Link prediction in heterogeneous networks based on multi-graph convolution

Predicting potential links in heterogeneous networks is essential for understanding complex relationships within networks. Previous studies mostly rely on shallow topological information, leading to inaccurate link predictions. To address this, we propose a multi-graph convolution-based method. First, we use Node2vec to learn structural node features, which are then combined with the original node features to form new feature representations. Next, we construct two additional graphs based on cosine similarity and Euclidean distance between nodes. These along with the original graph, undergo dual-channel convolution to learn node representations. The three learned feature matrices are weighted and fused using an attention mechanism, generating the final node features for link prediction. Experiments show significant improvement in prediction performance. Ablation studies and feature visualization further confirm the method’s effectiveness. This approach improves the accuracy of link prediction in heterogeneous networks and enhances understanding of node relationship patterns.