Subgraph-aware virtual node matching Graph Attention Network for entity alignment

Abstract

With the vigorous development of knowledge graph, its construction has become a hot issue recently. However, most knowledge graphs are incomplete, and it is worthwhile to devote much effort to further fuse the knowledge of different entities. Hence, entity alignment has surprisingly emerged as a key step of knowledge fusion to tackle the above issue. The existing GNN-based models simply aggregate entity embeddings, or incorporate auxiliary information of entities (e.g., relationship, time, etc.), which frustratingly fail to consider neighborhood structure of entities and the cross-graph matching information of the nodes between two knowledge graphs. Thus, to overcome the two challenges, we propose a novel Subgraph-aware Virtual Node Matching Graph Attention neTwork for entity alignment, named SVNM-GAT. Specifically, to strikingly capture the cross-graph matching interaction information of entities, we propose a new virtual node matching mechanism, which can obtain the long-range dependencies between virtual nodes and holistic nodes by scale-dot attention. Additionally, a subgraph discriminator is presented in SVNM-GAT to identify the subgraph structure and obtain weighted adjacent matrix by calculating subgraph structure coefficients. Finally, we adopt experiments to verify the effectiveness of our model on the cross-linguistic and temporal datasets, and all the results indicate that SVNM-GAT considerably outperforms the state-of-the-art methods.