A multi-channel attention graph convolutional neural network for node classification

Graph convolutional neural networks (GCNs) introduced the idea of convolution into graph neural networks. It has been widely used in graph data processing in recent years. However, the current GCNs framework is not suitable for the task of handling complex relational graphs. For example, in node classification, too much dependence on node features leads to an over-smoothing phenomenon and high similarity between nodes, which affects the effect of node classification. To address this issue, we provide a new multi-channel attention graph convolutional neural network for node classification called SM-GCN. We have improved the accuracy of node classification through the following two aspects of work. (1) Alleviating the problem of over-reliance on a single feature by learning node features and topological structure node embeddings and applying both combinations. (2) Alleviating the over-smoothing by introducing scattering embeddings of topological structures to achieve band-pass filtering of different signals. Then, use the attention mechanism to apply the critical weights of embedding. Extensive experimental results on multiple datasets based on performance metrics demonstrate that the proposed method has superiority in improving the accuracy from 1 to 10% compared with the cutting-edge method and provides a novel scenario for the problem.