通过对14篇推荐系统论文进行总结，其中包含2篇综述类论文，分别涉及推荐系统中的重排序技术[1]以及基于因果的推荐系统综述[2]。另外12篇算法类论文则包含去偏的协同过滤技术[3]、联邦推荐系统中的中毒攻击[4]、离线推荐中的选择偏差问题[5]、基于MLP结构的序列化推荐[6]、基于哈希学习的推荐算法[7,9]、结合对比学习的序列推荐[8]、异质图上的食谱推荐[10]、POI推荐[11,13]、股票分析和推荐的结合算法[12]、结合自监督与图神经网络的多行为推荐算法[14]。更具体的内容可参考下文的部分摘要内容。
WeiwenLiu,YunjiaXi,JiaruiQin,FeiSun,BoChen,WeinanZhang,RuiZhang,RuimingTang
PengWu,HaoxuanLi,YuhaoDeng,WenjieHu,QuanyuDai,ZhenhuaDong,JieSun,RuiZhang,Xiao-HuaZhou
Recently,recommendersystem(RS)basedoncausalinferencehasgainedmuchattentionintheindustrialcommunity,aswellasthestatesoftheartperformanceinmanypredictionanddebiasingtasks.Nevertheless,aunifiedcausalanalysisframeworkhasnotbeenestablishedyet.Manycausal-basedpredictionanddebiasingstudiesrarelydiscussthecausalinterpretationofvariousbiasesandtherationalityofthecorrespondingcausalassumptions.Inthispaper,wefirstprovideaformalcausalanalysisframeworktosurveyandunifytheexistingcausal-inspiredrecommendationmethods,whichcanaccommodatedifferentscenariosinRS.ThenweproposeanewtaxonomyandgiveformalcausaldefinitionsofvariousbiasesinRSfromtheperspectiveofviolatingtheassumptionsadoptedincausalanalysis.Finally,weformalizemanydebiasingandpredictiontasksinRS,andsummarizethestatisticalandmachinelearning-basedcausalestimationmethods,expectingtoprovidenewresearchopportunitiesandperspectivestothecausalRScommunity.
ChenxiaoYang,QitianWu,JipengJin,XiaofengGao,JunweiPan,GuihaiChen
DazhongRong,QinmingHe,JianhaiChen
YutaSaito,MasahiroNomura
MuyangLi,XiangyuZhao,ChuanLyu,MinghaoZhao,RunzeWu,RuochengGuo
Self-attentionmodelshaveachievedstate-of-the-artperformanceinsequentialrecommendersystemsbycapturingthesequentialdependenciesamonguser-iteminteractions.However,theyrelyonpositionalembeddingstoretainthesequentialinformation,whichmaybreakthesemanticsofitemembeddings.Inaddition,mostexistingworksassumethatsuchsequentialdependenciesexistsolelyintheitemembeddings,butneglecttheirexistenceamongtheitemfeatures.Inthiswork,weproposeanovelsequentialrecommendersystem(MLP4Rec)basedontherecentadvancesofMLP-basedarchitectures,whichisnaturallysensitivetotheorderofitemsinasequence.Tobespecific,wedevelopatri-directionalfusionschemetocoherentlycapturesequential,cross-channelandcross-featurecorrelations.ExtensiveexperimentsdemonstratetheeffectivenessofMLP4Recovervariousrepresentativebaselinesupontwobenchmarkdatasets.ThesimplearchitectureofMLP4Recalsoleadstothelinearcomputationalcomplexityaswellasmuchfewermodelparametersthanexistingself-attentionmethods.
FanWang,WeimingLiu,ChaochaoChen,MengyingZhu,XiaolinZheng
YixinZhang,YongLiu,YonghuiXu,HaoXiong,ChenyiLei,WeiHe,LizhenCui,ChunyanMiao
FangyuanLuo,JunWu,TaoWang
YijunTian,ChuxuZhang,ZhichunGuo,ChaoHuang,RonaldMetoyer,NiteshV.Chawla
XiaolinWang,GuohaoSun,XiuFANG,ShoujinWang,JianYang
HeyuanWang,TengjiaoWang,ShunLi,ShijieGuan,JiayiZheng,WeiChen