Dependency enhanced session-based recommendation based on GNN

Abstract

Session-based recommendations can predict the next item through the user's current session sequence. Session-based recommendation with graph neural network has become the mainstream method in this domain. However, the session sub-graph constructed by the session sequence will bring fake dependencies issue. Adjacent items may be noise, non-adjacent items may have true dependencies; Moreover, the information outside the target session is ignored. To solve these problems, a dependency enhanced session-based recommendation model based on graph neural network (GID-GNN) is proposed. Firstly, it uses GNN to learn session features, meanwhile uses self-attention to capture session items’ dependencies to obtain session level representation; Then a global graph is constructed to obtain effective global level information and global level representation; Finally, the prediction is made by combining two part representations. Experiments were tested on two real datasets, the results proved effectiveness of GID-GNN model.