Towards Robust Graph Neural Networks via Exclusive Group Sparse Feature Learning

Graph Neural Networks (GNNs) are widely used in graph learning tasks but often suffer from irrelevant and redundant noise, especially particularly in noisy environments, which leading to degrades performance degradation. This challenge underscores the need for effective feature selection to extract meaningful features while suppressing noises. Although attribute selection guided GNNs (AsGNNs) leverage group sparsity to mitigate this issue, they still struggle to effectively learn discriminative features for each class. To address these issues, we propose an exclusive group sparse feature learning method to enhance the robustness of GNNs, dubbed as EGS-RGNNs, which integrates exclusive and group sparsity-inducing regularizations to promote feature sharing, competition, and adaptive selection within each GNN layer. Building on this, we further introduce an efficient proximal gradient optimization algorithm to enable end-to-end training of EGS-RGNNs. Extensive experimental results on benchmark datasets show that EGS-RGNNs outperforms baseline methods in node classification tasks, demonstrating exceptional robustness in noisy environments.