A Multi-Information Enhanced Attention Network for Session-Based Recommendation

Session-based recommendations attempt to anticipate the subsequent item based on past interactions with anonymous users. Only a small subset of user interactions can be modeled in order to fully capture user preferences. Most of existing models have verified that attention mechanisms are useful for modeling user preferences. However, these attention-based models still have two limitations. First, these methods focus on pairwise item-item interactions. Categories and behaviors, which can provide side information to improve recommendation accuracy, are simply ignored. Second, irrelevant items (i.e., spurious user behaviors), which are not favorable for the final result, are difficult to identify and eliminate automatically. To solve the above difficulties, we propose a Multi-Information Enhanced Attention Network (MIEAN). Specifically, we capture attention to items in a session by combining behaviors with items and categories, respectively, and then design enhanced attention to decrease the effects of irrelevant items. We have conducted comprehensive experiments on two real and public datasets, JData and Tmall, and the experimental results show that the proposed model achieves state-of-the-art performance in the main test metrics and also prove that categories and behaviors are useful for modeling user preferences.