Classification of complex network based on improved hierarchical model

Inhalt:
For the node classification of complex network, the Hierarchical Graph Convolutional Networks for Semi-supervised Node Classifification has achieved good result on this problem, but the complex interactive information between nodes is not mined in that artical, and greatly increase the number of features of the node in the refining operation, which may be resulting in a low classification efficiency. In this article, we put forward the idea of combining the graph coding with feature selection based on hierarchical model to solve this problem. Firstly, the graph coding algorithm is used for the original network to fuse each node with its neighboring nodes to obtain the edge feature vector, and the edge feature vector is aggregated with the node representation matrix obtained from the hierarchical model using the splicing function to form a new node feature matrix. Secondly, a feature selection algorithm is used to reduce the feature dimension of the node, the feature selection autoencoder is used to select the feature to remove the irrelevant or redundant features. The experimental results show that, through our method, the node classification accuracy exceeds benchmark models by 2%-3% on the citation network dataset (Cora, Citeseer, Pubmed) and the knowledge graph dataset (NELL) respectively, and improved the node classification efficiency compared to benchmark models.