Abstract

Characterizing users’ interest accurately plays a significant role in an effective recommender system. The sequential recommender system takes successive user-item interactions and dynamic users’ preferences into account, which brings powerful hidden representations of users. Conventional methods of such patterns mainly include Markov Chains (MCs) and Recurrent Neural Networks (RNNs). Recently, the use of self-attention mechanisms and bidirectional architectures enables recommendations not only to capture long-term semantics but also to fuse information from both the left and right sides. However, a major limitation can be observed in previous strategies—unstable representations of users. Sequential models only mine the relationship between items that users have interacted rather than forming the whole representation of each user, which is often a constraint on the further improvement of recommendation performance. To address this limitation, we propose a mixed sequential recommendation by combining bidirectional self-attention with direct user-item match, which brings two benefits. For one thing, our model can learn bidirectional semantics from users’ behavioral sequences. For another, it produces more stable users’ representations and achieves better recommendation performance. Extensive empirical studies demonstrate our approach considerably outperforms various state-of-the-art sequential models.