[ChrisW09]

Fixes #92.

RepresentationSearchCV around a classifier did not behave like a classifier: it had no classes_, predict_proba, predict_log_proba or decision_function, and is_classifier(search) was False. Nested evaluation with a probability-based scorer such as cross_val_score(search, X, y, scoring="roc_auc") scored NaN on every fold, and an outer integer cv used KFold instead of StratifiedKFold.

Root cause

RepresentationSearchCV is a plain BaseEstimator that only delegated predict and score to best_estimator_. It did not take its estimator type from the wrapped estimator, and it did not delegate the rest of the classifier API, as scikit-learn's GridSearchCV does.

Fix

• RepresentationSearchCV.__sklearn_tags__ takes estimator_type, classifier_tags and regressor_tags from the tags of estimator.
• predict_proba, predict_log_proba and decision_function transform X with best_preprocessor_ and delegate to best_estimator_. As in BaseSearchCV, they are guarded with available_if (_estimator_has), which checks best_estimator_ after fit and the unfitted estimator before, so hasattr on the search reflects the estimator it wraps.
• classes_ is a property that returns best_estimator_.classes_.
• fit records n_features_in_, and feature_names_in_ when every column label of X is a string (scikit-learn's rule). A refit on input without string labels removes stale names.
• predict, score and the new methods share the transform through _best_representation.

Behaviour changes

• is_classifier(search) and is_regressor(search) follow the wrapped estimator. An outer cross_val_score / cross_validate with an integer cv therefore splits a classifier search with StratifiedKFold, so its outer folds and scores change. The search's own inner splitting is unchanged.
• New public methods predict_proba, predict_log_proba and decision_function (each only when the wrapped estimator provides it) and new attributes classes_, n_features_in_ and feature_names_in_. A search around a regressor exposes none of the delegated methods, and predict and score are unchanged.

Documentation

• RepresentationSearchCV class docstring: the search takes the estimator type of estimator; the Attributes section lists classes_, n_features_in_ and feature_names_in_. Docstrings for the delegated methods and classes_.
• docs/core_concepts/target_awareness.md ("Searching over representations"): new paragraph saying that the search takes the wrapped estimator's type, delegates classes_, predict_proba, predict_log_proba and decision_function, and can be evaluated in an outer stratified cross-validation with scoring="roc_auc".

Tests

• test_search_takes_the_estimator_type: a LogisticRegression search is a classifier and a Ridge search a regressor, and check_cv builds a StratifiedKFold for the former. Fails on main, passes here.
• test_classifier_search_delegates_to_the_best_estimator: on string labels, classes_ is ["neg", "pos"], and predict_proba, predict_log_proba and decision_function equal those of best_estimator_ on the best representation. Fails on main, passes here.
• test_delegated_methods_follow_the_wrapped_estimator: an SVC() search exposes decision_function but not predict_proba, before and after fit; a Ridge search exposes none of the delegated names. Fails on main, passes here.
• test_unfitted_classifier_search_raises_not_fitted: before fit there is no classes_, and predict_proba raises NotFittedError. Fails on main, passes here.
• test_nested_cross_val_score_with_roc_auc: nested cross_val_score with scoring="roc_auc" gives finite scores above 0.9. Fails on main, passes here.
• test_search_records_the_input_features: n_features_in_ and feature_names_in_ after a DataFrame fit; a refit on an array drops feature_names_in_. Fails on main, passes here.

Reproduction

Output of the script from #92 on this branch:

GridSearchCV            is_classifier=True, predict_proba=True, classes_=[0 1]
RepresentationSearchCV  is_classifier=True, predict_proba=True, classes_=[0 1]
nested CV roc_auc: [0.96286526 0.94967532 0.96631494]

Checks

• just test: 1444 passed, 59 skipped, 7 xfailed; coverage 93.72% (gate 90%).
• Minimum dependencies (CI job: Python 3.10, numpy 1.24.4, pandas 2.0.3, scipy 1.10.1, scikit-learn 1.6.0): 1431 passed, 72 skipped, 7 xfailed; scripts/quickstart.py passes.
• just docs: strict build succeeds.
• just check: all pre-commit and pre-push hooks pass (ruff format, ruff lint, prettier, file hygiene, pyright with 0 errors).
• Import direction between subpackages is unchanged.

Merge notes

• Base: main.
• The PR for RepresentationSearchCV with a classifier uses task="regression" for placement; string labels crash #77 (fix(search): default to the classification task for a classifier #115) appends tests to tests/compose/test_search.py after the same class_data fixture, which both branches add identically; git keeps one copy. The appended tests conflict: keep both blocks.
• The PR for RepresentationSearchCV cannot use group-based CV splitters and accepts no groups or fit parameters #91 (fix(search): pass groups to the splitter and fit params to the estimator #126) conflicts in three files. In the imports of pretab/compose/search.py, keep this PR's get_tags and available_if imports and RepresentationSearchCV cannot use group-based CV splitters and accepts no groups or fit parameters #91's _check_method_params. At the final refit in fit, keep RepresentationSearchCV cannot use group-based CV splitters and accepts no groups or fit parameters #91's .fit(x_all, y_arr, **fit_params) followed by this PR's n_features_in_ / feature_names_in_ lines. In docs/core_concepts/target_awareness.md, put RepresentationSearchCV cannot use group-based CV splitters and accepts no groups or fit parameters #91's paragraphs on groups and fit parameters first and this PR's paragraph after them. In tests/compose/test_search.py, merge the two import blocks and keep both test blocks. With this branch, RepresentationSearchCV with a classifier uses task="regression" for placement; string labels crash #77's and RepresentationSearchCV cannot use group-based CV splitters and accepts no groups or fit parameters #91's combined this way, the full test suite passes (1453 passed, 59 skipped, 7 xfailed).
• The two "ple" searches on string labels in these tests pass preprocessor_params={"task": "classification"}, because on main target-aware placement does not get the classification task from the search. Once RepresentationSearchCV with a classifier uses task="regression" for placement; string labels crash #77 lands this is redundant but harmless.
• After the PRs for Preprocessor.get_feature_names_out(input_features) fails after fitting on an ndarray, breaking Pipeline.get_feature_names_out() #65 (fix(preprocessor): accept input_features after fitting on a NumPy array #108) and Preprocessor fitted on a DataFrame rejects a same-width NumPy array at transform (and vice versa) #78 (fix(preprocessor): match array input by position after a named fit #116) land, the Preprocessor records feature_names_in_ itself, and the search could read it from best_preprocessor_. It gives the same value as the rule used here.

This change was previously part of #97, which is closed in favour of one pull request per issue.