Combine temporal information in session-based recommendation with graph neural networks

Abstract

With the rapid growth of Internet data, recommendation systems have become the basic technology to alleviate information overload. The session-based recommendation (SBR) is a challenging task, and its purpose is to predict user behavior based on anonymous sessions. The existing SBR methods need to be improved in capturing the complex conversion relationship of items. In contrast, Graph Neural Network (GNN) can capture the complex conversion between items by modeling sessions as graph structure data. However, current methods just sort the clicked items within a session based on time, without utilizing the temporal information between sessions, leading to poor recommendation performance. To improve the accuracy of session recommendation (SR), we propose an SBR model based on GNN with Combined Temporal (CT-GNN) information. The proposed CT-GNN model is built based on the time of session occurrence, and it can learn the temporal association relationship between session items to enrich the connection between items. More importantly, based on the diversity problem faced by current SRs, which refers to the tendency of users to engage with popular items and resulting in limited exposure for other items, the CT-GNN model uses the Local Item Representation Learning (LIRL) module to learn users' local preferences. Through the LIRL module, the CT-GNN can capture users' interests, boost item exposure, and increase recommendation diversity. The experimental results show that the CT-GNN model is superior to the state-of-the-art SBR methods, with MRR scores 1.1 %∼3.4 % higher than the best-performing baseline; recall scores 0.9 %∼2.3 % higher than the best-performing baseline, and alleviates the diversity problem faced by SR, greatly improving the speed of model convergence, and the training time is 13.6 %∼20.7 % of the best-performing baseline.