Dual-view graph neural network with gating mechanism for entity alignment

Entity alignment aims to connect equivalent entities between different knowledge graphs (KGs), which is an important step in knowledge fusion. The structural heterogeneity between KGs severely hinders the development of entity alignment. The existing researches mainly focus on alleviating the structural heterogeneity from the view of entity neighborhood heterogeneity, ignoring the important effect of the relation heterogeneity on it. To this end, we propose a Dual-view graph neural network (GNN) based on a gating mechanism named DvGNet, which comprehensively alleviates the structural heterogeneity of KG from the perspective of entity interaction and relation interaction. From the perspective of entity interaction, DvGNet gives important neighbors high weights to alleviate the heterogeneity of entity neighborhood. From the perspective of relation interaction, DvGNet obtains the relation matching degree between KGs according to the relation embeddings, so as to alleviate the relation heterogeneity. Furthermore, to learn the precise representation for the entity, we propose a concise and effective gating mechanism to aggregate the embeddings among the network layers. We conduct extensive experiments on three entity alignment datasets, as well as detailed ablation studies and analyses, demonstrating the effectiveness of DvGNet.