资源描述:
《Transfer Learning for Collaborative Filtering via a Rating-Matrix Generative Model》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、TransferLearningforCollaborativeFilteringviaaRating-MatrixGenerativeModelBinLilibin@fudan.edu.cnSchoolofComputerScience,FudanUniversity,Shanghai200433,ChinaQiangYangqyang@cse.ust.hkDept.ofComputerScience&Engineering,HongKongUniversityofScience&Technology
2、,HongKong,ChinaXiangyangXuexyxue@fudan.edu.cnSchoolofComputerScience,FudanUniversity,Shanghai200433,ChinaAbstractitemsbasedonacollectionoflike-mindedusers’ratingrecordsonthesamesetofitems.VariousCFmeth-Cross-domaincollaborativefilteringsolvesodshavebeenpr
3、oposedinthelastdecade.Forex-thesparsityproblembytransferringratingample,memory-basedmethods(Resnicketal.,1994;knowledgeacrossmultipledomains.InthisSarwaretal.,2001)findK-nearestneighborsbasedonpaper,weproposearating-matrixgenerativesomesimilaritymeasure.M
4、odel-basedmethods(Hof-model(RMGM)foreffectivecross-domainmann&Puzicha,1999;Pennocketal.,2000;Si&Jin,collaborativefiltering.Wefirstshowthat2003)learnprference/ratingmodelsforsimilaruserstherelatednessacrossmultipleratingmatri-(anditems).Matrixfactorizationme
5、thods(Srebro&cescanbeestablishedbyfindingasharedJaakkola,2003)findalow-rankapproximationfortheimplicitcluster-levelratingmatrix,whichisratingmatrix.Mostofthesemethodsarebasedonthenextextendedtoacluster-levelratingmodel.availableratingsinthegivenratingmatri
6、x.Thus,theConsequently,aratingmatrixofanyrelatedperformanceofthesemethodslargelydependsonthetaskcanbeviewedasdrawingasetofusersdensityofthegivenratingmatrix.anditemsfromauser-itemjointmixturemodelaswellasdrawingthecorrespondingHowever,inreal-worldrecomme
7、ndersystems,usersratingsfromthecluster-levelratingmodel.canrateaverylimitednumberofitems.Thus,theThecombinationofthesetwomodelsgivesratingmatrixisoftenextremelysparse.Asaresult,theRMGM,whichcanbeusedtofillthetheavailableratingdatathatcanbeusedforK-NNmissi
8、ngratingsforbothexistingandnewsearch,probabilisticmodeling,ormatrixfactorizationusers.AmajoradvantageofRMGMisthatareradicallyinsufficient.Thesparsityproblemhasitcansharetheknowledgebypoolingtherat-becomeamajorbottleneckformo