Broadcast Link Prediction Method for Heterogeneous Networks Based on Homogeneous Subgraph Partitioning

In recent years, the application of graph neural network methods to heterogeneous networks has emerged as a research focus. The link prediction task is one of the most crucial tasks in the domain of complex networks. Heterogeneous networks simultaneously possess complex structural information and diverse semantic information, thereby presenting substantial challenges for link prediction in heterogeneous networks. To this end, this paper proposes a broadcast link prediction approach based on homogeneous subgraph partitioning (Metapath Partitioned Broadcast Link Prediction, MPBLP). Specifically, the original heterogeneous network is initially partitioned into homogeneous subgraphs with distinct semantics through meta-paths. Secondly, within each homogeneous subgraph, broadcast message passing is carried out for the target node pair (u, v) to acquire the information of the paths between node pairs. Finally, the embedding representations in different homogeneous subgraphs are weighted and aggregated via an attention mechanism to obtain the final representation of the node pair and accomplish the link prediction task. Experiments conducted on multiple public heterogeneous network datasets indicate that the MPBLP method proposed in this paper significantly surpasses the existing baseline methods in terms of link prediction accuracy, demonstrating the effectiveness and theoretical value of this method in the field of heterogeneous network link prediction.