EC-Gate: Expansion Contribution-Aware Gating for Structure-Guided Message Passing in GNNs

Yutong Zheng, Xiangyu Zhao, Danilo Mandic

Keywords: GNN architectures

Abstract: The message passing framework has largely driven the success of GNNs, yet it faces potential limitations: over-squashing, over-smoothing and expressiveness constraints. A promising solution is structure-guided message passing, which leverages the graph structure to guide information flow and better capture long-range dependencies. We present EC-Gate, a lightweight plug-in that leverages Expansion Contribution (EC)—a layer-wise measure of how edges expand the receptive field—to drive group-wise gates that regulate message propagation. By concentrating capacity on structurally critical edges, EC-Gate can improve the sensitivity bound in large hidden dimensions, while limiting overfitting. EC-Gate delivers significant improvements across synthetic and molecular benchmarks. Remarkably, when implemented on a standard GCN backbone, it achieves state-of-the-art performance on PCBA and competitive results on Lipo and AqSol, showing that EC serves as a strong structural prior. Furthermore, the empirical analysis of gate activations reveals how EC-Gate modulates message passing in an anisotropic manner.

Submission Type: Full paper proceedings track submission (max 9 main pages).

Submission Number: 59