Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (16): 108-114.DOI: 10.3778/j.issn.1002-8331.2205-0266

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Hypergraph Convolutional Networks for User Micro-Behavior Session-Based Recommendation

YANG Xianpeng, LI Xiaonan, LI Guanyu   

  1. Faculty of Information Science & Technology, Dalian Maritime University, Dalian, Liaoning 116026, China
  • Online:2023-08-15 Published:2023-08-15

基于超图卷积网络的用户微行为会话推荐

杨显鹏,李晓楠,李冠宇   

  1. 大连海事大学 信息科学技术学院,辽宁 大连 116026

Abstract: Session-based recommender system(SBRS) is widely concerned by industry and academia, because it only uses short-term session information of users to make recommendations without using user profile and long-term history information. When modeling user session information, the existing session-based recommendation system constructs each session as an independent graph, ignores the correlation between items, only considers the single user commodity interaction information, and ignores the diversity of interactions (such as browsing, clicking, adding to shopping cart, etc.). To solve the above problems, this paper proposes a user micro-behavior session-based recommendation method based on hypergraph convolution network. The method first constructs the interaction sequence between users and items as a hypergraph to learn the high-order correlation between items in the session, and uses the hypergraph convolution neural network to get the embedding of the interactive product sequence. Then, a series of operations generated when users interact with goods are represented as “micro-behavior sequence” to enrich the diversity of interaction, and the embedding of micro-behavior sequence is obtained by using gated recurrent unit(GRU) network learning. Finally, the two are fused together to obtain a more fine-grained embedded representation for the SBRS. A large number of experiments in data sets Tmall and JDATA show that, compared with the baseline method, the recommendation accuracy of rating indicators [P@20] and [MRR@20] has been significantly improved.

Key words: hypergraphs, recommendation system, micro-behavior, session-based recommendation

摘要: 会话推荐系统因仅利用用户短期会话信息进行推荐,无需使用用户配置文件和长期历史信息等优点,受到工业界和学术界的广泛关注。现有的会话推荐系统建模用户会话信息时将每一个会话构建为独立的图,忽略了项目间的相关性,且仅考虑单一的用户商品交互信息,忽略了交互的多样性(如浏览、点击、加入购物车等)。针对上述问题,提出了一种基于超图卷积网络的用户微行为会话推荐方法。该方法将用户与项目的交互序列构建为超图,以学习会话中项目间的高阶相关性,利用超图卷积神经网络得到交互商品序列的嵌入;将用户与商品交互时产生的一系列操作表征为“微行为序列”,以丰富交互的多样性,利用门控循环单元(gated recurrent unit,GRU)网络学习得到微行为序列的嵌入;将两者融合到一起,使会话推荐系统得到更细粒度的嵌入表示。在数据集Tmall与JDATA中的大量实验表明,在评价指标[P@20]与[MRR@20]中,相比于基线方法推荐准确性有明显提升。

关键词: 超图, 推荐系统, 微行为, 会话推荐