Rethinking transition relationship between co-occurring items in graph neural networks for session-based recommendation

Abstract

Session-based recommendation aims to recommend items based on anonymous behavior sequences. A session is generally modeled as a sequential structure or graph structure to extract its session-level representation. However, directly modeling a session as a sequence ignores the co-occurrence relationship between pairwise items, and current graph-based methods do not consider information transition directions sufficiently when calculating the similarity of co-occurring items. In this paper, we propose a novel approach, dubbed Item Transition Relationship Graph Neural Networks (ITR-GNN), to model different transition relationships between co-occurring items in an effective way, for better extracting user intents of current sessions. In ITR-GNN, session sequences are modeled as directed unweighted graphs, and two different transition relationships between a pair of co-occurring items are distinguished by two different concatenations of their latent vectors, according to the direction of the edge between them. In addition, we notice that the data used for session-based recommendation contain many noisy labels, degrading recommendation effects. Thereby, we introduce a label distillation strategy to learn improved soft labels and replace potential noisy labels through a teacher–student network. Experimental results on three benchmark datasets demonstrate that the ITR-GNN outperforms state-of-the-art methods. Source code is available at https://github.com/cyq002/ITR-GNN.