Behavior-Aware Global-Enhanced Neural Modeling for Sequential Set Recommendation

In some scenarios, such as e-commerce, users typically purchase multiple items in a single transaction, rather than just one item. Recommender systems for such scenarios are expected to recommend multiple items in a single step (also known as next basket/set recommendation). This makes user interest modeling more complex than in the normal recommendation. In this paper, we find that existing methods suffer from insufficient information mining, including ignoring item transition patterns and limited local information modeling, resulting in incomplete and inaccurate representations of user interest learning, which ultimately affects performance. To address these problems, we propose a novel solution Behavior-Aware Global-Enhanced Graphs Networks (BAGE¹), which explores global item relationships in sequential purchases using contextualized structural and global information to infer user interests. Specifically, BAGE constructs two graphs, the transition graph and the exploratory graph, to learn structural and contextual information globally. Each graph contains multiplexed information representing different dependencies within item relations. A behavior-aware representation learning is proposed to extract user preferences from the two graphs. The dynamic behavior encoder integrates the graphical and sequential behavior information. To capture inherent characteristics and capture the correlations of user preferences across different graphs while maintaining their inherent characteristics, a contrastive cross-information learning module guides the learning process of user interests. An adaptive gating fusion module fuses sequential and graphical information by incorporating sequential patterns, item transitions, and contextual influences to produce a comprehensive representation of user interests. Experiments on three real-world datasets demonstrate the advantages of BAGE over existing state-of-the-art methods.