A community-and-role-aware graph neural network on link prediction

Link prediction is a crucial problem in complex networks, and graph neural networks provide an effective approach. However, many existing methods suffer from the representation learning in sparse networks. In reality, most real-world networks exhibit sparsity and contain a large number of small-degree nodes. Due to insufficient structural information, the ability to learn meaningful embeddings is significantly constrained for small-degree nodes. In this article, we propose a novel graph neural network method by constructing a two-layer graph attention network, with the community and role information incorporated into the graph attention network for the small-degree nodes. The aggregated information is not limited to direct neighbors, but is extended to more nodes with structural similarity, thereby greatly enriching the structural information of small-degree nodes. Experiments on eight real-world datasets demonstrate the effectiveness of the proposed method.