CDGCN: An Effective and Efficient Algorithm Based on Community Detection for Training Deep and Large Graph Convolutional Networks

Abstract

Graph convolution neural network (GCN) has become a critical tool to capture representations of graph nodes. At present, the graph convolution model for large scale graphs is trained by full-batch stochastic gradient descent, which causes two problems: over-smoothing and neighborhood expansion, which may lead to loss of model accuracy and high memory and computational overhead. To alleviate these two challenges, we propose CDGCN, a novel GCN algorithm. Through partitioning of the original graph with a community detection algorithm, the original graph is decoupled into many different communities. The decoupled communities restrict the message passing between certain communities, and the model can be trained by using mini-batch stochastic gradient descent. Meanwhile, to prevent the prediction accuracy of the model from decreasing due to over-decoupling, a shallow full-batch stochastic gradient descent GCN is added to the model. We have conducted extensive experiments on node classification tasks of multiple datasets. The experiments show that the model has excellent performance, especially in the dataset with a higher average clustering coefficient, and can achieve higher prediction accuracy.