3VFP: Three-View Feature Propagation Entity Alignment

Entity alignment (EA) is a technique used in the field of knowledge representation and data integration that aims to identify and match equivalent entities across multiple knowledge graphs or data sources, it can be used in data fusion seamlessly or other downstream applications. Existing EA-oriented methods are mainly TransE-based and graph neural network based representation learning and obtained good performance, but they still have gaps: such as cannot fully use graph structure information and be well combined with iterative training strategy. To overcome the above shortcomings, a novel model via noise-tolerance three-view feature propagation is proposed in this paper, in which we absorb entity-entity graph to learn feature from neighbor entity, and utilize entity-relation graph to learn feature from the relations between entities. We also utilize highway network to filter out noise during feature propagation. We test the model on the DBP15k dataset, results show that it can work well with iterative training strategy and the hits@1 and MRR raised dramatically.