Relationship-Enhanced Session-based Recommendation with Graph Neural Networks

Abstract

 The session-based recommendation systems analyze anonymous users' recent behavior data to infer their preferences and provide accurate recommendations. However, existing methods often fail to adequate­ly leverage the click order and frequency of items within a session, thus lacking the ability to fully cap­ture complex dependencies among items. To address these limitations, we propose a novel model named RESR-GNN. The key innovations of our work include: (1) a relation-enhanced session graph that incorpo­rates click information to differentiate the importance of items; and (2) an integrated soft attention and multi-layer self-attention mechanism to comprehensively model pairwise item preferences within the entire session. Extensive experiments on two public datasets demonstrate that RESR-GNN consistently outperforms state-of-the-art baseline models. On the Diginetica and Yoochoose1/64 datasets, the evalu­ation metrics P@20 and MRR@20 were respectively improved by 3.15%, 5.51%, 1.53% and 2.36%, which proved the effectiveness of the proposed model.