基于多粒度注意力网络的知识超图链接预测

在知识图谱中进行链接预测是图谱补全的有效方法,可以有效地改善知识图谱的数据质量.然而,现实生活中的关系往往是多元的,这些包含多元关系的知识图谱可称为知识超图(knowledgehypergraph,KHG).然而,现有的知识超图链接预测模型忽略了多元关系的平等性(多元关系中实体不存在先后关系)与整体性(多元关系缺少一个实体则不成立).针对以上问题,首先提出了一种知识超图多元关系表示模型,可以直接建模知识超图中的多元关系;然后研究了一种基于多粒度神经网络的链接预测方法(hyperedgepredictionbasedonmulti-granular attentionnetwork,HPMG).该模型将关系划分为多重粒度进行学习,从不同粒度联合完成知识超图的学习和预测,充分考虑了知识超图中不同维度多元关系的整体性.接下来,针对HPMG特征融合不充分的问题,提出了基于多粒度注意力网络的知识超图链接预测方法 HPMG+,结合全局和局部注意力,实现了不同特征的有区分融合,进一步提高了模型的性能.最后,真实数据集上的大量实验结果验证了所提方法的效果显著地优于所有基线方法.

Knowledge Hypergraph Link Prediction Based on Multi-granular Attention Network

Link prediction in knowledge graphs is the most effective method for graph complementation,which can effectively improve the data quality of knowledge graphs.However,the relationships in real life are often multiple,thus these knowledge graphs containing multiple relationships can be called knowledge hypergraphs,i.e.,KHGs.Unfortunately,the existing knowledge graph link prediction methods cannot be directly applied to knowledge hypergraphs,and the existing knowledge hypergraph link prediction models ignore the equality (there is no sequential relationship among the elements in a multivariate relationship) and completeness (a multivariate relationship is not valid if it lacks elements) of the real-life multivariate relationships.To address these problems,we firstly propose a knowledge hypergraph multivariate representation model,which can directly model the multivariate relationships in the knowledge hypergraph.Then,we study a multi-granularity neural network-based hypergraph prediction method HPMG,which divides the relations into multiple granularities for learning and prediction from different granularities jointly.Next,to address the problem of inadequate HPMG feature fusion,we propose HPMG⁺ based on multi-granularity attention network for link prediction of knowledge hypergraphs,which combines all and local attention to achieve differentiated fusion of different features and further improves the performance of the model.Finally,extensive experimental results on real datasets verify that the proposed methods significantly outperform all baseline methods in terms of hyper-edge prediction.