A Random Multi-channel Spectral Graph Convolutional Network

Abstract

Graph neural networks (GNNs) have been widely used to process graph data. However, the ability of most traditional graph models to integrate node features and topology is not ideal. In this case, the Adaptive Multi-channel Graph Convolution Networks (AMGCN) model is proposed to redesign the model structure. It can greatly improve the fusion ability of topology and node features in complex graphs with rich information. Nevertheless, this model cannot solve the problems of over-smoothness and non-robustness as the depth of the model increases. To address this issue, we propose A Random Multi-channel Spectral Graph Convolution Network (RMSGCN). It adopts a random propagation strategy to reduce potential noise interference from high-order neighbors. Moreover, this model introduces simple spectral graph convolution (SSGC) to enhance the model's ability in deep convolution. A large number of experiments on five data sets show that our model has better performance. It has better expressive power in deep convolution.