Reference Issues/PRs

Fixes #20952.

What does this implement/fix? Explain your changes.

This PR adds TransformedTargetClassifier, a classification counterpart to the regression-oriented TransformedTargetRegressor.

Task list

• Implement base class for both TransformedTargetRegressor and TransformedTargetClassifier
• Deprecate regressor in favor of estimator in TransformedTargetRegressor
• Add TransformedTargetClassifier to api reference
• Check docstring
• Update documentation wherever relevant
• Allow proper input validation
• Enable hyperparameter validation with _parameter_constraints
• Ensure consistency w.r.t. estimator tags
• Enable metadata routing
• Add test in tests/test_metaestimators_metadata_routing.py
• Make sure modified docs renders nicely
• Update changelog
• Add tests
• Update user guide
• Ensure compatibility with multi-label targets

Important points

• I created a base class BaseTransformedTarget in a similar spirit than what is done for the bagging estimators with BaseBagging. However, this introduces a breaking change for TransformedTargetRegressor: the estimator constructor argument is renamed from regressor to estimator. I deprecated the former in favor of the latter.
• While for TransformedTargetRegressor, the transformer is meant to be any sklearn transformer and therefore accept 2d inputs, things are not so clear for TransformedTargetClassifier. Natural transformers in that case would be label transformers: LabelEncoder, LabelBinarizer and MultipleLabelBinarizer, which will warn in case the input is 2d. To avoid that I can make sure the input is 1d if input_tags.two_d_array = False.
• As the historical issue Allow for Transformers on y #4143 is now closed, I updated the FAQ. I took the liberty of removing the reference to pipegraph that is referenced in the same section as it has not been maintained in the past 5 years. Please let me know if this is not fine, and I'll add it back.
• Following the discussion in TransformedTargetClassifier #20952, it seems to me the main use case is to provide the possibility of using LabelEncoder along with classifier that do not use it internally (so external to sklearn, eg. XGBoostClassifier). As it only make sense along with 3rd party classifiers, I feel I should probably not try and add an example to the gallery.
• I have added some tests but to make a proper test plan, I need to understand better what are the potential use cases. For example, should I ensure it works for multi-label targets? What are the allowed y like?
• I am happy to update the user guide, but would prefer to get some feedback from maintainers first.

Please don't hesitate to leave any feedback, I am here to learn :)