基于辅助行为去噪的多行为推荐

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）08-018-2383-07

doi：10.19734/j.issn.1001-3695.2025.01.0004

Multi behavior recommendation based on auxiliary behavior denoising

ChenWenhao，Chen Yuan ，Zhu Xiaei （ 4OOo54，China）

Abstract：Multi-behaviorrecommendation（MBR）systemsanalyzeuser preferences throughauxiliary behaviors toaleviate data sparsityand enhancerecommendationaccuracy.However，previous approaches thatused randomly initialized userand itemembeddingsfailed toprovidesufficientinformationvalueandignoredthenoiseinbehaviorembeddingagregationaswell asthenoiseinauxiliarybehaviorinteractionsequences.Toovercome theselimitations，thiswork introducedtheauxiliarybehavior denoising MBR model（ABD-MBR）.The framework firstlypre-trainedusersand items togeneratemeaningful initial embedings.Then itdesigneda projectionagregation module to mapauxiliarybhaviorsto thetargetbehavior space，minimizing noise interference during embedding fusion.Additionally，the model implemented bot- ∇⋅k and top ∇⋅k sampling modules to refineauxiliarybehavirinteractionsequences，fectivelysuppreingsequence-levelnoise.Finally，themodelemploydmultitask learning tooptimizeitsperformance.Experimentsonthreepubliclyavailabledatasets showthat，comparedtoMB-HGCN， the model achieves an average improvement 10.2% in HR@10 and 13.4% in NDCG@10 ，demonstrating its effectiveness. Key words：multi-behaviorrecommendation；auxiliary behavior；pre-training；embedded projection；multi-task learning

0 引言

随着互联网的快速发展，大量的物品信息呈指数级增加，这使得人们要花费大量时间和精力来对这些信息进行管理。（剩余17280字）