Refined node type graph convolutional network for recommendation

Recently, because of the remarkable performance in alleviating the data sparseness problem in recommender systems, Graph Convolutional (Neural) Networks (GCNs) have drawn wide attention as an effective recommendation approach. By modeling the user-item interaction graph, GCN iteratively aggregates neighboring nodes into embeddings of different depths according to the importance of each node. However, the existing GCN-based methods face the common issues that, they do not consider the node information and graph structure during aggregating nodes, such that they cannot assign reasonable weights to the neighboring nodes. Additionally, they ignore the differences in node types in the user-item interaction graph and thus, cannot explore the complex relationship between users and items, resulting in a suboptimal result. To solve these problems, a novel GCN-based framework called RNT-GCN is proposed in this paper. RNT-GCN integrates the structure of the graph and node information to assign reasonable importance to different nodes. In addition, RNT-GCN refines the node types, such that the heterogeneous properties of the user-item interaction graph can be better preserved, and the collaborative information of users and items can be effectively extracted. Extensive experiments prove the RNT-GCN achieved significant performance compared to SOTA methods.