Reproduce “Position: Don't be Afraid of Over-Smoothing And Over-Squashing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Niklas Kormann, Benjamin Doerr, Johannes F. Lutzeyer
This empirical study investigates the practical impact of over-smoothing and over-squashing on Graph Neural Networks across standard benchmarks. It finds that these phenomena may be less critical for performance than widely assumed, suggesting that performance limitations often stem from uninformative receptive fields rather than deep model issues.
Key insights
- Performance degradation in deep GNNs on tested benchmarks is often attributable to uninformative receptive fields, where relevant information is localized in shallow neighborhoods, rather than solely over-smoothing.
- The practical impact of over-squashing, as addressed by current mitigation techniques and measured by graph curvature, does not consistently limit predictive performance on commonly used datasets.
- A paradigm shift is needed in theoretical GNN research to focus on data- and task-centric analysis, quantifying the 'problem radius' and the localization of label-relevant information to guide research efforts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.