MHGNN: Hybrid Graph Neural Network with Mixers for Multi-interest Session-Aware Recommendation

The session-aware recommendation aims to predict the next interaction based on the user’s current and historical sessions. Despite the remarkable achievements of existing methods, they still have drawbacks in some aspects. Firstly, most existing methods only consider transition relationships between items within the current user’s sessions, while neglecting the valuable item transition patterns from other users and the useful preferences from similar users. Moreover, the issue of long-range dependencies within and across sessions has been ignored. To address these issues, we propose a novel Hybrid Graph Neural Network with Mixers for multi-interest session-aware recommendation, named MHGNN. MHGNN constructs a heterogeneous global graph and homogeneous session graphs to capture global contextual information and personalized behavior patterns. Furthermore, cross-session and cross-channel mixers are devised to capture cross-correlation features, solving the long-range dependency problem. Finally, multiple interests of the user are extracted and fused from the user’s sessions for personalized recommendations, while two interest-enhancing signals are designed to enhance the effectiveness. Extensive experiments on three real-world datasets demonstrate that MHGNN outperforms state-of-the-art models.