Granular concept-enhanced relational graph convolution networks for link prediction in knowledge graph

Link prediction is a task of completing absent triplets by leveraging existing triplets in KG and the ultimate goal is to mitigate the incompleteness and sparsity of KG in terms of content. As well known, Relational Graph Convolutional Networks (R-GCN) model is a promising method for link prediction due to its capability of the graph structure. However, R-GCN mainly relies on information from adjacent nodes, leading to shortcomings in the model for capturing a wider range of relational information. Meanwhile, Formal Concept Analysis (FCA) is increasingly being applied in various fields as an effective data analysis tool. Inspired by this, we integrate FCA into R-GCN to address the shortcomings of R-GCN mentioned above. Specifically, a formal context is first created according to the entities and relations, and granular concepts can be obtained by the formal context. Then the weights of the relational parameters in R-GCN are redistributed based on similarity of granular concepts. Further we develop granular concept-enhanced relational graph convolution networks (GCR-GCN) model, where granular concept can simulate the process of data conceptualization in human brain very well, so it has stronger interpretability compared to the black-box characteristic of R-GCN. Finally, experimental results demonstrate that the GCRGCN model improves the effectiveness of link prediction by effectively assigning different weights to entities with the same relation. In addition, the granular concept effectively improves the computational efficiency of the model and reduces the storage pressure of the model during computation.