Entity alignment based on informative neighbor sampling and multi-embedding graph matching

Entity alignment is an important and necessary step in the process of knowledge fusion, which aims to match entities with the same meaning in different knowledge graphs. In this paper, we propose a novel entity alignment method based on informative neighbors sampling and multi-embedding graph matching (Multi-EINS). The graphs are embedded by graph convolutional network and the informative neighbors sampling is used to extract the neighborhood region topological structure feature to enhance the entity embedding. Relation and attribution information are embedded to incorporate former entity embedding, resulting representation-level embedding. Semantic and the character information are considered from outcome-level by calculate the distances of entities. The distance matrixes of multi-embedding are fused and the graph matching algorithm is performed on the fused matrix to align entities from different knowledge graphs. Experimental results on real datasets show that our proposed model effectively solves the entity alignment problem and outperforms 14 previous methods by 1% to 3% at least.