Knowledge Graph Reasoning with Bidirectional Relation-Guided Graph Attention Network

Abstract

Graph convolutional neural networks (GCN) have demonstrated superior performance in graph data modeling and have been widely used in knowledge inference research in recent years. However, knowledge graph is a heterogeneous multi-relational connected graph with complex interactions between neighboring nodes, and most existing GCN-based methods aggregate neighborhood information with the same importance, which leads to the loss of important semantics in the context. In addition, most GAT-based methods consider the neighborhood as a whole and ignore the direction information of the relationship. To this end, we propose a bidirectional relation-guided graph attention network (BR-GAT), which utilizes a bidirectional self-attention mechanism to compute the importance of neighboring nodes, computes the importance of the neighborhood on the representation of relations through a relation-specific mechanism, and finally fuses the joint propagation of neighboring information to update the representations of entities and relations. We conduct link prediction experiments on three standard datasets, and the results demonstrate that BR-GAT outperforms several state-of-the-art models.