MinSR: Multi-level Interests Network for Session-Based Recommendation

Abstract

Session-based recommendation, which is the task of predicting user behavior based on anonymous sessions in the recommendation domain, has drawn increasing interests in recent years. Most existing studies focus on modeling a session as a sequence and only capturing the item-level dependency. Although newly impressive progress has been made to session-based recommendation utilizing graph neural networks, those methods are deficient in incorporating multi-level coupling relations and capturing the session-level information. In this paper, we propose a multi-level interests network (MinSR) based on Graph Neural Networks (GNN) and Attention mechanism, which can simultaneously integrate multi-level interests in the recommendation process and provide a framework for exploiting both current and global session relationships. On the aspect of the current session, we extract Current Preference (CP) and Interest Point (IP) for each session using graph neural network and attention network. On the aspect of the global session, we generate Global Tendency (GT) via self-attention graph pooling for the session graph. Finally, by inherently combining them in a unified framework, our method can take into account both current and global session dependencies. Extensive experimental results based on two real-world datasets demonstrate that the proposed MinSR achieves competitive results compared with the state-of-the-art approaches.