PtSGNN: Pre-Trained Graph Neural Network for Session-Based Recommendation

Session-based recommendation is to provide recommendation services to users by studying their anonymous sessions. Previous works consider each session separately and model the sequence of each session for recommendation, but these models can only perceive the items within a session and cannot consider the relevant connections and global information between sessions. Pre-training techniques can extract session features on global session information and weight the global information to node features, which can capture richer inter-session information. In model construction, sampling node features across sessions is considered to learn hidden representations between different sessions. In this paper, we propose a new method PtSGNN (Pre-trained GraphSAGE Neural Network). PtSGNN combines pre-training on the global graph with cross-session information. The information of the global graph was extracted by pre-training. It is then combined with a graph network capable of extracting cross-session information, and finally generates session representations for recommendation. Extensive experiments on two real-world benchmark datasets show significant and consistent improvements of PtSGNN over state-of-the-art methods.

I. Introduction

Session-based Recommendation Systems (SBRS) have been proposed as a solution to the inaccuracies prevalent in traditional recommendation systems due to insufficient user information. By delving into the interactional behaviors between users and systems, SBRS can capture the latest interests and needs of users, thereby better reflecting their real-time preferences and behavioral patterns [1].