面向用户多层次多样化推荐的端到端学习框架

打开文本图片集

关键词：多样化推荐；个性化多样化需求；多层次多样化约束；推荐系统；端到端框架中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2026）04-018-1120-09doi： 10.19734/j. issn.1001-3695.2025.08.0280

End-to-end learning framework for multi-level diversified recommendation

Wen Wen1，Guo Xiaotong1，Feng Yali1†，Zheng Jiabi 2，3 ，Hao Zhifeng4 （1.SchooloftercegdoUestofolag;.Clfopueeo PolytechnicalUesityagzhtatKybotofoeaetgUsi jing210023，China；4.Shantou University，Shantou Guangdong 5150o0，China）

Abstract：Diversifiedrecommendationaims toallviatetheinformationcocoonefectandimproveusersatisfactionbyincreasingthediversityamongitems intherecommendationlistandreducing low-qualityrepetitiverecommendations.However，existingdiversifiedrecommendationmethodsoutinelydisregardthepersonalized diversityrequirements whenpromoting recommendationlistdiversityMoreover，mostexisting methodsstillrelyonstagedlearningorstagedoptimizationstrategy，whichlimit theimprovementofrecommendationperformance.Therefore，this paper proposedanend-to-end learning framework for multileveldiversifiedrecommendation（MLDR）from the perspectiveof personalizedrequirements.By introducing constraints that captureusers’multi-level diversifiedrequirements intocollborativefiltering modelssuchasmatrix factorizationorgraph euralnetworks，anddesigningobjectivefunctionsthatbalancedbothrecommendationaccuacyanddiversity，itachievedhereasonable captureofuserpreferences.Extensive experimentsonfourdatasets demonstrate that MLDR performs significantlybetterthanthe state-of-the-artbaselines，anditimproves thediversityofbase models whilepreservingrecommendationaccuracy.

Keywords：diversifidrecommendation；personalizeddiversityrequirements；multi-level diversifiedconstraints；recommen dation system；end-to-end framework

0 引言

推荐系统作为信息过滤工具，通过分析用户交互历史，挖掘用户潜在偏好，从而帮助用户定位感兴趣的产品或服务。（剩余21059字）