evaluation of different features about squidpy HOT 6 CLOSED

I have implemented an exemplary notebook of how to do this in the feature_evaluation branch

from squidpy.

List of feature evaluations to implement

• embedding and clustering evaluation of features
  -> to analyse if img features have the information to cluster spots/crops

  • compare embeddings/clusterings of different features + on different crop scalings
  • compute silhouette scores

• hyper parameter search over features, crop params, ..?.. for "optimal" clustering
  -> get best features or crop options for optimising silhouette score or nmi with expression based clustering as reference
  -> measure how good our clustering and how similar our clustering to expression clustering can be

• visual comparison between img feature and gene expression clusterings (mostly finished by Hannah, repeat with more features)
  -> to analyse if img crops provide similar information as gene expressions

  • see feature_evaluation.ipynb

• visual feature sanity check (Louis)
  -> check if computed feature results are reasonable

  • plot example crops for different feature values (crops for minimal cosine similarities)
  • keep the function flexibel such that you can explore easily.

• entropy of features
  -> measure that tells us if single features are informative,
  -> you could also select features based on that measure

• correlation between features
  -> identify redundant features and correlative clusters of features
  -> could also interesting for selecting a smaller set of interesting features

from squidpy.

if this is done, can you close this @LouisK92 ?

from squidpy.

not yet done, I am working on this with an example notebook.

from squidpy.

I think we can close it @hspitzer ?

from squidpy.

yeah, we can close it. There are some nice ideas here that we did not implement, because we decided to forgo the entire feature evaluation thing. Maybe we'll come back to this in the future - but certainly not now.

from squidpy.