Use check_array to validate y

[betatim]

Reference Issues/PRs

closes #25073 (more precisely this PR combined with #25080 closes it)

What does this implement/fix? Explain your changes.

Uses check_array in _check_y so that we get the same behaviour for converting pandas special data types (that can represent missing values) like Int64 as for X. This is done in the part of the code around pandas_requires_conversion.

Is this what you had in mind?

cc @glemaitre

[adrinjalali]

Would this hack still work after this change?

https://stackoverflow.com/questions/49581104/sklearn-gridsearchcv-not-using-sample-weight-in-score-function/49598597#49598597

[betatim]

I've not tried to run the code from the SO answer, so technically "I don't know". From a quick read of the code it seems that it relies on the y being passed around as a data frame so that it can extract indices from it and then use those to index the sample weights.

From a super high level view, this change replaces a y = np.asarray(y) with y = check_array(y). So I assume that the SO answer should still work. Because np.asarray() would also remove the "data frame-ness".

[jjerphan]

The CI jobs fail because some tests have y be one-element arrays.

One solution is to adapt check_array regarding

scikit-learn/sklearn/utils/validation.py

Lines 926 to 933 in cbfb6ab

	if ensure_min_samples > 0:
	n_samples = _num_samples(array)
	if n_samples < ensure_min_samples:
	raise ValueError(
	"Found array with %d sample(s) (shape=%s) while a"
	" minimum of %d is required%s."
	% (n_samples, array.shape, ensure_min_samples, context)
	)

probably by:

• changing the default value of ensure_min_samples to 0
• changing the check for ensure_min_samples > 1

Alternatively, we can change this call to:

Suggested change

	y = check_array(
	y, ensure_2d=False, dtype=dtype, input_name="y", force_all_finite=False
	)
	y = check_array(
	y,
	ensure_2d=False,
	ensure_min_samples=0,
	dtype=dtype,
	input_name="y",
	force_all_finite=False,
	)

The first proposal is more sensible to me while probably necessitating more changes and adaptation in cascade whilst the second proposal is relatively simple but probably semantically not always correct.

[betatim]

For example one test test_classification_report_output_dict_empty_input checks that you can call classification_report with [] for both y_true and y_pred. So I think we need to allow ensure_min_samples=0.

[jjerphan]

LGTM given that tests pass.