Enhancing Session-Based Social Recommendation with Multi-interest Global Popularity Preferences

Session-based Social Recommendation (SSR) has emerged as a critical research area in the field of recommender systems with the explosion of information. However, current SSR methods typically focus on utilizing session information through the aggregation of time series data, while their use of social information generally relies solely on leveraging social relationships of neighboring users to assist in embedding learning. Previous approaches overlook the influence of collective user preferences on individual users’ final choices. To address this question, we propose a Global Popularity Preference Enhanced SSR model that improves recommendation performance by capturing and incorporating multi-interest global popularity trends. First, we employ a heterogeneous graph neural network (HGNN) enhanced with a relational attention mechanism to aggregate long-term user interest representations and user-item embeddings, taking into account both interactions with different frequencies and social information. We incorporate the varying strengths of social relationships during the message-passing process to better model user influence. Furthermore, based on the global item embeddings, we extract global popularity, which reflects the collective user preferences over a specific period. Our model combines session-based time series information, user and item embeddings, and global popularity to aggregate session embeddings for more accurate recommendation predictions. Extensive experiments on three datasets demonstrate that our method outperforms the latest algorithms.