Cross-Session Aware Temporal Convolutional Network for Session-based Recommendation

Abstract

Recent advancements in Graph Neural Networks (GNN) have achieved promising results for the session-based recommendation, which aims to predict a user's actions based on anonymous sessions. However, existing graph-structured recommendation methods only focus on the internals of a session and neglect cross-session effect which contains valuable complement information for more accurately learning the taste of the user in the current session. Meanwhile, the graph structure lacks the sequential position information so that different sequential sessions can be constructed as the same graph, inevitably limiting its capacity of obtaining an accurate vector of a session representation. In order to solve the above limitations, we propose Cross-session Aware Temporal Convolutional Network (CA-TCN) model. For the cross-session aware aspect, CA-TCN builds a global-item graph and a session-context graph to model cross-session influence on both items and sessions. Global-item graph explores the global cross-session influence on items by building relevant item connections among all sessions. Session-context graph explores the complex cross-session influence on sessions by building the connections between the current session and other sessions with similar user intents and behavioral patterns as the current session. And, we connect items and sessions with hierarchical item-level and session-level attention mechanism. Besides, compared with the GNN, TCN can perform convolution operation on multi-hops items and maintain sequence information in the process of convolution. Extensive experiments on two real-world datasets show that our method outperforms state-of-the-art methods consistently.