A HETEROGENEOUS GRAPH NEURAL NETWORK FOR MINING FREQUENT RELATION PATTERNS OF KNOWLEDGE GRAPH

Due to the difficulties in modeling knowledge graph with current mining algorithms, aimed at mining frequent relation patterns and distribution of each structure, a graph neural network model was designed to describe the heterogeneous structure within the scope of nodes. The model took relations as the input of node features, retained the original structure information by using the autoencoder and multi-head attention mechanism, and designed the translation layer of feature structure to map the same structure to the same space, so as to obtain the frequent heterogeneous structure. Experiments show that this model can mine the relation patterns and the distribution of each structure in the graph faster. In addition, it has a stable performance in the link prediction task, which verifies the feature expression ability, and is even better than some joint learning models in heterogeneous graphs with many relationship types.