Time Series Enhanced Graph Neural Networks for Session-based Recommendation

In recent years, Graph Neural Networks (GNNs) have received much attention in many fields. Many researchers have also tried applying GNNs to session-based recommendation(SBR), which predicts the next click action of anonymous users based on historical behavior sequences. However, most existing approaches suffer from two problems. Firstly, they cannot capture the temporal order of information propagation within a session well. The second is that users' interests may differ in different sessions, and most methods focus on extracting information from a single session without considering the relationships between sessions. For the first problem, we propose a Time Series Enhanced Graph Neural Networks(TSGNN), which assigns different labels to the edges of the session graph and learns intra-session item representations according to the order of information propagation. For the second problem, we construct multiple sessions as global graphs to achieve inter-session item representation learning that explicitly captures dynamic user interests. Finally, the learned item representations at both levels are combined with position vectors using a soft attention mechanism. We have conducted various experiments on two benchmark datasets showing that the TSGNN model outperforms state-of-the-art methods.