feat: Enable EstimatorTransformer.get_feature_names_out via ClassNamePrefixFeaturesOutMixin mixin

[DMZ22]

Description

EstimatorTransformer inherits TransformerMixin, which does not supply get_feature_names_out, so the method simply isn't there:

>>> EstimatorTransformer(LinearRegression()).get_feature_names_out(None)
AttributeError: 'EstimatorTransformer' object has no attribute 'get_feature_names_out'

The practical symptom is the pipeline case, which fails even after fitting:

>>> make_pipeline(StandardScaler(), EstimatorTransformer(LinearRegression())).fit(X, y).get_feature_names_out()
AttributeError: Estimator estimatortransformer does not provide get_feature_names_out.

The transformer's output is not one-to-one with the input columns — transform returns one column per target for a multi-output estimator and reshapes to a single column otherwise — so OneToOneFeatureMixin would be wrong here. The sklearn convention for this shape is ClassNamePrefixFeaturesOutMixin plus a _n_features_out, which is what this adds; n_outputs_ is recorded during fit.

single output   -> ['estimatortransformer0']
two targets     -> ['estimatortransformer0', 'estimatortransformer1']
in a pipeline   -> ['estimatortransformer0']
before fit      -> NotFittedError   (rather than AttributeError)

In every case the number of names equals transform(X).shape[1], which is asserted in the tests rather than assumed. A side benefit is that set_output(transform="pandas") now yields a properly-named DataFrame instead of failing.

@koaning — this is the change you green-lit on #533 back in 2022 ("Yes please!"); @CarloLepelaars offered then but no PR followed, and nothing open touches this file, so I hope it's helpful rather than stepping on anyone.

Fixes #533

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)

Checklist:

• My code follows the style guidelines (ruff) — ruff check and ruff format --check both clean
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation — the method is inherited API, so nothing in the docs asserts its absence; happy to add a line if you'd like one
• I have added tests that prove my fix is effective or that my feature works
• I have added tests to check whether the new feature adheres to the sklearn convention — the existing parametrize_with_checks suite still passes, and the names/width equality is asserted
• New and existing unit tests pass locally with my changes

tests/test_meta/ goes from 1073 passed to 1110 passed with the same 15 pre-existing collection errors (missing pyarrow locally). Reverting only estimator_transformer.py while keeping the tests fails 37 of them, so they do guard the change.

---

One thing I found while measuring the output width, which I have not touched because it is a separate bug and a behaviour change:

transform does output.reshape(-1, 1) whenever multi_output_ is False. With predict_func="predict_proba" and a 1-D y, predict_proba returns (n_samples, n_classes), so the reshape multiplies the rows:

X has 60 rows, 3-class y
  estimator_.predict_proba(X) -> (60, 3)
  transform(X)                -> (180, 1)      # 60 samples in, 180 rows out

The class docstring records this too — its example is 150 samples with predict_func="predict_proba" and states Shape of transformed data: (300, 1). Happy to open a separate issue, or a PR, if you'd like it fixed.

[CarloLepelaars]

Appreciate you picking this up. I forgot about it

[FBruzzesi]

Thanks for the contribution @DMZ22 - I think the changes are totally fine.

The only issue I see is that ClassNamePrefixFeaturesOutMixin was introduced in scikit-learn 1.2, and we currently allow scikit-learn 1.0+ . @koaning what's your take in bumping the minimum scikit-learn version dependency?

[koaning]

I've been behind on reviews, sorry about that.

I think I am cool with the version bump. It's been a few years at this point and the main thing that's important is that it doesn't increase the maintainance burden for us.