Position-aware context attention for session-based recommendation

Abstract

In session-based recommendation scenarios where user profiles are not available, predicting their behaviors is a challenging problem. Previous dominant methods to solve this problem are RNN-based models. Recently, attention mechanisms that allow higher parallelization have shown significant improvement on this issue. However, none of the existing attention-based methods explicitly takes advantage of both the position information and context information in a sequence. We assume that one item usually exhibits different levels of importance when it appears in different positions in a sequence. Therefore, a position-aware context attention (PACA) model is proposed as a remedy, which improves the recommendation performance by taking into account both the position information and the context information of items. PACA introduces positional vectors to model the position information and utilizes a pooling function to generate the context feature vectors. Then the two vectors are combined to generate the attention weight for each item in a session. To further improve the performance, we use a multi-head method to combine several parallel attention modules. Extensive experiments on two real-world datasets show that the proposed attention model is able to achieve very promising performance in comparison with the state-of-the-art methods. Finally, we visualize the positional vectors to explicitly analyze the importance of each position in a sequence.