融合动态邻域选择与多兴趣建模的图神经网络推荐模型

打开文本图片集

关键词：推荐系统；图神经网络；强化学习；多头注意力机制

中图分类号：TP391.3 文献标志码：A 文章编号：1001-3695（2026）01-007-0060-09

doi：10.19734/j.issn.1001-3695.2025.06.0185

Graph neural network recommendation model integrating dynamic neighborhood selection and multi-interest representation

WangChong，Gu Chengtong†，Fu Xiang，HongXin （School ofBusiness，Guilin University ofElectronic Technology，Guilin Guangxi541OO4，China）

Abstract：Recommendation systems often relyongraph neural networks to model complex interactions between users and items.However，existing methods generalladopt staticorrandomneighborhoodsampling strategies，whichnotonlyeasilyintroduce noisy information butalsofail toadapttothedynamicchanges ofuser interests.To addressthese isses，this paper proposed DNGM.Onthe user side，themodeladopteda multi-head atention mechanism.Itused multiple independent ttetionheads tofocusondiferentfeaturesubspaces inparallel.Themodelfurthercapturedusers’multi-dimensionalinterestrepresentations.Ontheitem side，themodeloptimizedtheneighborhood selection strategyacording touser interestsandrecommendationgoalsviatheactor-criticreinforcementlearningframework．Itrealizeddynamicagregationofneighborhdnformation，effectivelysuppressedoiseinterference，andimprovedepresentationqualityExperimentalesultsontheepublicdatasets，namely MovieLens-1M，Book-Crossng，andLast.FM，show thatthe proposed model outperforms existing mainstream models in terms of AUC，accuracy （ACC），and other evaluation metrics. Specifically，AUC increases by up to 2. 81% and ACC by up to 1.32% ，demonstrating its effectiveness and robustness.

Key words：recommendation system（RS）；graph neural network；reinforcement learning；multi-head attntion mehanism

0 引言

随着互联网与信息技术的迅速发展，用户能够轻松接触到海量的在线内容，涵盖新闻、影视、教育、电商等多个领域。（剩余22313字）