Multi-perspective knowledge graph completion with global and interaction features

Duantengchuan Li,Fobo Shi,Xiaoguang Wang,Chao Zheng,Yuefeng Cai,Bing Li

DOI: https://doi.org/10.1016/j.ins.2024.120438

IF: 8.1

2024-03-09

Information Sciences

Abstract:Knowledge graphs are multi-relation heterogeneous graphs. Thus, the existence of numerous multi-relation entities imposes a tough challenge to the modelling of the knowledge graph. Some recent works represent the property of corresponding entities and relations by generating embeddings. They attempted to identify the missing entities by translation operations or semantic matching. However, the expressiveness of these approaches depends on the entity (relations) embedding. The heterogeneity of entities leads to the difficulty of balancing uniform embedding dimension settings on complex and sparse relational entities, as high-dimensional embedding leads to the overfitting of sparse relational entities, and low-dimensional embedding leads to the underfitting of complex relational entities. We introduce a multi-perspective knowledge graph embedding model with global and interaction features (MGIF) to alleviate these issues. This achieved knowledge transfer from complex relational entities to sparse relational entities through the multi-view features. In particular, to overcome the local limitations of convolution neural networks, the global features shared between entities (relations) and entities (relations) are incorporated in the MGIF. The performance of MGIF is experimentally evaluated on several datasets. The experimental effects demonstrate that MGIF can efficiently model complicated entities and accomplish state-of-the-art complex relationship prediction results on most evaluation metrics.

computer science, information systems