搜索
PersonalizedNewsRecommendationBasedonClickBehaviorJiahuiLiu,PeterDolan,ElinRønbyPedersenGoogleInc.1600AmphitheatreParkway,MountainView,CA94043,USA{jiahui,peterdolan,elinp}@google.comABSTRACTOnlinenewsreadinghasbecomeverypopularasthewebprovidesaccesstonewsarticlesfrommillionsofsourcesaroundtheworld.Akeychallengeofnewswebsitesistohelpusersfindthearticlesthatareinterestingtoread.Inthispaper,wepresentourresearchondevelopingpersonalizednewsrecommendationsysteminGoogleNews.Foruserswhoareloggedinandhaveexplicitlyenabledwebhistory,therecommendationsystembuildsprofilesofusers’newsinterestsbasedontheirpastclickbehavior.Tounderstandhowusers’newsinterestschangeovertime,wefirstconductedalarge-scaleanalysisofanonymizedGoogleNewsusersclicklogs.Basedontheloganalysis,wedevelopedaBayesianframeworkforpredictingusers’currentnewsinterestsfromtheactivitiesofthatparticularuserandthenewstrendsdemonstratedintheactivityofallusers.Wecombinethecontent-basedrecommendationmechanismwhichuseslearneduserprofileswithanexistingcollaborativefilteringmechanismtogeneratepersonalizednewsrecommendations.ThehybridrecommendersystemwasdeployedinGoogleNews.ExperimentsonthelivetrafficofGoogleNewswebsitedemonstratedthatthehybridmethodimprovesthequalityofnewsrecommendationandincreasestraffictothesite.AuthorKeywordsPersonalization,usermodeling,newstrend.ACMClassificationKeywordsH.3.3.InformationSearchandRetrieval:Informationfiltering.GeneralTermsAlgorithms,Design,ExperimentationINTRODUCTIONNewsreadinghaschangedwiththeadvanceoftheWorldWideWeb,fromthetraditionalmodelofnewsconsumptionviaphysicalnewspapersubscriptiontoaccesstothousandsofsourcesviatheinternet.Newsaggregationwebsites,likeGoogleNewsandYahoo!News,collectnewsfromvarioussourcesandprovideanaggregateviewofnewsfromaroundtheworld.Acriticalproblemwithnewsservicewebsitesisthatthevolumesofarticlescanbeoverwhelmingtotheusers.Thechallengeistohelpusersfindnewsarticlesthatareinterestingtoread.Content-basedrecommendationisatechnologyinresponsetothischallengeofinformationoverloadingeneral.Basedonaprofileofuserinterestsandpreferences,systemsrecommenditemsthatmaybeofinterestorvaluetotheuser.Content-basedmethodsplaysacentralroleinrecommendersystems,asitisabletorecommendinformationthathasnotbeenratedbeforeandaccommodatestheindividualdifferencesbetweenusers[3,8].Thistechniquehasbeenappliedinvariousdomains,suchasemail[16],news[4,5,20],andwebsearch[15,18].Inthedomainofnews,thistechnologyparticularlyaimsataggregatingnewsarticlesaccordingtouserinterestsandcreatinga“personalnewspaper”foreachuser.Anaccurateprofileofusers'currentinterestsiscriticalforthesuccessofcontent-basedrecommendationsystems.Somesystems[1,19]requireuserstomanuallycreateandupdateprofiles.Thisapproachplacesanextraburdenonusers,somethingveryfewarewillingtotakeon.Instead,systemscanconstructprofilesautomaticallyfromusers'interactionwiththesystem.Inthispaper,wedescribeourresearchondevelopingapersonalizednewsrecommendationsystembasedonprofileslearnedfromuseractivityinGoogleNews.TheGoogleNewswebsite,availableat’newsinterestschangeovertimeandhoweffectiveitwouldbetousethepastuseractivitiestopredicttheirfuturebehavior.Permissiontomakedigitalorhardcopiesofallorpartofthisworkforpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmadeordistributedforprofitorcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthefirstpage.Tocopyotherwise,orrepublish,topostonserversortoredistributetolists,requirespriorspecificpermissionand/orafee.IUI’10,February7–10,2010,HongKong,China.Copyright2010ACM978-1-60558-515-4/10/02...$10.00.31Tounderstandthisissue,weconductedalargescaleloganalysisofGoogleNewsuserstomeasurethestabilityofusers’newsinterests.Wefoundthattheirinterestsdovaryovertimebutfollowtheaggregatetrendofnewsevents.Basedonthesefindings,wedevelopaBayesianmodeltopredictthenewsinterestsofanindividualuserfromtheactivitiesofthatparticularuserandthenewstrenddemonstratedinactivitiesofagroupofusers.Torecommendnewsstoriestousers,thesystemtakesintoaccountofthegenuineinterestsofindividualusersandthecurrentnewstrend.Therefore,theuserwillreceivenewstailoredtoherinterestswithoutmissingtheimportantnewsevents,evenwhenthoseeventsdonotstrictlymatchtheuser’sparticularinterests.Wecombinedthecontent-basedmethodwiththecollaborativefilteringmethodpreviouslydevelopedforGoogleNews[7]togeneratepersonalizedrecommendationsfornewsaccess.Thehybridmethodwasevaluatedinaliveexperiment:asubsetofthelivetrafficatGoogleNewsusedthehybridmethod;theresultshowedsignificantimprovementovertheexistingcollaborativefilteringmethod.Theexperimentonlivetrafficalsorevealedanumberofinterest