Pair discriminability/concordance as a way of assessing embedding/model

[bdpedigo]

Discriminability using pairs as the "groups": Code for discriminability here https://hyppo.neurodata.io/reference/discrim.html Code for omni on maggot graph here https://github.com/neurodata/maggot_models/blob/9f24b3239bb0cc426c8556b124c75a612b9af34f/notebooks/137.7-BDP-omni-clust.py#L690

Experiment:

• compute discriminability from a distance matrix in the embedded space (euclidean or cosine)
• sweep over different params and compute discriminability stat
  • first embedding dimension (for omni)
  • last embedding dimension (final SVD)
  • whether or not to do the graph matching step first
    • look at whether pairs that are graph matched instead of real are less likely to be close to "pair"

[bdpedigo]

could be used as a semi-principled way of doing dimension selection on this data? does it say anything about what running nonpar/semipar would yield?

[bdpedigo]

feels like there are n different RDPG models, and this could say which of them provide a level of coarseness where pairs are the same?

[bdpedigo]

Jovo says https://pubmed.ncbi.nlm.nih.gov/10458943/ "used hemisphere concordance to evaluate methods'