivid-ualusersessionsfortraining,while,duetoprivacyreasons, 5.4VisitForcomparisonpurposes,wedenethevisitadfactorwhichrepresentstherandomwalk(norestart)visitprobabilityofanodefromthe\begin"node.Thisadfactorisnotrelatedtotheconversionnodebutsimplycapturesthelikelihoodoftheeventhappeninginauserconversionpath.Formally,theadfactorisdenedasarowvectorhit(b;)solvinghit(b;)=hit(b;)^M+eb:SimilartoPropositon1,onecanshowthatlim!0@ppr(b;) @=hit(b;):5.5MarginalIncrease(Passthrough)WeintroducetheMI(j)adfactor,whichrepresentsthemarginalincreaseintheprobabilityofhittingconversionfromthe\begin"nodebifweincreasetheweightofalledgesoutgoingfromthenodejby",allocatedinproportiontothecurrentedgeweights.WerstcomputeMI(j)inthecon-textofPersonalizedPageRankwitharestartprobability.Let@ppr(b;c) @wjibethemarginalin uenceofanedgeweightwjionthePPRofthe\begin"nodefortheconversionnode.TheadfactorMI(j)canbeproxiedbythefollowingsum-mation(notethatsensitivitywithrespecttoeachoutgoingedgeisweightedproportionallytothecurrentedgeweight).MI(j)=Xiwji@ppr(b;c) @wji;Werstobservethat,inthesettingofrandomwalkswithrestart,@ppr(b;c) @wijcanbecomputedbyusingthefollowingProposition:Proposition2.Themarginalin uenceofanedgeweightonthePPRofthesinkcwithrestartatthesourcenodebisgivenbyaproductofthePPRofthestartnodeoftheedge(i)withrestartatthesourcenodebandthePPRofthesinknodecwithrestartattheendnodeoftheedge(j):@ppr(b;c) @wij=1 ppr(b;i)ppr(j;c):Usingthisproposition,wederiveasimpleclosed-formfor-mulaforMI(j).Proposition3.Themarginaleectofincreasingtheweightoftheoutgoingedgesofanodej,MI(j),isequalto:MI(j)=ppr(b;j)ppr(j;c) :(1)Theproofsarelefttotheappendix.Anadvantageofthisclosed-formformulaisthatitletsusapplyfastalgorithmsforcomputingtheMIadfactor.Theabovepropositionalsoimpliesthat,astendstozero,theMIadfactorcanbecomputedasfollows:Proposition4.Themarginaleectofincreasingtheweightoftheoutgoingedgesfromanodejonthehittingprobabilityofconversion,MI(j),isequalto:MI(j)=lim!0MI(j) =hit(b;j)hit(j;c):(2)Table3:Correlationmatrixforadfactors.=0:05 PPREv.Conv.LastAdVisitPass PPR1.00000.97330.7285-0.06570.1321 Ev.Conv.0.97331.00000.7064-0.06140.1370 LastAd0.72850.70641.0000-0.05420.0991 Visit-0.0657-0.0614-0.05421.00000.8006 Pass0.13210.13700.09910.80061.0000 TheaboveobservationimpliesthattheMIadfactoristhesameasthepassthroughadfactorwhichisthechangeinhit(b;c)ifweremovethenodejfromthegraphandredirectallincomingedgestothisnodetothe\null"node,orformally,Pass(i)=hit(b;i)hit(i;c):Wecallthecorrespondingadfactorthe\passthrough"orPassadfactor,sinceitcapturesthevalueofrandomwalkspassingthroughtheevaluatednode.5.6TheRemovalEffectInthispart,weintroducetheadfactorRE(i)foreachnodeiwhichisdenedasthechangeintheprobabilityofhittingconversionstartingfromthe\begin"nodebifwere-movenodeifromthegraph.Intuitively,thisadfactorcap-turesthechangeintheprobabilityofreachingconversionifweremoveanodei,ortheincomingedgesofnodei.Simi-lartotheMIadfactor,wedenetheREadfactorinthecontextofarandomwalkwithrestartprobabilityasRE(i)=Xjwji@ppr(b;c) @wji;UsingProposition2,wecanderiveasimilarclosed-formformulaforRE(i)whichisshowntobeequaltoPass(i).Proposition5.TheREandREadfactorscanbecom-putedasfollows:RE(i)=ppr(b;i)ppr(i;c) :(3)RE(i)=lim!0RE(i) =hit(b;i)hit(i;c):(4)Theproofsarelefttotheappendix.ThePassadfactorisaproxyforbothMIandREadfactors,i.e.,MI(i)=RE(i)=Pass(i).Asaresult,weonlyreportempiricalresultsforthepassthrough(Pass)adfactor,andthiswillimplythesameresultsfortheMIandREadfactors.5.7EfcientAlgorithmsComputationaleciencyiscrucialforsuccessfulapplica-tionofanydataminingalgorithmtotherealworldadvertis-ingdata.Fortunately,alladfactorsintroducedintheprevi-ousSectioncanbeecientlycomputedonlargescalegraphsusingparallelmachines.ThekeyistheobservationthatthePageRankcontributionvectors(cprppr(;c))canbee-cientlycomputedusingalocalalgorithmwhichadaptivelyexaminesonlyasmallportionoftheinputgraphnearaspec-iedvertex[4].Wehaveimplementedthealgorithmof[4]inthedistributedcomputingenvironmentofMapReduce[11]usingthePregelframework[21].InthePregelframework,everynodeinthegraphcanperformitslocalcomputationandthenodesinteractwitheachotherbyaperiodicex-changeofmessages.Thecprcomputationcanbeachievedbyasimplelocalalgorithm(Algorithm1),inwhicheverynodehasstateconsistingoftwovariables:thecurrentlyac-cumulatedcprandtheresidualvalueobtainedfromother Figure3:ROCcurveforbrandpredictionforclicks(top)andimpressions(bottom) Figure4:ROCcurveforclickpredictionforbrandqueries(top)andnonbrandqueries(bottom)Figure4showsthatforbrandqueriestheLastAdadfac-torandthePPRadfactorareexcellentsignalsoftheclickattribute;fornon-brandqueriestheypredictwellforrecallofuptoapproximately0.6,afterwhichtheprecision-recallcurvebecomes at(untiltheeventualjumpto(1.0,1.0)).Thisweirdbehaviorcanbeattributedtothefactthat,inourdataset,weobserveanumberofraregeneric(non-brandspecic)queries,forwhichclickontheaddoesnotleadtoanyconversion.Weemphasizethatsuchqueriesarerare,thereforeusingthePassadfactorltersthemoutandthePassadfactorshowsexcellentpredictivepowerforthenon-brandqueriesinFigure4.Next,wecomparethePPRandtheLastAdadfactorsbyusingadierencebetweentheadrankinbothmodelsasapredictorforbrand/non-brand,click/impressionand Figure5:ROCcurvefor(PPRrank-lastadrank)predictor Figure6:PPRbehaviorasafunctionofdistancetoconversion Figure7:ROCcurveforbrandpredictionbydier-entmodelsbroad/exactattributes.ResultsinFigure5suggestthatthedierenceisagoodpredictorofthebrandattribute;manualinspectionshowsthatthisisbecausethePPRad-factorranksbrandnodesevenhigherthantheLastAdad-factor.Ontheotherhand,theROCcurvesforclickandphrase(broad)/exactmatchpredictionareclosetodiagonalsuggestingthatthereisnosystematicdierenceinrankingsofclicks/impressionsandbroad/exactmatchnodesbybothmethods.Next,weplotthePPRbehaviorasafunctionofthedis-tancetoconversioninFigure6.Thegurewasconstructedusingthedatafromthe\BackwardMarkov"model,inwhicheveryuserqueryisrepresentedbymultiplenodesinthegraph,dependingonitsdistancefromtheconversionnode.Thus,foreveryqueryq,onecancalculatelog(ppr(qatdistanced;c))log(ppr(qatdistance1;c))9:Figure6showsaverageoftheseresultsfordierentgroups.NotethatforclicksthePPRvaluesatdistancetwoandmorefromconversionaresignicantlysmallerthanthePPRval-uesatdistanceone,suggestingthatiftheuserclicksontheadanddoesnotconvertimmediately,thelikelihoodtocon-vertinthefuturegoesdown.Consistentwithourpriorob-servations,thePPRbehaviorforbrandandnon-brandclicksissimilar.Ontheotherhand,fornon-brandimpressionsthePPRgrows(oratleastdoesn'tdecrease)withdistancetoconversion,indicatingthatexposuretotheadvertiseradonnon-brandqueriescanhavealong-termpositivecorrelationwiththelikelihoodoftheusertoconvert.Forbrandim-pressions,thePPRdecreaseswithdistancetoconversion,althoughnotasstronglyasforclicks,againsuggestingthat,iftheusersearchesforabrandanddoesnotconvertsoonenough,thelikelihoodofconvertinginthefuturegoesdown.Finally,Figure7investigateshowdierentgraphstruc-turesaecttherelationshipbetweenthePPRadfactorandthebrandattributeofthenode.Weconsiderthreemodels:\SimpleMarkov",\BackwardMarkov"and\RSMarkov1". 9Thelogrepresentationwasprimarilychosentoreducesen-sitivitytooutliers. For\SimpleMarkov"modelwesimplyplottheROCcurveusingthePPRadfactorasapredictorforbrand.For\Back-wardMarkov"modelweusetheobservation(fromFigure6)thatthePPRofbrandnodesdecayswithdistancefromcon-versionwhilethePPRofnonbrandnodesgrowswithdis-tancefromconversion,thusweuse10Xd=2(log(ppr(qatdistanced;c))log(ppr(qatdistance1;c)))asthepredictor.Forthe\RSMarkov1"model,wewouldexpectsimilarbehaviortoholdandthePPRofthebrandnodesinthe\search"state(beforetheuserclickedonanyads)tobelowerthanthatinthe\interested"state(aftertheuserclickedonsomead),whilethePPRofthenonbrandnodesinthe\interested"statetobehigherthanthatisthe\search"state;thus,weconstructthepredictoraslog(ppr(qinsearchstate;c))log(ppr(qininterestedstate;c)):Figure7showsthatthe\RSMarkov1"modelhasthebestpredictivepowerforthebrandattribute,conrmingthein-tuitionthatuserclickshavedierentvalueinbrandandnon-brandcontexts.7.DATAMININGTOOLSProducingintuitiveandconcisereportsfromhugeamountsofdataisthedesiredgoalforadvertisers.Inthissection,wediscussvariousdataminingtoolsthatcanbedevelopedusingouradfactorsdenedonthegraphmodelsintroducedinthepaper.Identifyingthetopkadswiththelargestadfactor.Thesimplesttypeofreportonecanthinkofisidentifyingthetopkadswiththelargestadfactorsinaparticulargraphmodel.Thevalueofsuchreportdependsontheparticularadfactorandtheunderlyinggraphmodelused.Forinstance,inSection6,weshowthatthePPRadfactor(inthe\SimpleMarkov"graph)hashighbutnotperfectcorrelationwiththebrandattributeoftheadimpression.IdentifyingthetopkimpressionswiththePPRadfactorcanthusbeinterpretedasidentifyingthetopkimpressions(userqueries)withthehighest\branding"impact:manyoftheseuserquerieswillinfacthavetheadvertisernameorthebrandnameinthem,butsomewillnotandthuscanbethoughtofas\shadowbrands".Duetoprivacyissues,wecannotshowanactualexampleofsuchoutput.AnotherinterestingexampleofthetopkoutputisthetopkimpressionsforthePassadfactor;theycanbethoughtofasthetopkuserqueriesassociatedwiththelargestrevenuetotheadvertiserifonetakesintoaccountthelong-termcorrelationsinthegraph.Interest-ingly,fromFigure3,weknowthattheseareuncorrelatedwiththebrandattribute.Identifyingadswithsignicantlong-termeectsnottakenintoaccountbytheLastAdmodel.AdvertiserstraditionallyrelyontheLastAdreports(click-throughandconversionrates)toevaluateadeectiveness.Analternativereportwesuggestwouldbetolookforadsthatarerankedhigherwithadfactorstakingintoaccountthelong-termcorrelations(likePPR)thanwiththeLastAdadfactor.Onesuchcriteriaforrankingisthedierencebe-tweenthePPRadfactornoderankandtheLastAdadfactornoderank(propertiesofthiscriteriaareshowninFigure5).AnothercriteriaistolookforimpressionsthateitherhavethePPRgrowingwiththedistancefromconversion(\Back-wardMarkov"model)orhavethePPRinthe\search"statehigherthaninthe\interested"state(\RSMarkov1"model).AsFigure7shows,thesecondapproachismorebiasedto-wardsthebrandimpressionsthantherstone.Identifyingthetopkadswiththemaximummarginalincrease(MI)adfactor.Apopularadvertisingobjectiveistomaximizetheexpectednumberofuserconversionsgivenacertainbudgetconstraint.Inarandomwalkmodel,wecanrestateitasmaximizingtheprobabilityofhittingtheconversionnodestartingfromthe\begin"node.Aheuristicwaytoidentifythevaluablenodestoinvestonistoexamineasmallincreaseonthebidofaquery(orkeyword)whichwillhavethemaximumef-fectontheprobabilityofhittingtheconversionnode.Thissmallincreaseonthebidresultsinasmallincreaseintheweightofincomingedgesforthisquery.Thus,inordertoidentifyqueriesforwhichtheincreaseintheirbidresultsinthemaximummarginalincreaseintheprobabilityofhittingconversion,wecanlookforthetopkadswiththemaximumMIadfactor,andtheadvertisermayconsiderincreasingthebidforquerieswithlargeMI.10ExplainingtheadfactorvalueTheadfactorassignedtoanyparticularnodeinthegraph,mustbeeasilyexplainedbyalocalstructurearoundthisnode.Forinstance,forthePPR(PageRankcontribution)adfactorofanodeu,onecanalwaysconsiderthetop-mneighbornodesthroughwhichtheexplainednodeucon-tributesthelargestfractionofthePageRanktothecon-versionnode.Formally,letubethepermutationofallneighborsofthenodeuthatarrangestheminthedecreas-ingorderofwuvppr(v;c).Wedeneu(m)asthesetoftherstmnodesinthepermutation.Thesenodescanbethoughtofasthemostlikelynextactionsoftheuseraftertheexplainedaction,assumingthattheuserisgoingtoconvert.ItisstraightforwardtoextendthepushbackAlgorithm1tokeeptrackofthetopmcontributorsforeverynode,thusonecanecientlygeneratecorrespondingreports.Thereportsarebestrepresentedasgraphplots,constructedstartingataseednodesandgoinguptoadistancedinagraphofthetop-mneighborsofeverynode.Duetoprivacyreasonsaswellasthespacelimit,weomitanexampleofsuchplot.8.CONCLUSIONSOnlineadvertisershaveaccesstoaggregatestatisticssuchastheclickthroughrate,theconversionrateortheliftoftheiradcampaigns,butseektounderstandmoresophisti-catedcorrelationsinusers'trajectoriesofadsseenandusers'actions.Hereweintroduceanalternativedataminingtech-niqueforaggregatinguser-leveladvertisingdata.Wedenevariousadgraphstomodelthedatapertinenttothead-vertiserandproposeseveraladfactorsbasedonstationaryprobabilitiesofsuitablerandomwalkstoquantifytheim-pactofadevents.Weshowanecdotalevidenceforadfactorsanddescribetheirpotentialdataminingapplications.Ourresearchintroducesnovelprimitivesforminingadvertiser-specicdataforadeects.Adfactorswedenearesuccinct,easytocomputeandcapturemanystructuralpropertiesof 10Weemphasizethatthisisonlyaheuristic.Anypracticalbiddingstrategyshouldalsotakeintoaccountthecurrentbiddinglandscape(howmuchdotheadvertiserandthecom-petitorsbid)foraparticularkeyword.