DeHier: decoupled and hierarchical graph neural networks for multi-interest session-based recommendation

Ronghua Lin, Feiyi Tang, Chengzhe Yuan, Hao Zhong, Weisheng Li, Yong Tang

Abstract: Session-based recommendation is a crucial task for learning the user’s intent and predicting their next clicks within a short session. Due to the diversity of user interests across their historical sessions, multi-interest session-based recommender systems have emerged accordingly. However, most existing methods suffer from random item embedding initialization and neglect interest propagation across sessions. In this paper, we propose a multi-interest session-based recommender system named DeHier, which consists of three parts: (1) decoupled item global embedding, (2) multi-interest item embedding, and (3) next-clicked item prediction. In the decoupled item global embedding module, DeHier learns coarse-grained item representation, subsequently capturing user long-term preferences. In the multi-interest item embedding module, DeHier employs hierarchical graph neural network (GNN) layers to propagate diversified user interests from historical sessions to current sessions and learn user short-term preferences. The user long-term preferences will be integrated into short-term preferences to make final recommendations. To demonstrate the effectiveness of DeHier, we conduct extensive experiments on three real-world datasets. The experimental results indicate that our proposed method is superior to the chosen baselines. Specifically, DeHier achieves up to 4.12% and 7.43% improvements of HR@5 and MRR@10, respectively, on the LastFM dataset.