ENH check_array returns numeric array w/ mixed typed dataframes

[thomasjpfan]

Reference Issues/PRs

Related to #22231

What does this implement/fix? Explain your changes.

This PR allows us to better handle DataFrames with boolean dtypes, when dtype=None. It also refactors the code slightly to make it easier to reason about. With the following snippet:

import pandas as pd
from sklearn.utils.validation import check_array

df = pd.DataFrame(
    {"bool": [True, False, True], "int": [1, 2, 3]},
    columns=["bool", "int"],
)

array = check_array(df, dtype=None)

• On main, the dtype of array is object
• With this PR, the dtype of array is int64.

Another example is:

import pandas as pd
from sklearn.utils.validation import check_array

df = pd.DataFrame(
    {"bool": [True, False, True], "int": [1, 2, 3]},
    columns=["bool", "int"],
)

array = check_array(df, dtype="numeric")

• On main, the dtype of array is float.
• With this PR, the dtype of array is int64.

Any other comments?

I think the semantics of dtype=None is weird when the dataframe has mixed dtypes. There is really no way to "preserve the original dtype". With this PR, the semantics become "use np.result_type to make a good guess".

CC @glemaitre

[glemaitre]

LGTM

[thomasjpfan]

I am merging this PR, because it already has two +1s.