Over-Squashing

Over-squashing occurs when message-passing graph neural networks compress information from many distant nodes into a limited number of representations, making long-range dependencies hard to learn. Researchers diagnose bottlenecks using graph curvature, effective resistance, and transport-based measures, then mitigate them through graph rewiring, hierarchical models, or broader communication mechanisms. These methods are tested on long-range graph benchmarks, heterophilous graphs, and spatio-temporal forecasting tasks. A central tension is that improving connectivity can make information mix too quickly, worsening oversmoothing and increasing the need for careful normalization and edge-budget choices.