基于四元数图神经网络的知识图谱嵌入

针对当前大多数知识图谱嵌入方法对实体和关系的表示能力低、难以处理复杂关系的问题，提出一种基于四元数图神经网络的知识图谱嵌入方法，用于解决知识图谱的链路预测问题。该方法为了包含更丰富的关系信息，将四元数引入到知识图谱嵌入中对实体和关系建模，并考虑两者之间的共现关系。模型利用勒维图变换将知识图谱中的实体和关系转换为图网络中的节点，采用两者的共现关系构建图中的边;将四元数图神经网络(quaternion graph neural networks,QGNN)作为编码器模块，学习图节点的四元数嵌入;利用四元数空间内的哈密顿乘积构造评分函数对生成三元组进行排序。实验结果表明，所提模型能够很好地捕捉到实体与关系之间潜在的相互依赖关系，在知识图谱嵌入方面优于现有的嵌入模型。

Knowledge graph embedding based on quaternion graph neural network

Aiming at the problem that most current knowledge graph embedding methods have low representation capabilities for entities and relationships and are difficult to handle complex relationships, a knowledge graph embedding method based on quaternion graph neural network is proposed to solve the link prediction problem of knowledge graph. In order to include richer relationship information, this method introduces quaternions into the knowledge graph embedding to model entities and relationships, and considers the co-occurrence relationship between them. Firstly, the model transforms the entities and relationships in the knowledge map into nodes in the graph network by using the Levi graph transformation, and constructs the edges in the graph by using the co-occurrence relationship between them. Secondly, quaternion graph neural networks (QGNN) is used as the encoder module to learn the quaternion embedding of graph nodes. Finally, the Hamiltonian product in quaternion space is used to construct a scoring function to rank the generated triples. The experimental results show that the proposed model can capture the potential interdependence between entities and relationships, and is better than the existing embedded model in knowledge graph embedding.