An Improved Graph Neural Network Method Using Relative Position Information for Session-based Recommendation

Abstract

Session-based recommendation mainly solves the recommendation problem in the anonymous scene, which is a challenging task. In recent years, most methods based on graph neural network (GNN) have ignore the location information of neighboring items. So we propose a graph aggregation method that introduces relative location information to capture this information. Specifically, we use two methods to learn item embedding, the location graph aggregation method is mainly used to capture the location relationship information between neighbors, and common graph aggregation method is mainly used to capture higher-order relationship information between items. Finally, we construct a session recommendation model and demonstrate the effectiveness of the proposed method on three datasets.