Leveraging Local Consecutive Information for Session-Based Recommendation

Session-based recommendation is the task of predicting the user's next actions based on the session history without any information about the long-term user profile. Inspired by the fact that a session can be modeled as a sequence, various methods have been proposed using transformer-based neural networks to encode sessions. Recent studies show that user behaviors can be influenced by groups of consecutive items in a session. Although the conventional transformer architecture can capture such relations, it requires much better data. To address this problem, we propose a new attention mechanism that can encode information of consecutive items and uses several transformer models to capture such information at multiple levels. We then use a weighted aggregation layer to fuse the results at different levels. We empirically evaluate the proposed method on two benchmark datasets. The results show that our model outperforms other transformer-based methods on common evaluation metrics.