Abstract: Graph Neural Networks (GNN), as a powerful graph representation method, have been widely used in graph representation learning, and have achieved a good result in knowledge graph completion. However, for knowledge graph with various types of nodes and links, most homogeneous graph neural networks fail to consider the heterogeneity of knowledge graph and the interaction and aggregation of rich node information, which brings great difficulties to the design of heterogeneous graph neural networks. To solve the above problems, a Graph Transformation Neural Networks (GTN) model is proposed, which embeds multiple types of nodes and relation into the graph to generate a new graph structure, learn the importance of different meta-paths, and solve the problem of heterogeneity of the graph. The composite information between entity node-relation edge types formed by the knowledge graph embedding method is then aggregated to represent the node features on the new graph structure to generate effective links, complete the available links between unconnected nodes in the original graph, and complete the link prediction task of knowledge graph completion. GTN can learn the importance between nodes and their meta-path based on neighbors, which shows that links between nodes in the graph structure can be generated. In order to alleviate the over-fitting phenomenon caused by the weak generalization ability of the model in the experiment, GTN added the R-Drop structure in the training process. Based on the datasets of the two commonly used link prediction tasks, FB15k-237 and WN18RR, comparative experiments and ablation experiments are conducted on GTN, and the results prove the effectiveness of this method.