DGE-ASR: Two-view Graph Neural Networks for Knowledge Graph Completion

Knowledge graph completion aims to address knowledge gaps caused by difficulties in collecting real-world facts. This paper proposes a novel dual-view graph neural network model, DGE-ASR, which designs a dual isomorphic graph attention encoder. It separately constructs isomorphic graphs centered on entity nodes and relation nodes, and then aggregates important neighbor information on both graphs using graph attention networks. Although the node types in the two graphs differ, modeling the entity and relation perspectives separately captures the heterogeneity of the knowledge graph to some extent. During the encoding phase, quaternion operations are used to enhance the GNN layers, effectively modeling the complex interactions between entities and relations, as well as between relations and entities. Additionally, skip connections and residual connections are designed to mitigate the issues of oversmoothing and gradient vanishing in graph neural networks. Experimental results on six datasets (FB15K-237, CoDEx, LitWD1K) show that DGE-ASR outperforms existing GNN- based models and tensor decomposition methods.