Latent Semantic Models for Collaborative Filtering

《Latent Semantic Models for Collaborative Filtering》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、LatentSemanticModelsforCollaborativeFilteringTHOMASHOFMANNBrownUniversityCollaborativefilteringaimsatlearningpredictivemodelsofuserpreferences,interestsorbehaviorfromcommunitydata,thatis,adatabaseofavailableuserpreferences.Inthisarticle,wedescribeanewfamilyofmodel-bas

2、edalgorithmsdesignedforthistask.Thesealgorithmsrelyonastatisticalmodellingtechniquethatintroduceslatentclassvariablesinamixturemodelsettingtodiscoverusercommunitiesandprototypicalinterestprofiles.Weinvestigateseveralvariationstodealwithdiscreteandcontinuousresponsevar

3、iablesaswellaswithdifferentobjectivefunctions.Themainadvantagesofthistechniqueoverstandardmemory-basedmethodsarehigheraccuracy,constanttimeprediction,andanexplicitandcompactmodelrepresentation.Thelattercanalsobeusedtomineforusercommunitites.Theexperimentalevaluations

4、howsthatsubstantialimprovementsinaccucracyoverexistingmethodsandpublishedresultscanbeobtained.CategoriesandSubjectDescriptors:H.3.3[InformationStorageandRetrieval]:InformationSearchandRetrieval—informationfiltering;I.5.3[PatternRecognition]:Clustering—algorithmsGenera

5、lTerms:Collaborativefiltering,recommendersystems,machinelearning,mixturemodels,latentsemanticanalysis1.INTRODUCTIONContent-basedfilteringandretrievalbuildsonthefundamentalassumptionthatusersareabletoformulatequeriesthatexpresstheirinterestsorinfor-mationneedsintermofin

6、trinsicfeaturesoftheitemssought.Insomecases,however,itmaybedifficulttoidentifysuitabledescriptorssuchaskeywords,topics,genres,etc.thatcanbeusedtoaccuratelydescribeinterests.Yetinothercases,forexample,inelectroniccommerce,usersmaybeunawareoratleastinattentiveoftheirint

7、erest.Inbothcases,onewouldliketopredictuserprefer-encesandrecommenditemswithoutrequiringtheusertoexplicitlyformulateaquery.Collaborativefilteringisatechnologythatiscomplementraytocontent-basedfilteringandthataimsatlearningpredictivemodelsofuserpreferences,Thisworkwassp

8、onsoredbyNSF-ITRgrants,awardnumbersIIS-0085836andIIS-0085940.Author’saddress:DepartmentofComputerScience,Box1910,BrownUniversity,Pr