bench: benchmark SheafNN over simplicial complex-structured datasets #17
Description
Description
This is where most of the novelty in this model lies, its the extension of spectral diffusion approaches to supervised learning on higher dimensional complexes. There are plenty of other attempts at supervised learning over combinatorial complexes like these, so there might be a non-synthetic dataset that appears in the literature worth testing. beyond this some ideas include: brain decoding tasks using fMRI data, mech interp of transformer models, social hierarchy learning tasks, etc. These are just some ideas, but ideally, for this bench, we need:
- labelled data over a fixed simplicial complex structure
- at the very least 2-simplices exist, ideally 3+
- must have a non trivial topology. meaning more than simply a single n-simplex with its respective lower dimensional supports.