融合文本增强与深度强化学习的交互式推荐方法

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）12-014-3637-07

doi：10.19734/j. issn.1001-3695.2025.04.0155

Interactive recommendation method combining text enhancement and deep reinforcement learning

Zhang Xiaoyua，Mei Hongyanb†，Hu Siyua，Zhao Entongʰ （a.Schoolofosdfotiongeing，.offeLng UesitfoogouLo China）

Abstract：Toaddressthedata-sparsitybotleneckandtheinefciencycausedbylargediscreteactionspacesininteractive recommendersystems（IRS），thispaperproposedaninteractiverecommendationmethodthattightlyintegrated textenhancement with deepreinforcement learning（TDIRS）.The methodfirstleveragedrich textualinformation togenerateexpresive embedingsand builtadeepreinforcement-learning model poweredbymulti-headatention tocaptureusers’evolving preferences，，thereby mitigating sparsity.Simultaneously，themethoddesignedadynamiccandidate-action generatorthatfuseditem popularitywithembeddingrepresentations toidentifyitemsthat trulyinterested heuserandtoshrinktheactionspace，bostingrecommendationeficiency.Extensiveexperimentsonthemusicdatasetandothersconfirm thatTDIRSoutperformsall baselines on core metrics，achieving up to a improvement of 7percentage points in HR@10 and validating the effectiveness of the approach.

Key words：textual information；multi-headatention mechanism；deepreinforcement learning；interactiverecommendation

0 引言

与传统推荐系统[1]中的推荐被视为一步预测任务不同，交互式推荐系统（IRS）[2-4]中的推荐被视为一个多步问题。（剩余15844字）