| Abstract | This report consists of two parts associated with graph neural networks: generalization and graph structural learning. We first study the Rademacher complexity of GNNs, as one of independent-algorithm generalization measurements. In addition, we also give upper bounds of the uniform stability of proximal SGD of $L_p$-regularized GNN, which is also used as generalization ability of some specific algorithm. Importantly, inspired by our theoretical findings, we propose a new graph structure learning to generate a clean adjacency matrix for downstream robust representation and learning. Several experiments over real graph data is implemented to show comparable performances of the proposed method on GNNs. |
