Attention-Based Relational Graph Convolutional Network for Knowledge Graph Reasoning

In recent years, with the rapid growth of knowledge graphs, knowledge reasoning technology is in great demand for research. The knowledge graph is a heterogeneous network with a graph structure. Graph Convolutional Network (GCN) is an extension of traditional Convolutional Neural Network (CNN) in non-Euclidean space, very suitable for processing complex graph data. In this paper, a attention-based relational graph convolutional network (AR-GCN) is proposed. When aggregating neighbor information, the weight of neighbor nodes is adaptively assigned through the attention mechanism, so that nodes can focus on different neighbor information and enhance the accuracy of feature representation. According to the topological characteristics of different knowledge graphs, two attention mechanisms are proposed. The experimental results show that AR-GCN outperforms R-GCN in entity classification and link prediction tasks, further showing that it has stronger characterization ability.