ENH Adds feature_names_out to impute module

[thomasjpfan]

Reference Issues/PRs

Continues #18444
Fixes #21200

What does this implement/fix? Explain your changes.

Adds feature_names_out to the impute module.

Any other comments?

There is an edge case I am concerned about:

from sklearn.impute import SimpleImputer
import pandas as pd
import numpy as np

marker = np.nan
X = np.array(
    [
        [marker, 2],
        [2, marker],
        [6, 3],
        [1, 2],
    ]
)
X_df = pd.DataFrame(X, columns=["a", "indicator_a"])

imputer = SimpleImputer(add_indicator=True)
imputer.fit_transform(X_df)
imputer.get_feature_names_out()
# array(['a', 'indicator_a', 'indicator_a', 'indicator_indicator_a'], dtype=object)

For this edge case, we can add sklearn to the prefix resulting in sklearn_indicator_a?

[lorentzenchr]

@thomasjpfan Thanks for your dedication for feature names!

[lorentzenchr]

What is our rule for when to prefix and when not? The indicator output column of SimpleImputer is prefixed with "indicator_".

[thomasjpfan]

I do not think we have a rule. I went with SimpleImputer prefixing "indicator_" because it needed to distinguish between the imputed features and the indicator.

Looking at this again, I think we can go with MissingIndcator to add the prefix. This way SimpleImputer only needs to combine them.

[ogrisel]

I also think this is fine to preserve the original names for the imputed features.

[ogrisel]

Thanks for the PR! Apart from the previous review comments, I would like to use a more explicit prefix:

[ogrisel]

LGTM!

[lorentzenchr]

LGTM
Remark: The convention introduced here is tor prefix with prefix = self.__class__.__name__.lower().

[lorentzenchr]

Before merging, should we wait for #21334 and then also use _ClassNamePrefixFeaturesOutMixin or _generate_get_feature_names_out?

[thomasjpfan]

Before merging, should we wait for #21334 and then also use _ClassNamePrefixFeaturesOutMixin or _generate_get_feature_names_out?

I do not think we need to wait. This PR is slightly different. MissingIndicator prefixes the input feature names, i.e. missingindicator_myfeature, while #21334 and _generate_get_feature_names_out generates all new names, i.e. pca0, pca1, etc.

[timotk]

FYI: This PR has been merged into the main branch, but is NOT included in the 1.0.1 release. I thought it was included because it was linked in the 1.0.1 release, but the specific commit is actually dropped and not picked:

drop 8f621ad ENH Adds feature_names_out to impute module (#21078)

Took me quite a while to figure that out.

[ademyanchuk]

FYI: This PR has been merged into the main branch, but is NOT included in the 1.0.1 release

It is still not included 1.0.2. Should it be this way?

[timotk]

@ademyanchuk Yes, it should be this way. I believe scikit-learn follows the process of semantic versioning:

1. The first number increases with major new functionality/changes
2. the second number for minor new features/changes.
3. third number is reserved for fixes/patches, so no new features.

[ademyanchuk]

Thank you, @timotk
Make sense, I have already switched to nightly for now 😊

[lorentzenchr]

As info: Scikit-learn does not follow SemVer. But new features as this one, are usually rolled out in a minor version like 1.1.0 or 1.2.0. Bugfix versions like 1.0.1 are for bugfixes only.