Graph alternate learning for robust graph neural networks in node classification

The real-world graphs are full of noises and perturbation. However, recent studies show that the existing graph neural networks (GNNs) are usually sensitive to the quality of the input graph. In this work, we propose a graph alternate learning (GAL) framework to alternately train dual models, i.e., prediction network to learn the graph structure for node classification tasks and graph regularization network to enhance the robustness of GNNs. The adoption of dual models, which learn and teach each other collaboratively at the entire training process, drives the formation of a better graph structure. Furthermore, a node feature selection method is integrated into the GAL network to reduce the influence of node attack. Lastly, in order to evaluate the anti-attack ability of GAL, we devise a smooth input graph adversarial attack, called Smooth-Attack , which can degrade the node classification performance of graph convolutional networks (GCN) and is considered to be a form of over-smoothing. Experiments show that our proposed GAL model can keep superiority on benchmark datasets under edge and node perturbation, and GAL is highly robust against Smooth-Attack .