Graph-enhanced and collaborative attention networks for session-based recommendation

Session-based recommendation uses short interaction sequences of anonymous users to predict the next item most likely to be clicked, and many methods have been proposed. However, there are still problems with the existing methods. Existing approaches can be divided into two groups based on data organization: (1) graph-based methods using graph neural networks to capture complex item transformations; (2) sequence-based approaches using self-attention networks to capture chained user interest patterns. Both methods are only applicable to specific kinds of user interest patterns due to the characteristics of the neural networks they use and cannot be adaptively used in all scenarios. Moreover, the recent approaches capture collaborative information from other sessions by constructing global graphs, etc., in order to enrich the current session, which can compromise personalized modeling due to the introduction of items that are not relevant to the current user. This work proposes a graph-enhanced and collaborative attention network (GCAN) to solve the above problems. In GCAN, graph-enhanced attention is designed to model user interest over item-specific subsequences with the help of a graph mask and distance bias, which include item transformations mined in session graphs and chained user interest in session sequences. In addition, collaborative attention is proposed to model the item representation within the current session at the collaborative level by exploiting the collaborative information from all sessions. Extensive experiments on three real benchmark datasets show that GCAN significantly outperforms state-of-the-art methods.