《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (4): 1061-1068.DOI: 10.11772/j.issn.1001-9081.2024030393

• 人工智能 • 上一篇    下一篇

基于多视图多尺度对比学习的图协同过滤

党伟超, 温鑫瑜(), 高改梅, 刘春霞   

  1. 太原科技大学 计算机科学与技术学院,太原 030024
  • 收稿日期:2024-04-08 修回日期:2024-05-26 接受日期:2024-05-29 发布日期:2024-08-15 出版日期:2025-04-10
  • 通讯作者: 温鑫瑜
  • 作者简介:党伟超(1974—),男,山西运城人,副教授,博士,CCF会员,主要研究方向:智能计算、软件可靠性
    温鑫瑜(1998—),男,山西晋中人,硕士研究生,主要研究方向:推荐系统
    高改梅(1978—),女,山西吕梁人,副教授,博士,CCF会员,主要研究方向:网络安全、密码学
    刘春霞(1977—),女,山西大同人,副教授,硕士,CCF会员,主要研究方向:软件工程、数据库。
  • 基金资助:
    山西省自然科学基金资助项目(202203021211194);太原科技大学博士科研启动基金资助项目(20202063);太原科技大学研究生教育创新项目(SY2022063)

Multi-view and multi-scale contrastive learning for graph collaborative filtering

Weichao DANG, Xinyu WEN(), Gaimei GAO, Chunxia LIU   

  1. College of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan Shanxi 030024,China
  • Received:2024-04-08 Revised:2024-05-26 Accepted:2024-05-29 Online:2024-08-15 Published:2025-04-10
  • Contact: Xinyu WEN
  • About author:DANG Weichao, born in 1974, Ph. D., associate professor. His research interests include intelligent computing, software reliability.
    WEN Xinyu, born in 1998, M. S. candidate. His research interests include recommendation system.
    GAO Gaimei, born in 1978, Ph. D., associate professor. Her research interests include network security, cryptography.
    LIU Chunxia, born in 1977, M. S., associate professor. Her research interests include software engineering, database.
  • Supported by:
    Shanxi Provincial Natural Science Foundation(202203021211194);Doctoral Research Start-up Fund of Taiyuan University of Science and Technology(20202063);Graduate Education Innovation Project of Taiyuan University of Science and Technology(SY2022063)

摘要:

针对图协同过滤推荐方法存在的单一视图局限性和数据稀疏性问题,提出一种基于多视图多尺度对比学习的图协同过滤(MVMSCL)模型。首先,根据用户-项目交互构建初始交互图,并考虑用户-项目中存在的多种潜在意图,以构建多意图分解视图;其次,利用高阶关系改进邻接矩阵,以构建协同邻居视图;再次,去除不重要的噪声交互,以构建自适应增强的初始交互图和多意图分解视图;最后,引入局部、跨层和全局3种尺度的对比学习范式生成自监督信号,从而提高推荐性能。在Gowalla、Amazon-book和Tmall 3个公共数据集上的实验结果表明,MVMSCL的推荐性能均优于对比模型。与最优基线模型DCCF(Disentangled Contrastive Collaborative Filtering framework)相比,MVMSCL的召回率Recall@20分别提升了5.7%、14.5%和10.0%,归一化折损累计增益NDCG@20分别提升了4.6%、17.9%和11.5%。

关键词: 推荐系统, 协同过滤, 图神经网络, 多视图, 对比学习

Abstract:

A Multi-View and Multi-Scale Contrastive Learning for graph collaborative filtering (MVMSCL) model was proposed to address the limitations of single view and the data sparsity in graph collaborative filtering recommendation methods. Firstly, an initial interaction diagram was constructed on the basis of user-item interactions, and multiple potential intentions in user-item interactions were considered to build multi-intention decomposition view. Secondly, the adjacency matrix was improved using high-order relationships to construct a collaborative neighbor view. Thirdly, the irrelevant noise interactions were removed to construct the adaptively enhanced initial interaction diagram and multi-intention decomposition view. Finally, contrastive learning paradigms with local, cross-layer, and global scales were introduced to generate self-supervised signals, thereby improving the recommendation performance. Experimental results on three public datasets, Gowalla, Amazon-book and Tmall, demonstrate that the recommendation performance of MVMSCL surpasses that of the comparison models. Compared with the optimal baseline model DCCF (Disentangled Contrastive Collaborative Filtering framework), MVMSCL has the Recall@20 increased by 5.7%, 14.5% and 10.0%, respectively, and the Normalized Discounted Cumulative Gain NDCG@20 increased by 4.6%, 17.9% and 11.5%, respectively.

Key words: recommendation system, collaborative filtering, Graph Neural Network (GNN), multi-view, contrastive learning

中图分类号: