CARE: Context-aware attention interest redistribution for session-based recommendation

Session-based recommendation (SBR) faces the challenge of modeling user behavior patterns within limited session sequences to predict the next item in anonymous sessions accurately. Recent models in this field have evaluated the effectiveness of attention mechanisms. These models focus on learning item representations by modeling transitions between items. They typically use the item embedding, such as the last-clicked item, as a query to calculate attention-based interest distribution, capturing user behavior preferences. However, it is crucial to acknowledge that the distribution of user interest weights within a session cannot be only attributed to a specific embedding. Such an assumption may overlook the complexity of user interest distribution patterns, which could negatively affect the accuracy of recommendations. To address the limitation, we propose the C ontext-aware A ttention interest RE distribution ( CARE ) model. A novel modified attention mechanism designed to effectively encode diverse and evolving interest distribution patterns by leveraging the session contextual information. Our method incorporates a learning context-aware relation- score encoder trained to capture dynamic interest distributions effectively within each session. Additionally, a session-length aware attention mechanism is introduced, which adaptively models the varying interests and their redistribution across different session contexts. Extensive experiments conducted on three public datasets demonstrate that the CARE model outperforms existing state-of-the-art session-based recommendation algorithms in performance. The effectiveness of CARE in accurately capturing diverse and evolving patterns of user interest distribution is also highlighted.