资源描述:
《Web-Scale Bayesian Click-Through Rate Prediction for Sponsored Search》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、Web-ScaleBayesianClick-ThroughRatePredictionforSponsoredSearchAdvertisinginMicrosoft’sBingSearchEngineThoreGraepelTHOREG@MICROSOFT.COMJoaquinQuiñoneroCandelaJOAQUINC@MICROSOFT.COMThomasBorchertTBORCHER@MICROSOFT.COMRalfHerbrichRHERB@MICROSOFT.COMMicrosoftResearchLtd.,7JJThomsonAvenue,CambridgeCB3
2、0FB,UKAbstractofCTRpredictionisabsolutelycrucialtoSponsoredSearchadvertisingbecauseitimpactsuserexperience,WedescribeanewBayesianclick-throughrateprofitabilityofadvertisingandsearchenginerevenue.(CTR)predictionalgorithmusedforSponsoredRecognisingtheimportanceofCTRestimationforonlineSearchinMicros
3、oft’sBingsearchengine.Theadvertising,managementatBing/adCenterdecidedtorunalgorithmisbasedonaprobitregressionmodelacompetitiontoenticepeopleacrossthecompanytothatmapsdiscreteorreal-valuedinputfeaturestodevelopthemostaccurateandscalableCTRpredictor.probabilities.ItmaintainsGaussianbeliefsoverTheal
4、gorithmdescribedinthispublicationtiedforfirstweightsofthemodelandperformsGaussianplaceinthefirstcompetitionandwonthesubsequentonlineupdatesderivedfromapproximatecompetitionbasedonpredictionaccuracy.Asamessagepassing.Scalabilityofthealgorithmisconsequence,itwaschosentoreplaceBing’spreviousensuredt
5、hroughaprincipledweightpruningCTRpredictionalgorithm,atransitionthatwascompletedprocedureandanapproximateparallelinthesummerof2009.implementation.WediscussthechallengesarisingfromevaluatingandtuningthepredictorThepapermakesthreemajorcontributions.First,itaspartofthecomplexsystemofsponsoreddescrib
6、estheSponsoredSearchapplicationscenario,thesearchwherethepredictionsmadebythekeyroleofCTRpredictioningeneral,andtheparticularalgorithmdecideaboutfuturetrainingsampleconstraintsderivedfromthetask,includingaccuracy,composition.Finally,weshowexperimentalcalibration,scalability,dynamics,andexploratio
7、n.resultsfromtheproductionsystemandcompareSecond,itdescribesanewBayesianonlinelearningtoacalibratedNaïveBayesalgorithm.algorithmforbinaryprediction,subsequentlyreferredtoasadPredictor.Thealgorithmisbasedonageneralised