基于双层图注意力网络的邻域信息聚合实体对齐方法

针对知识图谱中存在部分属性信息对实体对齐任务影响程度不一致以及实体的邻域信息重要程度不一致的问题，提出了一种结合双层图注意力网络的邻域信息聚合实体对齐方法（two-layer graph attention network entity alignment，TGAEA）。该方法采用双层图神经网络，首先利用第一层网络对实体属性进行注意力系数计算，降低无用属性对实体对齐的影响；随后，结合第二层网络对实体名称、关系和结构等信息进行特征加权，以区分实体邻域信息的重要性；最后，借助自举方法扩充种子实体对，并结合邻域信息相似度矩阵进行实体距离度量。实验表明，在DWY100K数据集上，TGAEA模型相较于当前基线模型，hit@1、hit@10和MRR指标分别提升了4.18%、4.81%和5%，证明了双层图注意力网络在邻域信息聚合实体对齐方面的显著效果。

Neighborhood information aggregation entity alignment method based on double layer graph attention network

This paper proposed TGAEA, which combined neighborhood information aggregation with two layers of graph attention networks to tackle challenges in entity alignment tasks within knowledge graphs. Initially, the method utilized the first-layer graph neural network to calculate attention coefficients for entity attribute embedding vectors, aiming to mitigate the impact of irrelevant attributes on entity alignment outcomes. Based on the attribute embedding, the second-layer graph neural network was employed to weight the embedding vectors of entity names, relationships, and structural information, thereby distinguishing the importance of various information within the entity''s neighborhood. Additionally, it utilized the bootstrap method to iteratively expand the seed entity pair, and completed the entity distance measurement by combining the neighborhood information similarity matrix. The experimental results show that the TGAEA model is significantly superior to the current advanced baseline model in the DWY100K dataset, compared to the best method, hit@1 hit@10 and MRR indicators increased by 4.18%, 4.81%, and 5%. These findings emphasize the substantial impact of the two-layer graph attention network on aggregating neighborhood information for entity alignment.