GLNBench – Benchmarked GNNs under label noise and found they collapse
Researchers released GLNBench, a benchmark for graph neural networks under label noise, revealing that robustness methods collapse on wide backbones and fail to scale. The study found that geometry-based methods like ERA…