Spatio-Temporal Attentive Network for Session-Based Recommendation

Abstract

Session-based recommendation aims to predict the user’s next click behavior based on the existing anonymous session information. Existing methods either only utilize temporal information of the session to make recommendations or only capture complex item transitions from spatial perspective to recommend, they are insufficient to obtain rich item representations. Besides, user’s real purpose of the session is also not emphasized. In this paper, we propose a novel session-based recommendation method, named Spatio-Temporal Attentive Session-based Recommendation, STASR for brevity. Specifically, we design a hybrid framework based on Graph Neural Network (GNN) and Gated Recurrent Unit (GRU) to obtain richer item representations from spatio-temporal perspective. During the process of constructing corresponding session graph in GNN, an individual-level skipping strategy, which considers the randomness of user’s behaviors, is proposed to enrich item representations. Then we utilize attention mechanism to capture the user’s real purpose involved user’s initial will and main intention. Extensive experimental results on three real-world benchmark datasets show that STASR consistently outperforms state-of-the-art methods on a variety of common evaluation metrics.