Interest Driven Graph Structure Learning for Session-Based Recommendation

Huachi Zhou, Shuang Zhou, Keyu Duan, Xiao Huang, Qiaoyu Tan, Zailiang Yu

Abstract: In session-based recommendations, to capture user interests, traditional studies often directly embed item sequences. Recent efforts explore converting a session into a graph and applying graph neural networks to learn representations of user interests. They rely on predefined principles to create edges, e.g., co-occurrence of item pairs in the sequence. However, in practice, user interests are more complicated and diverse than manually predefined principles. Adjacent items in the sequences may not be related to the same interest, while items far away from each other could be related in some scenarios. For example, at the end of shopping, the user remembers to purchase items associated with the one purchased at the beginning. While using predefined rules may undermine the quality of the session graph, it is challenging to learn a reasonable one that is in line with the user interest. Sessions are diverse in length, the total number of interests, etc. Signals for supervision are not available to support graph construction. To this end, we explore coupling the session graph construction with user-interest learning, and propose a novel framework - PIGR. It recognizes items with similar representations learned based on sequential behavior and preserves their interactions. Related items reside in the same induced subgraph and are clustered into one interest. A unified session-level vector is retrieved from the different granularity of interests to guide the next-item recommendation. Empirical experiments on real-world datasets demonstrate that PIGR significantly outperforms state-of-the-art baselines.