[MRG] Nearest Centroid Divide by Zero Fix

[trewaite]

Reference Issues/PRs

#18324

What does this implement/fix? Explain your changes.

Checks to see if X features are all 0 variance. If so, raises value error and will avoid divide by 0.

Any other comments?

Not sure if this is desired behavior but seems like best way to handle, let me know if the error message is specific enough.

[cmarmo]

Thanks @Trevor-Waite for your pull request. Do you mind adding a test checking that the error is thrown when necessary?
When you are ready for review, please change [WIP] to [MRG] in the title of your PR. Thanks!

[thomasjpfan]

Thank you for the PR @Trevor-Waite !

[thomasjpfan]

We would been to be careful here with using variance directly:

import numpy as np

X = np.empty((10, 2))
X[:, 0] = -0.13725701
X[:, 1] = -0.9853293

X_means = X.mean(axis=0)
var = (X - means)**2
var = var.sum(axis=0)

var
# array([7.70371978e-33, 1.23259516e-31])

We can use np.ptp instead to look for constant features:

np.ptp(X, axis=0)
# array([0., 0.])

[trewaite]

No problem! Good catch, changed logic to np.ptp(X).sum() == 0 and changed test case to match your example.

[thomasjpfan]

I suspect we still need the axis=0 to do ptp for each feature:

Suggested change

	if np.ptp(X).sum() == 0:
	if np.all(np.ptp(X, axis=0) == 0):

Also avoiding the sum avoids more numerical instability.

[trewaite]

Apologies, pushed before I noticed your review...will change to np.all methodology.

[thomasjpfan]

Please add an entry to the change log at doc/whats_new/v0.24.rst with tag |Fix|. Like the other entries there, please reference this pull request with :pr: and credit yourself (and other contributors if applicable) with :user:.

[rth]

Thank you @trewaite !