Hierarchical Social Similarity-guided Model with Dual-mode Attention for session-based recommendation

Session-based recommendation models users’ interests in sessions to make recommendations. Many previous studies have shown that users usually have similar interests to their friends, and are easily influenced by friends. However, these studies also tend to ignore the fact that users’ interests may merely be similar to certain friends’ interests in certain aspects. To address the above issues, we propose a novel Hierarchical Social Similarity-guided Model with Dual-mode Attention (HMDA) for Session-based Recommendation. Specifically, we first calculate the item-level similarity between users and their friends to select influential friends. We then compute the aspect-level similarity to explore the aspect difference between users’ interests and friends’ interests. Under the guidance of the item-level and aspect-level similarity, HMDA is capable of effectively and accurately aggregating the social influence exerted by friends on users, and further combining users’ individual interests to enhance recommendation performance. In addition, we design a dual-mode attention mechanism to capture the internal dependence and mutual dependence between the long-term and short-term interests of users. The proposed model is extensively evaluated on three real-world datasets. Experimental results demonstrate that our model outperforms the state-of-the-art baseline methods.