Knowledge graph relation prediction based on graph transformation

Knowledge graph relation prediction aims to predict the missing relation between entities. Many existing graph neural network (GNN)-based relation prediction models suffer from over-parameterization, and some models cannot effectively learn the correlation between relations for the relation prediction task. In order to solve the above problems, we propose a knowledge graph relation prediction model based on graph transformation. We use two kinds of graph transformation and a parallel fusion model to learn the semantic information, which effectively reduces the number of parameters and reduces the loss of semantic information compared to the Levi graph. Then, we utilize the self-attention mechanism to learn the correlation between relations, and combine it with the DistMult scoring function to complete the relation prediction task. Experiments on four real-world datasets WN18RR, CoDEx-S, Kinship, and FB15K-237 show that our model achieved a better balance between the number of parameters and prediction performance compared to existing GNN-based models on most datasets.