The Dual- Intent Group Recommendation Combining Hypergraph Convolution and Channel Attention

With the growing popularity of online groups in social networks, group discovery has become a key factor in enhancing user experience and fostering platform development. However, existing recommendation methods face challenges in handling user participation in group discovery tasks, as they often overlook individual user interests and item preferences. To address this issue, we propose an innovative model called DIGE, which aims to understand the different intents behind users' participation in groups, including social intents and personal interest intents. Our model separately models these two types of intents and integrates them for prediction to comprehensively capture user behavior and group characteristics. In the model design, we introduce a hypergraph structure to better preserve group information, while employing a novel structural refinement method to reveal the complexity of user behavior and interests. By incorporating the ECA (Efficient Channel Attention) mechanism, we further enhance the model's focus on important features. Additionally, we observe the phenomenon of intent overlap in real-world scenarios and design a self-supervised learning loss to promote consistency in recommendations. The model was compared against baseline models such as CFAG on three public datasets-Mafengwo, Weeplaces, and Steam-and experimental results demonstrate that our model significantly outperforms existing methods.