SGCN: A Robust Subsampling-Based GCN for Large-Scale Network with Out-of-Distribution Nodes
提出SGCN框架，通过多子图并行训练和性能感知加权，解决大规模图卷积网络的计算可扩展性和对分布偏移的敏感性问题，在交通、供应链等场景有应用潜力。
Large-scale systems, including transportation networks, supply-chain structures and energy dispatch systems, typically exhibit high-dimensional graph structures with dynamic topology, heterogeneous connectivity and substantial noise. These characteristics create persistent challenges for Graph Convolutional Networks (GCNs) in terms of computational scalability and sensitivity to distributional shifts. To address these limitations, this study introduces Subsampling-based Graph Convolutional Network (SGCN), a unified framework that combines multi-subgraph parallel training with performance-aware weighting for parameter aggregation. By decomposing full-graph learning into a sequence of size-controlled subgraph training tasks, SGCN removes the dependence on global neighborhood expansion and naturally supports large-scale parallel execution across GPUs, which significantly reduces memory consumption and training time. Moreover, the performance-aware weighting adaptively down-weights subgraphs that contain out-of-distribution(OOD) nodes or structural noise, enabling the model to obtain stable global representations without relying on explicit detection procedures. Experiments on the Cora, Pubmed and Amazon Computers datasets show that SGCN achieves predictive performance comparable to or exceeding that of standard GCNs and widely used sampling-based methods, while providing substantial computational efficiency gains. Under a variety of OOD node corruption levels and structural perturbation settings, SGCN demonstrates consistently stronger robustness, particularly in high-noise environments. Overall, SGCN provides a scalable and robust training framework for deep graph models in operations research. It can be applied to problems such as traffic prediction, routing, network flow optimization, supply chain risk analysis, and power grid monitoring. Future work will focus on better subgraph scheduling, adaptive sampling for dynamic graphs, and extensions to heterogeneous, multilayer, and temporal graphs, so as to further improve its use in large-scale industrial systems.