基于双曲图注意力网络的知识图谱链路预测方法

现有的大多数知识表示学习模型孤立地看待每个知识三元组，未能发现和利用实体周围邻域特征信息，并且将树状层级结构的知识图谱嵌入到欧式空间，会带来嵌入式向量高度失真的问题。为解决上述问题，该文提出了一种基于双曲图注意力网络的知识图谱链路预测方法(HyGAT-LP)。首先将知识图谱嵌入到负常数曲率的双曲空间中，从而更契合知识图谱的树状层级结构；然后在所给实体领域内基于实体和关系两种层面的注意力机制聚合邻域特征信息，将实体嵌入到低维的双曲空间；最后利用得分函数计算每个三元组的得分值，并以此作为判定该三元组成立的依据完成知识图谱上的链路预测任务。实验结果表明，与基准模型相比，所提方法可显著提高知识图谱链路预测性能。

Link Prediction in Knowledge Graphs Based on Hyperbolic Graph Attention Networks

Most existing knowledge representation learning models treat knowledge triples independently, it fail to cover and leverage the feature information in any given entity’s neighborhood. Besides, embedding knowledge graphs with tree-like hierarchical structure in Euclidean space would incur a large distortion in embeddings. To tackle such issues, a link prediction method based on Hyperbolic Graph ATtention networks for Link Prediction in knowledge graphs (HyGAT-LP) is proposed. Firstly, knowledge graphs are embedded in hyperbolic space with constant negative curvature, which is more suited for knowledge graphs’ tree-like hierarchical structure. Then the proposed method aggregates feature information in the given entity’s neighborhood with both entity-level and relation-level attention mechanisms, and further, embeds the given entity in low dimensional hyperbolic space. Finally, every triple’s score is computed by a scoring function, and links in knowledge graphs are predicted based on the scores indicating the probabilities that predicted triples are correct. Experimental results show that, compared with baseline models, the proposed method can significantly improve the performance of link prediction in knowledge graphs.