Collaborative-filtering-recommendation-algorithm-b

PhysicaA396(2014)66–76ContentslistsavailableatScienceDirectPhysicaAjournalhomepage:∗,ShiquanSun,CheLiuSystemsEngineeringInstitute,SchoolofElectronicandInformationEngineering,Xi’anJiaotongUniversity,Xi’an,ShaanxiProvince,Chinahighlights•Weuseitemdomainfeaturestoconstructuserpreferencemodels.•WecombineuserpreferencemodelswithCFforpersonalizedrecommendation.•Weusethemulti-attributedecisionmakingmethodtocalculatetheuserpreference.•Ourmethodintegratesdomaincharacteristicsintoapersonalizedrecommendation.•Ourmethodaidstodetectingtheimplicitrelationships(missedbyCF)amongusers.articleinfoArticlehistory:Received14December2012Receivedinrevisedform1October2013Availableonline21November2013Keywords:PersonalizedrecommendationCollaborativefilteringUserpreferencemodelItemdomainfeatureabstractPersonalizedrecommendationisaneffectivemethodforfighting‘‘informationoverload’’.However,itsperformanceisoftenlimitedbyseveralfactors,suchassparsityandcold-start.Someresearchersutilizeuser-createdtagsofsocialtaggingsystemtodepictuserpreferencesforpersonalizedrecommendation,butitisdifficulttoidentifyuserswithsimilarinterestsduetothedifferencesbetweenusers’descriptivehabitsandthediversityoflanguageexpression.Inordertofindabetterwaytodepictuserpreferencestomakeitmoresuitableforpersonalizedrecommendation,weintroduceaframeworkthatutilizesitemdomainfeaturestoconstructuserpreferencemodelsandcombinesthesemodelswithcollaborativefiltering(CF).Theframeworknotonlyintegratesdomaincharacteristicsintoapersonalizedrecommendation,butalsoaidstodetectingtheimplicitrelationshipsamongusers,whicharemissedbytheconventionalCFmethod.Theexperimentalresultsshowourmethodachievesthebetterresult,andprovetheuserpreferencemodelismoreeffectiveforrecommendation.CrownCopyright©2013PublishedbyElsevierB.V.Allrightsreserved.1.IntroductionTherapiddevelopmentofinformationtechnologyandthecurrentgrowthandpopularityoftheInternethavefacilitatedanexplosionofinformationthathasexacerbatedtheinformationoverloadproblem[1].Asoneofthemostusefulmethods,personalizedrecommendation,whichwasfirstproposedinthe1990s[2,3],adoptsknowledgediscoverytechniquessuchasdataminingandmachinelearningtodiscoveruserinterestsaccordingtouserbehaviorandthentomakerecommendations[4–6].Atypicalapplicationofpersonalizedrecommendationisinelectroniccommerce,suchasbookrecommendationsinAmazon.com[7],movierecommendationsinNetflix.com[8],videorecommendationsinTiVo.com[9],andsoon.Anefficientrecommendationsystemnotonlyprovidesappropriaterecommendationsforusers,butalsohelpstheserviceprovidergainsubstantialprofits.Mainstreamrecommendationalgorithmscanbedividedintofourcategories[10]:content-based(CB),collaborativefiltering(CF),network-based(NB),andhybridrecommendation(HR).TheCBmethodrecommendsobjectsthataresimilar∗Correspondingauthor.Tel.:[phone].E-mailaddresses:[e-mail](J.Zhang),[e-mail](Q.Peng).0378-4371/$–seefrontmatterCrownCopyright©2013PublishedbyElsevierB.V.Allrightsreserved.(2014)66–7667tothosepreviouslypreferredbythetargetuser.However,thismethodcannotfilteraudio,image,orvideoinformation[10].CFhasbeenthemostsuccessfulrecommendationsystemtechnology[11].InCF,wemakerecommendationaccordingtotheassumptionthatuserswhohavethesimilarperformanceswouldliketochoosethesimilaritems.However,theperformanceofCFissignificantlylimitedbythesparsityofdata[10].NBrecommendationutilizesrelationshipsbetweenusersanditemsorrelationshipsamonguserstoconstructanetwork,andthenanalyzesthenetworktodeterminerecommendationsforusersindirectly.However,the‘‘cold-start’’problemcouldnotbesolved[10].Finally,HRiscurrentlythemostpopularapproachanditcombinesatleasttworecommendationalgorithmstodeterminearecommendation[10].Manyscholarshaverecentlyintegratedvariouskindsofinformationintotherecommendationsystemtoimproveperformance.Suchinformationincludestags,time,trustrelationships,browserecords,socialnetworks,andsoon.Forin-stance,ZhengandLiinvestigatedtheimportanceandusefulnessoftagsandtimeinformationwhenpredictinguserprefer-encesandconsequentlyexaminedhowsuchinformationcouldbeexploitedtobuildaneffectiveresource–recommendationmodel[12].Yinetal.consideredthelatentvalueoftrustrelationshipstoconstructatrustpreferencenetworktomakerec-ommendations[13].Kardanetal.introducedaninnovativearchitectureforarecommendationsystemthattookadvantageofcollaborativetaggingandconceptmaps[14].Zhangetal.proposedarecommendationapproachthatcombinedcontentandrelationanalysesinasinglemodeltoestimatetherelationsamongusers,tags,andresourcesfortag,item,anduserrecommendations[15].Addinginformationtotherecommendationsystemnotonlyimprovestheperformance,butalsoenhancestheunderstandingofwhichfactorsinfluencerecommendations.Thesocialtaggingsystemiscurrentlypopularwithscholarswhoutilizeuser-createdtagstodepictuserpreferencesforpersonalizedrecommendation.Kime