Link Prediction in Knowledge Graph with Feature Enhancement

Knowledge graph link prediction aims to predict the unknown relationship between a pair of known entities or predict another unknown entity that has some relationship with a known entity. In the existing Multi-relational Graph Convolutional Networks, the influence differences of different neighbor nodes are not considered during entity and relationship aggregation, and the lack of semantics is caused after feature fusion. To solve this problem, this paper proposes a link prediction framework CompAECA (CompGCN Attention ECA, CompAECA) based on Multi-head Attention, Multi-relational Graph Convolutional Network and improved channel attention. The framework can realize variable weight aggregation by integrating Multi-head Attention into the Multi-relational Graph Convolutional Network, so as to consider the differences of different neighbor nodes. And an improved channel attention mechanism is introduced after feature fusion. This channel attention adds two two-dimensional convolutions to extract richer feature information for feature enhancement. The problem of equal aggregation of neighbor information and semantic loss caused by feature fusion in Multi-relational Graph Convolutional Networks is solved. Experimental results on three standard link prediction datasets show that the proposed framework achieves advanced results.