基于关系图注意力网络的链路预测方法

Link Prediction Method Based on Relational Graph Attention Networks

知识图谱已成为人工智能领域热门的研究方向,但大规模知识图谱往往不健全,因此链路预测任务被提出用于补全知识图谱。然而,现有基于图神经网络的链路预测方法忽略了关系特征,导致更新的节点特征不准确。针对上述问题,提出关系图注意力网络(Relational Graph Attention Networks,RGAT)用于链路预测。RGAT不仅学习节点特征,还考虑了关系类型和方向。RGAT将关系假设成头实体到尾实体的转换,利用关系转换操作把知识图谱的异质邻域转换成同质邻域,以便于图注意力网络能够准确传递信息。为了验证方法的有效性,在FB15k-237和WN18RR上进行实验,实验结果表明提出的模型RGAT能进行有效的链路预测。

Knowledge graphs have become a popular research direction in the field of artificial intelligence,but large-scale knowledge graphs are often incomplete.However,existing graph neural network-based link prediction methods ignore relational features,resulting in inaccurate updated node features.Regarding the problem above,Relational Graph Attention Networks(Relational Graph Attention Networks,RGAT)are proposed for link prediction.RGAT not only learns node features,but also considers relation types and directions.RGAT assumes the relation as the transformation from head entity to tail entity,and uses the relation to convert the heterogeneous neighborhood of the knowledge graph into a homogeneous neighborhood,so that the graph attention network can accurately transmit information.To verify the effectiveness of the method,experiments are conducted on FB15k-237 and WN18RR,and results show that the proposed model RGAT can perform effective link prediction.