Local and global structure-aware contrastive framework for entity alignment

Entity alignment (EA) seeks to identify equivalent real-world entities across different knowledge graphs. Recently, integrating graph neural networks (GNNs) with graph augmentation techniques to aggregate local structural information of entities has been proven effective for EA. However, stacking multiple GNN layers to capture higher-order neighbors often leads to oversmoothing of entity embeddings and the introduction of noise from irrelevant neighbors. In this paper, we propose a novel approach, Local and Global Structure- Aware Contrastive Framework (LGEA), to effectively learn the mutual information between the local and global structures of entities. Specifically, we propose a graph augmentation method using Singular Value Decomposition to capture the global structure. In the Global Structure-Aware Encoder module, we design a Residual Gated Unit to reduce noise and mitigate oversmoothing. LGEA incorporates contrastive learning to maximize the consistency between local and global embeddings. Additionally, we introduce a Degree-Aware Relation Encoder to integrate relational semantic information, enriching the entity embeddings. Extensive experiments on established EA benchmarks demonstrate that our method significantly outperforms previous approaches.