Subgraph Plug-in Boosts up Graph Neural Networks

ICLR 2026 Conference Submission16998 Authors

Keywords: Subgraph partitioning, mutual information, perturbation theory, graph neural network, plug-in method

TL;DR: SP partitions and extracts subgraphs around high-centrality nodes to augment GNNs, theoretically preventing rank collapse from repeated message passing and consistently boosting graph classification and regression accuracy across diverse benchmarks.

Abstract: Message-passing neural networks (MPNNs) often collapse into a one-dimensional subspace because repeated neighborhood aggregation amplifies the dominant eigenvector of the normalized adjacency matrix, erasing local distinctions and limiting graph classification performance. 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. Guided by these insights, we propose the Subgraph Plug-in (SP), a lightweight, architecture-agnostic module that 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 any base GNN without altering its architecture or incurring significant computational overhead. Across 11 graph-classification benchmarks and 13 GNN variants, we evaluate each backbone with and without SP, yielding 110 model–dataset pairs; SP improves performance in 94 of 110. Beyond these, on ZINC and OGBG-MolHIV, we conduct head-to-head comparisons against 11 methods, including augmentation modules, recent GNNs, and subgraph-based methods. SP achieves the best results among augmentation and subgraph-based approaches and remains competitive with recent GNNs, supporting its role as a widely applicable, cost-effective plug-in that preserves feature diversity and amplifies discriminative substructures. performance.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 16998