In recent years, GNN has developed rapidly, and the performance of related models in knowledge graph link prediction tasks has been significantly improved. To explain the performance improvement, the researchers need to extract the subgraph patterns learned by the GNN. However, the interpretation accuracy of existing GNN interpreters in typical multi-relational graph data scenarios such as knowledge graphs has not been verified, and related tools have not yet been implemented, resulting in difficulty in extracting interpretation subgraphs. In response to this problem, this paper proposed a knowledge graph link prediction model that converted a multi-relational knowledge graph into a uni-relational graph. The model combined the entities in the knowledge graph into new nodes and the relationship as a new node. Node features generated a new graph with only a single relationship, and trained a denoising autoencoder on the new graph to obtain link prediction capabilities, and finally used a GNN interpreter to generate subgraph explanations. Experiments on three benchmark datasets show that the relative AUC index of the link prediction model based on single-relational transformation is significantly improved compared with GraIL without transformation. Finally, this paper selected the FB15K-237 dataset to conduct explanatory subgraph extraction experiments, and verified the effectiveness of the model in directly extracting link prediction explanations.

[2023-08-21] Accepted Paper /doc/2023.06.0260/v1

[2024-02-05] Printed Article /doc/2023.06.0260/v2