基于时间块动态图神经网络的序列推荐方法

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）08-009-2312-08

doi：10.19734/j.issn.1001-3695.2024.12.0518

Time-block-based dynamic graph neural network for sequential recommendation

Peng Zihang¹， Zhang Quangui2†，Jin Haibo1，Liu Yixin1，Qi Yuxin1 （1.SchoolofreclUsityL;hlfc& ChongqingUniversity ofArts& Sciences，Chongqing 40216O，China）

Abstract：Dynamicgraph-basedsequentialrecommendationisacurentresearchhotspotintherecommendationsystemdo main.Existing methods typicallconstructed dynamicgraphsateach timestampof user-item interaction sequences，which strugled toaddress noisecaused byocasionaluser behaviors andfailed to effctivelycaptureusers’periodic preferences.To address thesechallenges，thispaper proposedaTBDGNNforsequentialrecommendation.The methodfirstlydivided temporal sequences intomultipletimeblocksbasedonthedistributionofuser-iteminteraction historydata.Thenitconstructeddyamic graphswithineachtimeblocktomodeltemporalevolutionofuserbehaviors.Aditionally，itdesignedatime-block-levelgraph neural network framework tomitigate theimpactofocasional interactions（e.g.，usermisoperations）andcaptureusers’periodicbehaviors through time-blockpartitioning.ExperimentalresultsdemonstratethatTBDGNNsignificantlyoutperformsthe DGEL baseline on core metrics in datasets including MovieLens，achieving a maximum improvement of 8.7% in hit@10 . The findings validate the model's efectiveness in dynamic recommendation and periodic behavior modeling.

Key Words：sequential recommendation；dynamic graph neural network；time block；periodic preferences

0 引言

传统的推荐系统主要依赖于用户的静态偏好与物品间的相似性，通过协同过滤等手段为用户提供个性化推荐[1，2]。（剩余20349字）