GGAT: Knowledge Graph Embedding Model via Global Information

Abstract

Recently, knowledge graph embedding model based on Graph Attention Network (GAT) has shown great potential in link prediction task. However, the existing GAT based models ignore the global information in the neighborhood. We propose GGAT, a knowledge graph embedding model based on global information. The encoding ability of GGAT is enhanced by using global information. Meanwhile, we employ multi-head attention mechanism to improve GGAT's perception of the interaction between entities in the neighborhood. In addition, GGAT uses residual structure to improve the stability of the model and the ability to perceive remote semantic connections. Experiments on two link prediction benchmarks demonstrate the proposed key capabilities of GGAT. Moreover, we set a new state-of-the-art on a knowledge graph dataset.