Graph Convolutional Neural Network based on the Combination of Multiple Heterogeneous Graphs

Collaborative filtering algorithms based on graph neural networks have achieved good performance in many scenarios of recommender systems (RSs), most of which utilize the raw user-item interaction information for recommendation. However, user-item interaction information is usually sparse, limiting the recommendation performance of the algorithms. Therefore, how to fully mine the sparse user-item interaction information and improve the accuracy of the recommendation algorithms is an important issue worth studying. To address this issue, this paper proposes a graph convolutional neural network algorithm based on the combination of multiple heterogeneous graphs (GCN-MHG), which constructs a heterogeneous fully adjacent graph by connecting user-user and item-item nodes to each other (i.e., homogeneous nodes are connected to each other), and design a heterogeneous full adjacency auxiliary (HFAA) layer to aggregate information on this graph. The HFAA layer not only expands the influence of distant nodes on the target node, making the target node can receive global information, but also alleviates the information bottleneck phenomenon. In addition, we find that retaining the node features in the process of aggregating information at some layers (i.e., adopting self-connection of nodes at partial layers) will not only avoid the damage to the algorithms' performance caused by the features without semantic information such as ID of nodes in the scenarios of RSs, but also strengthen the influence of non-one-hop neighbor nodes. Experiments on three real datasets demonstrate that the proposed GCN-MHG performs better than the state-of-the-art algorithms. At the same time, several ablation experiments are also designed, and the rationality of GCN-MHG is analyzed theoretically and experimentally. Codes are available at https://github.com/huangheyuanbj/GCN-MHG.