We seek to modeling hierarchical relations in biological data by representing the data as a hypergraph, and learning simultaneous, joint embeddings for the ndoes and hyperedges. In particular, this allows us to update our node features using higher-order relations present in our data, and this allows us to learn representations of the higher order relations themselves.
In the context of spatial data, the higher order information is defined by encoding spatial adjacency/proximity in hyperedges.
We develop a hypergraph scattering transform, and support learning the appropriate hyper-diffusion scales in our wavelets.
TODO