Recommendation Based on Graph Neural Network with Structural Identity

With the development of graph neural networks (GNN), some researchers use interaction records to construct graphs and use GNN to model and capture the information on the neighborhood of user nodes or item nodes, so as to make good use of cross-user collaborative signals in the recommendation. However, using GNN to model the graph structure is limited, because it can only capture the information carried by neighbors within a certain range of the central node, called community information, but can hardly capture structural identity such as the roles of the nodes in the graph, leading to sub-optimal recommendation performance. In order to solve the limitation, we design a Structural-Identity-aware Graph Neural Network, called SIGNN, which can capture both the structural identity that independent of nodes' position in the graph and the community information around the neighbors of nodes, to make full use of the information in the interaction graph and make a better recommendation. The experimental results on three real-world datasets show that SIGNN outperforms existing methods, and the effects of each component are verified by ablation experiments.