Degree aware based adversarial graph convolutional networks for entity alignment in heterogeneous knowledge graph

Entity alignment, as the vital technique for knowledge graph construction and integration, aims to match entities that refer to the same real-world identity in different knowledge graphs (KGs). Recently, much effort has been devoted to embedding-based methods for entity alignment. For most of such methods, the entity with a high degree is hard to be aligned with its equivalent counterpart with a low degree. This degree difference between equivalent entities poses a great challenge for entity alignment. To solve this problem, a novel entity alignment framework that integrates a graph convolutional network (GCN) based embedding initializer and a degree aware generative adversarial network is proposed. In particular, the embedding initializer utilizes a GCN with highway gates to generate the preliminary embedding of entities based on their topological characteristics in the KGs. By alleviating the relevance between embeddings and degree features, the degree aware GAN mitigates the impact of degree difference and generates the final degree level irrelevant alignment result. To quantify the heterogeneity between the KGs, an evaluation metric called heterogeneity entropy of degree (HED) that represents the degree difference is defined in this paper. Based on HED, the impact of KG heterogeneity on the performance of the proposed DAGCN model is investigated based on WK3l-15k and DBP15k datasets. The experimental results show that the proposed degree aware adversarial graph convolutional network (DAGCN) outperforms other state-of-the-art methods over all metrics, especially when the HED is large.