Subgraph Plug-in Boosts up Graph Neural Networks

Message-passing neural networks often collapse into a one-dimensional subspace because repeated neighborhood aggregation amplifies the dominant eigenvector of the normalized adjacency matrix, attenuating local distinctions. In this paper, we theoretically analyze this phenomenon using perturbation theory to trace the eigenvector amplification process and mutual information bounds to quantify the resulting loss of discriminative signals, and propose a lightweight, architecture-agnostic module called subgraph plug-in (SGP) to boost up any graph neural network (GNN) without altering its architecture or incurring significant computational overhead. It selects the top- \(\kappa \) nodes by centrality, extracts their \(\tau \) -hop neighborhoods as interpretable subgraphs, and concatenates the resulting subgraph embeddings with the global representation of baseline GNN. Experiments on 13 benchmarks, covering both classification and regression tasks with diverse backbones, show that the proposed SGP consistently improves the performance of base models in all settings and is competitive with representative regularization-based augmentation methods. On ogbg-molhiv and ZINC, SGP achieves gains of up to 6 ROC-AUC points and 10 to 12% relative MAE reduction, confirming that the plug-in can reliably boost up off-the-shelf GNNs by exploiting highly-central substructures.