Robust Unsupervised Graph Representation Learning via Mutual Information Maximization

Recent studies have shown that GNNs are vulnerable to adversarial attack. Thus, many approaches are proposed to improve the robustness of GNNs against adversarial attacks. Nevertheless, most of these methods measure the model robustness based on label information and thus become infeasible when labels information is not available. Therefore, this paper focuses on robust unsupervised graph representation learning. In particular, to quantify the robustness of GNNs without label information, we propose a robustness measure, named graph representation robustness (GRR), to evaluate the mutual information between adversarially perturbed node representations and the original graph. There are mainly two challenges to estimate GRR: 1) mutual information estimation upon adversarially attacked graphs; 2) high complexity of adversarial attack to perturb node features and graph structure jointly in the training procedure. To tackle these problems, we further propose an effective mutual information estimator with subgraph-level summary and an efficient adversarial training strategy with only feature perturbations. Moreover, we theoretically establish a connection between our proposed GRR measure and the robustness of downstream classifiers, which reveals that GRR can provide a lower bound to the adversarial risk of downstream classifiers. Extensive experiments over several benchmarks demonstrate the effectiveness and superiority of our proposed method.