MPAN: Multi-Parallel Attention Network for Session-based Recommendation

Abstract A powerful session-based recommender can typically explore the users’ evolving interests (i.e., a combination of her long-term and short-term interests). Recent advances in attention mechanisms have led to state-of-the-art methods for solving this task. However, there are still three main limitations. First, most of the attention-based methods only utilize the last clicked item to represent users’ short-term interests which ignores the temporal information and behavior context. Second, to learn users’ long-term interests, most existing models employ the vanilla attention method, which has difficulty in capturing the diversity of long-term interests. Third, current studies typically assume that long-term interests and short-term interests are equally important ignoring their user-specific importance. Therefore, we propose a Multi-Parallel Attention Network (MPAN) model for Session-based Recommendation. Specifically, we propose a novel time-aware attention mechanism to learn users’ short-term interests by capturing contextual information and temporal signals simultaneously. Next, we design a refined multi-head attention mechanism to extract the diverse long-term interests from different latent subspaces. Besides, we introduce a gated fusion method that adaptively integrates users’ long-term and short-term interests to generate a hybrid interest representation. Experiments on three real-world datasets show that MPAN achieves noticeable improvements than state-of-the-art methods.