Benchmarking Sheaf Neural Networks for Inductive Tasks
Sheaf neural networks transfer to inductive graph tasks, but the strongest matched baselines still beat them.
The paper benchmarks 1,890 controlled SNN runs across 14 inductive datasets. It finds restriction-map design matters most inside the sheaf operator, with general maps preferred. Larger stalk dimensions add capacity, but not better long-range reach. The authors say surrounding GNN architecture choices explain more performance variation than the sheaf-specific design space itself. ArXiv · AI/CL/LG's note
The paper benchmarks 1,890 controlled SNN runs across 14 inductive datasets. It finds restriction-map design matters most inside the sheaf operator, with general maps preferred. Larger stalk dimensions add capacity, but not better long-range reach. The authors say surrounding GNN architecture choices explain more performance variation than the sheaf-specific design space itself. ArXiv · AI/CL/LG's note
score 4