An Important Sampling over Graph Convolutional Network for Community Detection

According to graph theory, the complex network is transformed into graph structure data to process the convolutional neural network efficiently and conveniently. The convolution kernel is defined by extending the convolution operation on the image to the graph structure data, and the convolution layer is used for the graph. Coarse-graining and pooling operations extract features of complex networks with irregular data. When training the network with stochastic gradient descent, an important sampling method is designed to change the distribution of samples to reduce variance, thereby saving gradient calculation time. Experimental results show that this method can obtain better community discovery accuracy with lower computational complexity in social networks, citation networks, and knowledge graph datasets, reducing computational time compared with the existing graph convolutional network. In addition, the memory footprint can be extended to use in large-scale complex networks.