矩阵分解技术应用到推荐系统培训讲学.ppt

《矩阵分解技术应用到推荐系统培训讲学.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、矩阵分解技术应用到推荐系统1、PaperBackgroundFriday,September17,202121.YehudaKoren,YahooResearch2.RobertBellandChrisVolinsky,AT&TLabs-Research3.PaperpublishedbytheIEEEComputerSocietyinAugust20094.AuthorwonthegrandNetflixPrizeCompetitioninSeptember20092、IntroductionModernconsumersareinundatedwithchoices

2、.MoreretailorhavebecomeinterestedinRS,whichanalyzepatternsofuserinterestinproductstopridepersonalizedrecommendationsthatsuitauser'staste.NetflixandAmazon.comhavemadeRSasalientpartoftheirwebsites.Particularlyuserfulforentainmentproductssuchasmovies,music,andTVshows.3、RecommenderSystemStra

3、tegiesContentFilteringCollaborativeFiltering1.Neighborhoodmethodsuser-orienteditem-oriented2.LatentFatorModelFriday,September17,202143.1、ContentsFilteringCreateaprofileforeachuserorproducttocharacterizeitsnature.Needtogatherexternalinformation.Aknownsuccessfulrealizationofcontentfilter

4、ingistheMusicGenomeProject,whichisusedfortheInternetradioservicePandora.com.Friday,September17,202153.2、CollaborativeFilteringAnalyzerelationshipsbetweenusersandinterdep-enciesamongproductstoidentifynewuser-itemas-Socitions.Disadvantages:coldstartTwoprimaryareas:neighborhoodmethodsuser-o

5、rienteditem-orientedLatentfactormodelsFriday,September17,202163.2.1、NeighborhoodmethodsCenteredoncomputingtherelationshipsbetweenitemsorusers.Theitem-orientedapproachevaluatesauser’spreferenceforanitembasedonratingsof“neighboring”itemsbythesameuser.Theuser-orientedapproachidentifieslike-

6、mindeduserswhocancomplementeachother’sratings.Friday,September17,20217Example:3.2.2、LatentFactorModelsFindfeaturesthatdescribethecharacteristicsofratedobjects.Itemcharacteristicsanduserpreferencesaredescribedwithnumericalfactorvalues.Assumption:Ratingscanbeinferredfromamodelputtogetherfr

7、omasmallernumberofparameters.Friday,September17,202194、MatrixFactorizationMethodsCharacterizebothitemsandusersbyvectorsoffactorsinferredfromitemratingpatterns.RSrelyondifferenttypesofinputdata.Strength:incorporationofadditionalinformation,implicitfeedback.Implicitfeedback