This paper introduces S³GNN, a novel method designed to mitigate the Oversqueezing (OSQ) phenomenon in Message-Passing Neural Networks (MPNNs) by efficiently incorporating global information mixing, resulting in significant error reduction and parameter savings.
Message-passing neural networks (MPNNs) frequently face an information bottleneck when capturing long-range dependencies, leading to the Oversqueezing (OSQ) phenomenon. While spectral filtering has shown promise in achieving global information mixing, practical implementation often faces restrictive theoretical assumptions. The authors propose S³GNN, a solution that mitigates OSQ by lightweightly reintroducing omitted components with substantially lower computational complexity, maintaining effective stability constraints. Extensive experiments across diverse domains, including long-range benchmarks, KGQA, and mesh-based fluid dynamics, demonstrate that S³GNN achieves up to an order-of-magnitude error reduction while using up to 50% fewer parameters. The work suggests a practical path toward more efficient and powerful long-range graph learning models.