Graph Attention Network with Relational Dynamic Factual Fusion for Knowledge Graph Completion

Mei Yu, Yilin Zuo, Wenbin Zhang, Mankun Zhao, Tianyi Xu, Yue Zhao, Jiujiang Guo, Jian Yu

Abstract: Knowledge graph completion (KGC) completes knowledge graphs integrity by predicting missing entities in triples. Existing KGC models have achieved excellent results, especially graph attention network models (GATs). Existing GATs ignore the factual correlation between different relations in the same pair of entities based on the global graph structure. To solve this problem, we propose a model RISDF based on the graph attention network with relational dynamic factual fusion. Specifically, we first extract relational diversified information, including relational context and relational paths, and model them to fully obtain the neighborhood information of relations. Then we introduce the point mutual information and the related impact factor to capture the factual correlation between relations. We also design a dynamic factual fusion cage to combine the dynamic dependence and the factual correlation between relations to form a composite coefficient, which is used as the weight of the relation aggregation to realize relational dynamic fact fusion. In addition, in order to fully share the neighborhood information of relations, we fuse the sum of relational context embeddings and relational path embeddings weighted by the combination coefficient. Experimental results on benchmark datasets show that our model outperforms several state-of-the-art models available in KGC.