Skip to content
Book Open access

Over-squashing as Transport Congestion: A Sandpile Dynamics Perspective

Aug 2026 · Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 · pp. 4264-4275 · 0 citations · 5 references

Abstract

Message-passing graph neural networks (MP-GNNs) are widely used for learning on relational data. However, their performance drops on tasks requiring long-range interactions due to over-squashing, where exponential information compression overwhelms fixed-width embeddings. While existing analyses often attribute this to geometric bottlenecks under linear diffusion assumptions, thresholded nonlinearities in GNNs motivate a load-release view akin to Abelian sandpiles. Using the discrete sandpile model as a structural proxy, we show that graph bottlenecks force large stabilization cost, effectively creating zones of high transport congestion. We characterize stabilization-invariant equivalence classes induced by the reduced Laplacian and derive cut-based lower bounds linking bottlenecks to unavoidable stabilization effort. The resulting theory is discrete, whereas our implementation is a continuous vector-valued surrogate. The theory identifies the relevant design factors, namely capacity and cut size. Guided by these insights, we propose a differentiable Sandpile Stabilization Layer (SSL) and congestion-aware objectives designed to redistribute excess load and manage stabilization costs. Experiments on long-range benchmarks, together with congestion and collision diagnostics, show that targeting sandpile-identified bottlenecks mitigates representation collapse and improves over standard baselines. Project Page: https://sandpile-gnn.github.io/

Read PDF

Similar papers

Preprint Aug 2026

Pair-Centric Graph Rewiring for Over-Squashing via Optimal Transport-Guided Communication Alignment

PairAlign is proposed, a pair-centric graph rewiring framework that makes this question explicit through demand-support shortage and introduces an Optimal Transport-guided rewiring mechanism to coordinate the finite edge budget for pair-level structural compatibility and shortage-target coverage.

Yan Wang, Chuan-Xian Ren · 0 citations
#machine learning Open access Jan 2025

DeltaGNN: Graph Neural Network with Information Flow Control

DeltaGNN is introduced, to the best of the authors' knowledge, among the first scalable (featuring linear computational and memory complexity overhead) and generalizable (capable of effectively handling graphs with diverse homophily, density, and topology) architectures for long-range and short-range interaction detection.

Kevin Mancini, Islem Rekik · 2 citations
Jul 2026

Schreier-Coset Graph Rewiring

This work introduces a novel method Schreier-Coset Graph Rewiring, a group-theoretic rewiring method that augments the input graph with a Schreier-Coset graph derived from a special linear group, creating a low-resistance bypass for long-range communication.

Aryan Mishra, Randy Martinez, Lizhen Lin · 0 citations
#graph neural networks Preprint Aug 2026

CoRe-GNN: Multilevel Message passing on Coarsened graphs

CoRe-GNN is proposed, which performs both propagations in parallel at each layer: a coarsened inter-cluster term capturing long-range structure, and a local intra-cluster term preserving per-node discriminability.

Antonin Joly, Nicolas Keriven, Aline Roumy · 0 citations
Open access Aug 2026

Exposition on Over-squashing Problem of GNNs: Current Methods, Benchmarks and Challenges.

Graph-based message-passing neural networks (MPNNs) have achieved remarkable success in both node and graph-level tasks. However, several identified problems, including over-smoothing (OSM), limited expressive power, and oversquashing (OSQ), still restrain the performance of MPNNs. In particular, the latest identified problem, OSQ, reveals that MPNNs generally fail to maintain their learning accuracy with tasks that require long-range dependencies between node pairs. In this work, we present an exposition of the OSQ problem by summarizing its various formulations in the current literature and categorizing existing solutions into three different types. In addition, we also discuss the alignment between OSQ and expressive power plus the trade-off between OSQ and OSM. Furthermore, we summarize the empirical methods proposed by existing works to verify the efficiency of OSQ mitigation approaches, together with illustrations of their computational complexities. Lastly, we identify some open questions that are of interest for further exploration of the OSQ problem.

Dai Shi, Andi Han, Lequan Lin et al. · 0 citations
Preprint Aug 2026

SNAP-tFDP: Massively Scalable Graph Layouts via Sparse Negative Sampling

Comprehensive evaluations on 12 large-scale graphs demonstrate that the proposed negative sampling-based algorithm outperforms state-of-the-art algorithms in neighborhood preservation and cluster separation and reduces memory consumption by 72% on average.

Xin Chen, Shuowei Hou, Yifan Wang et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.