GNN-Oversquashing-Oversmoothing-Risk-Dataset

Simulated GNN information-degradation (oversquashing + oversmoothing) risk for 648 configurations across 6 architectures (plain GCN, PairNorm, MADReg, curvature-based graph rewiring, sheaf diffusion, wide GIN), 3 graph topologies, 3 heterophily levels, and width, grounded in the 2019-2023 GNN depth-limitation literature. Forward-looking task predicts severe degradation at deployment depth from a cheap shallow prototype.

About Dataset

Simulates two documented GNN failure modes and the trade-off between them: curvature-based graph rewiring relieves oversquashing but is formally shown to worsen oversmoothing (increased connectivity → more mixing), modeled explicitly rather than as a strict improvement. 648 configs across 6 architectures, 3 topologies, 3 heterophily levels. Ten sources (2019-2023) including Alon & Yahav's foundational bottleneck paper, Topping et al.'s curvature/rewiring method, Oono & Suzuki's oversmoothing proof, and a paper directly formalizing the trade-off.

Purpose of Dataset

Lets A Researcher Choosing A Gnn Architecture For A Long-range Task Estimate, From A Cheap Shallow Prototype, Whether The Deep Model Actually Needed Will Hold Up.