Faster Training of Neural Networks for Recommender Systems.pdf

《Faster Training of Neural Networks for Recommender Systems.pdf》由会员上传分享，免费在线阅读，更多相关内容在应用文档-天天文库。

1、FasterTrainingofNeuralNetworksforRecommenderSystemsbyWendyKogelAThesisSubmittedtotheFacultyoftheWORCESTERPOLYTECHNICINSTITUTEinpartialfulllmentoftherequirementsfortheDegreeofMasterofScienceinComputerSciencebyMay2002APPROVED:ProfessorCarolinaRuiz,ThesisAdvis

2、orProfessorSergioA.Alvarez(BostonCollege),ThesisAdvisorProfessorLeeA.Becker,ThesisReaderProfessorMichaHofri,HeadofDepartmentAbstractInthisprojectweinvestigatetheuseofarticialneuralnetworks(ANNs)asthecorepredictionfunctionofarecommendersystem.Inthepast,resea

3、rchconcernedwithrec-ommendersystemsthatuseANNshavemainlyconcentratedonusingcollaborative-basedinformation.Welookattheeectsofaddingcontent-basedinformationandhowalteringthetopologyofthenetworkitselfaectstheaccuracyoftherecommendationsgenerated.Inparticular,

4、weinvestigateamixtureofexpertstopology.WecreatetwoexpertclustersinthehiddenlayeroftheANN,oneforcontent-baseddataandanotherforcollaborative-baseddata.Thisgreatlyreducesthenumberofconnectionsbetweentheinputandhiddenlayers.Ourexperimentalevaluationshowsthatthis

5、newarchitectureproducesthesameaccuracyofrecommendationasthefullyconnectedcongurationwithalargedecreaseintheamountoftimeittakestotrainthenetwork.Thisdecreaseintimeisagreatadvantagebecauseoftheneedforrecommendersystemstoproviderealtimeresultstotheuser.Acknowl

6、edgmentsIwouldliketothankProf.CarolinaRuizandProf.SergioAlvarezforalltheirhelpthroughoutthewholeofmymaster'sthesiswork.Prof.LeeA.Beckerfortakingthetimetoreadandcritiquemythesis.Prof.MichaelCiaraldiforhishelpinguringoutWeka.TheKnowledgeDiscoveryandDataMining

7、ResearchGroup(KDDRG)atWPIfortheirhelpwithWekaandinputonmythesispresentation.iContents1Introduction12BackgroundandRelatedWork52.1NeuralNetworks.................................52.2ErrorBack-Propagation.............................72.3RecommenderSystems.......

8、.......................92.3.1Content-basedandCollaborativeRecommendation..........102.4Weka........................................113ANewANNArchitectureforRecommenderSystems133.1NetworkArchitec