RNS Limitations

The large-scale sampler benchmark uses only three main datasets, leaving some structural explanations tentative. RNS can underperform in structurally extreme settings such as very small dense graphs or highly imbalanced binary-label datasets. RNS is less clearly suited to architectures with global attention. SGFormer under RNS underperforms full-graph training on OGBN-ARXIV and POKEC, plausibly because all-pairs global attention interactions are truncated. Rare predictive patterns may be disrupted by random partitioning.