Improvements to the NoiseWinnow transformer

[oesteban]

What would you like to see added in this software?

As Satra pointed out in nipreps/nireports@4ddb626#r102410149:

it was funny to see this pop up in my email. i would say this was a reasonable approach at the time, but one could parameterize the orthogonality/unrelatedness in more ways than corrcoef/auc. one can bring in information theoretic metrics, independence (via dcor), etc.,. it would also be useful to parameterize the margin (the bounds within which this noise vector is seen as useless) instead of fixing it as i did back then. finally given all the new feature importance methods (lofo, shap, etc.,.) that didn't exist back then, it may be useful to empirically evaluate that this works.

Do you have any interest in helping implement the feature?

Yes

Additional information / screenshots

No response