Knowledge Graph Completion Method Based on Interactively Connected Graph Attention Network

Knowledge graph provides underlying support for many intelligent information service applications, including intelligent search, public safety, finance, medical care and other fields. However, the existing knowledge graph is usually incomplete, and knowledge graph completion has become an urgent problem to be solved. The existing knowledge graph completion method models ignore the important information rich in neighbor nodes and relationships, and often simply splice neighbor nodes and relationships together, ignoring the different importance of different relationships and neighbor nodes to nodes. To solve this problem, this paper proposed a knowledge graph completion method (ICGAT) based on the interactive connection graph attention network. The method firstly finds out the potential relationship by finding two-hop neighbor nodes, and expands the triples of each node. Then it fuses the relationship in each triplet with the features of the node, and adopts the method of interactive connection between nodes and neighbor nodes, using 4 space vectors to represent the interactively connected relationship. Finally, the vector of interactive connection was input into the graph attention network to obtain the weight of relations and neighbor nodes to the node, so as to illustrate its importance. In order to effectively represent triples of complex relationships such as one-to-many, many-to-many, etc. , this method used the RotatE model as a pre-training model. The experimental results in the link prediction task show that the performance of the mean rank and HR@10 indicators of the ICGAT method in the WN18RR and FB15k-237 datasets have been improved to a certain extent, indicating that ICGAT can improve the accuracy of the link prediction task.