PAC-Bayesian Generalization Bounds for Graph Convolutional Networks on Inductive Node Classification

Abstract: Graph neural networks (GNNs) have achieved remarkable success in processing graph-structured data across various applications. A critical aspect of real-world graphs is their dynamic nature, where new nodes are continually added and existing connections may change over time. Previous theoretical studies, largely based on the transductive learning framework, fail to adequately model such temporal evolution and structural dynamics. In this paper, we presents a PAC-Bayesian theoretical analysis of graph convolutional networks (GCNs) for inductive node classification, treating nodes as dependent and non-identically distributed data points. We derive novel generalization bounds for one-layer GCNs that explicitly incorporate the effects of data dependency and non-stationarity, and establish sufficient conditions under which the generalization gap converges to zero as the number of nodes increases. Furthermore, we extend our analysis to two-layer GCNs, and reveal that it requires stronger assumptions on graph topology to guarantee convergence. This work establishes a theoretical foundation for understanding and improving GNN generalization in dynamic graph environments.