DISS-CF: direct item session similarity enhanced collaborative filtering method for recommendation

Recently, as Graph Convolutional Network (GCN) achieves great success in graph representation tasks, GCN-based Collaborative Filtering (CF) methods have been widely applied in recommender systems and achieved state-of-the-art performance. However, to our best knowledge, these methods only applies an undirected bipartite graph containing no direct relations between items which has been verified to have better performance in our experiments. Moreover, these methods do not consider the impact of the sequential information from historical interaction sequences. To explore the effect of direct relations between items and the importance of sequential information, in this paper, we propose a new method called Direct Item Session Similarity Enhanced Collaborative Filtering (DISS-CF)- a general improved method for GCN-based recommender systems. DISS-CF calculates the direct similarities among items by applying session windows and generates an enhanced directed weighted interaction graph instead of an undirected bipartite graph. This work conducts comprehensive experiments by integrating the DISS-CF into several popular GCN-based recommender system models on different real-world benchmark datasets. The results validate the effectiveness of direct item relations and the importance of sequential information from historical interactions.