作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (5): 112-121. doi: 10.19678/j.issn.1000-3428.0064216

• 人工智能与模式识别 • 上一篇    下一篇

基于图卷积网络融合群组关系的社会化推荐方法

陈昱瑾1,2,3, 王晶1,2,3, 武志昊1,2,3, 赵耀帅3,4, 林友芳1,2,3   

  1. 1. 北京交通大学 计算机与信息技术学院, 北京 100044;
    2. 交通数据分析与挖掘北京市重点实验室, 北京 100044;
    3. 民航旅客服务智能化应用技术重点实验室, 北京 101318;
    4. 中国民航信息网络股份有限公司, 北京 101318
  • 收稿日期:2022-03-17 修回日期:2022-05-17 发布日期:2022-08-19
  • 作者简介:陈昱瑾(1998-),女,硕士研究生,主研方向为深度学习、推荐系统;王晶、武志昊,副教授、博士生导师;赵耀帅(通信作者),高级工程师、硕士;林友芳,教授、博士生导师。
  • 基金资助:
    中国民航信息网络股份有限公司与民航旅客服务智能化应用技术重点实验室基金项目(K20L00070)。

Social Recommendation Method Integrating Group Relationships Based on Graph Convolution Network

CHEN Yujin1,2,3, WANG Jing1,2,3, WU Zhihao1,2,3, ZHAO Yaoshuai3,4, LIN Youfang1,2,3   

  1. 1. School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China;
    2. Beijing Key Laboratory of Traffic Data Analysis and Mining, Beijing 100044, China;
    3. Key Laboratory of Intelligent Application Technology for Civil Aviation Passenger Services, Beijing 101318, China;
    4. TravelSky Technology Limited, Beijing 101318, China
  • Received:2022-03-17 Revised:2022-05-17 Published:2022-08-19

摘要: 协同过滤推荐系统普遍面临交互数据稀疏,社会化推荐通过引入用户社交信息来缓解数据稀疏问题。现有社会化推荐方法主要关注好友关系,即用户间形成的直接社交关系,但社交数据的稀疏性限制了该类方法的性能表现。由用户加入兴趣小组所形成的群组关系数量繁多且富有价值,然而目前较少有研究关注这种关系,仅有的方法多采用矩阵分解等传统方法建模,对用户协同兴趣和社交影响的表达不够深入。为提升推荐质量,进一步研究群组关系,从缓解社交数据稀疏性的角度论证其在辅助推荐方面的作用,并基于建模能力更强的图卷积网络学习用户、项目与群组之间的高阶关系,分别设计出以间接和直接方式利用群组关系的推荐方法IGRec-Trans和IGRec-Direc,探索更合理的群组关系融合方式。在真实数据集上的实验结果表明,所提方法能有效提升推荐性能,相比最优基准方法DiffNet++,在HR@10和NDCG@10指标上最高可提升4.55%和3.98%,在冷启动用户推荐任务上NDCG@10指标最高可提升18.6%。

关键词: 社会化推荐, 群组关系, 图卷积网络, 表示学习, 注意力机制

Abstract: Collaborative filtering recommender systems typically experience the problem of interaction data sparsity. The social recommendation is proposed to alleviate this issue by introducing users' social information. Existing social recommendation studies primarily focused on the direct connections between users, such as friendship and correlation between users. However,the sparsity of social data limits the performance of these recommender systems. It should be noted that the user-group relationship,another type of valuable information in social networks, which is formed by users joining the interest groups they are interested in, has not received widespread attention. The few existing group-information-enhanced methods have only attempted to integrate it using traditional methods, such as matrix factorization. The collaborative interests and social influence of users have not been modeled satisfactorily. It further studies the user-group relationship to improve recommendation performance, demonstrates its effectiveness to assist in the recommendation from the perspective of alleviating the sparsity of social data, and employs a more robust high-order GCN-based model to learn the representation of users, items, and groups. Specifically, this study proposes two GCN-based methods called IGRec-Trans and IGRec-Direc,which integrate user-group relationships into a social recommendation in indirect and direct manners, respectively. The experimental results for real datasets show that compared with the optimal DiffNet++ model in the benchmark methods, the better method between the two methods improves HR@10 and NDCG@10 by up to 4.55% and 3.98%, respectively. In addition, NDCG@10 can be improved by up to 18.6% on a cold-start recommendation.

Key words: social recommendation, user-group relationship, Graph Convolution Network(GCN), representation learning, attention mechanism

中图分类号: