Multi-interest-aware Adaptive Self-attention Network for Session-based Recommendation

Session-based Recommendation (SBR) is to predict the next interactive item based on a group of anonymous users’ temporary behavior sequences. Despite the superior performance of existing methods for SBR, there are still several limitations: (i) Existing methods tend to predict the users’ newest interest but less consider users’ multiple interests, which is vital to obtain long-term preferences. (ii) Also, existing methods take all items in the sequences as wanted ones, ignoring the impact of noisy items, leading to suboptimal performance. To address these limitations, a Multi-interest-aware Adaptive Self-attention Network (MASN) for SBR is proposed, which explicitly considers users’ multiple interests while reducing misclicked items’ impact. Firstly, aiming at the diversity of user interests, multiple interest representations of users are learned from temporary sequences; Secondly, to reduce the impact of misclicked items, the importance score of each item is learned by an improved adaptive self-attention mechanism combined with the interest representation, which also reflects the current intention; Then, the overall preference of users is learned by Gated Recurrent Unit and soft-attention mechanisms; Finally, the intention and the preference are integrated to perform recommendation. The experimental results show that MASN outperforms nine representative baselines across four benchmark datasets.