MNT Handle NaNs in scipy dev rankdata

[lesteve]

One of the failure in the scipy-dev build #23626 (turns out it fixes all the scipy-dev failures for some reason ...)

This is a change in scipy 1.10.dev: scipy/scipy#16140

import numpy as np

from scipy.stats import rankdata

print(rankdata([1, 2, np.nan]))

scipy 1.10.dev: array([ 1., 2., nan])
scipy 1.9: array([ 1., 2., 3.])

Note: the change in scipy breaks backward-compatibility for the rank in the cv_results_ attribute. For nan scores, the associated rank will be np.iinfo(np.int32).min i.e. -2147483648. I am not sure how much the ranks are used in general (I would guess not used very much) and how acceptable such a breaking change is.

[thomasjpfan]

Thank you for the fix. LGTM

[thomasjpfan]

I removed the "Breaking Change" label because this preserves the behavior in scikit-learn when SciPy 1.10 is released. In a sense, this PR is fixing something in scikit-learn that would have been broken.

[jjerphan]

OK, I added them regarding @lesteve's last remark:

Note: the change in scipy breaks backward-compatibility for the rank in the cv_results_ attribute. For nan scores, the associated rank will be np.iinfo(np.int32).min i.e. -2147483648. I am not sure how much the ranks are used in general (I would guess not used very much) and how acceptable such a breaking change is.

[thomasjpfan]

As for @lesteve's comment, we can think about backporting this change to 1.1.3 so that 1.1.X works with SciPy 1.10 with the same behavior as SciPy 1.9. Note, by the time SciPy 1.10 is released, I suspect we would have already released 1.2, which would include this PR.

[ogrisel]

Thanks for the fix @lesteve. Let's add a changelog entry targeting scikit-learn 1.2.

I we ever decide to backport this to 1.1.3 we will move the entry then.

[lesteve]

OK I have added a what's new entry

[ogrisel]

Thanks for the fix!

[lesteve]

Great, the scipy-dev build may be green tomorrow morning which hasn't happened in a while 🤞