摘要
多行为推荐(multi-behavior recommendation,MBR)在互联网平台中愈发重要,但现有方法仍面临两大挑战:a)无法刻画用户不同行为下的复杂兴趣偏好;b)难以建模不同行为间的相互关系。基于此,提出一种对比学习增强的多行为超图神经网络模型(multi-behavior hypergraph neural network model enhanced with contrastive lear-ning,MBHCL),在建模用户复杂多类型交互的同时,结合对比学习捕获行为间共性与差异,以获取更优嵌入表示,缓解冷启动与数据稀疏问题。具体地,MBHCL首先构建用户-项目多行为交互超图,以刻画用户对项目不同维度的偏好;其次设计三个对比任务整合单行为表示,通过捕捉行为间的共性与差异获取全面用户兴趣偏好。最终,MBHCL在四个真实场景数据集上进行对比实验。结果表明,在Tmall和BeiBei数据集上,HIT和NDCG指标有至少4.8%的提升,在Kuairand和Yelp数据集上,HIT和NDCG指标至少提升3.6%,并通过消融实验验证了各模块的有效性,同时显著改善了冷启动用户推荐效果。
Multi-behavior recommendation(MBR)systems are increasingly important in internet platforms but face two critical limitations:a)failure to characterize users’complex preferences under diverse behaviors,b)difficulty modeling inter-behavior relationships.This study proposed a multi-behavior hypergraph neural network model enhanced with contrastive lear-ning(MBHCL)to address these issues.The method constructed user-item hypergraphs for multi-behavior interactions,capturing users’multi-dimensional preferences.It designed three contrastive tasks to integrate single-behavior representations through commonality-difference modeling,obtaining optimized embeddings to alleviate cold-start and data sparsity problems.Experiments on four real-world datasets(Tmall,BeiBei,Kuairand,Yelp)demonstrate MBHCL’s effectiveness.The model achieved minimum 4.8%HIT and NDCG improvements on Tmall and BeiBei datasets,and 3.6%enhancements on Kuairand and Yelp datasets.Ablation tests verified all components’contributions,with cold-start recommendations showing significant performance gains.
作者
王光
李佳欣
Wang Guang;Li Jiaxin(College of Software,Liaoning Technical University,Huludao Liaoning 125105,China)
出处
《计算机应用研究》
北大核心
2025年第8期2304-2311,共8页
Application Research of Computers
基金
国家自然科学基金资助项目(62173171)。
关键词
推荐系统
多行为推荐
图神经网络
超图
对比学习
自监督学习
recommendation system
multi-behavior recommendation
graph neural network
hypergraph
contrastive lear-ning
self-supervised learning
