KNOWLEDGE GRAPH COMPLETION METHOD BASED ON IMPROVED GRAPH CONVOLUTION NEURAL NETWORK

Knowledge graph completion aims at mining and predicting hidden entity relationships in knowledge graph. Most traditional methods use graph convolutional neural network to complete knowledge graph. However, the traditional methods mainly use the information of neighbor nodes to update the central node, ignoring the relationship information to some extent. Therefore, an improved graph convolutional neural network is proposed as an encoder. It combined neighbor nodes and relationships by using cyclic correlation strategy, and used attention mechanism to distinguish the degree of their contribution to the central node. InteractE model was used as the decoder to achieve end-to-end network training. The experiments on WN18RR, FB15K-237 and YAGO3-10 data sets show that, compared with the current best model, the MRR and Hits@N indexes are improved to a certain extent.