RepresentationSearchCV lacks predict_proba, classes_ and classifier tags, so nested roc_auc CV returns NaN #92

Description

RepresentationSearchCV wrapping a classifier does not behave like one. It is a plain BaseEstimator with only predict and score:

• no classes_, predict_proba, predict_log_proba or decision_function;
• its estimator type is unset, so sklearn.base.is_classifier(search) is False.

As a result, cross_val_score(search, X, y, scoring="roc_auc") returns all-NaN scores, with only a UserWarning ("Scoring failed"). Outer cross-validation with an integer cv also uses plain KFold instead of StratifiedKFold. scikit-learn's search meta-estimators (GridSearchCV, RandomizedSearchCV) delegate all of these to the refitted best estimator.

Affected code

• pretab/compose/search.py:34: class RepresentationSearchCV(BaseEstimator).
• pretab/compose/search.py:123-134: only predict and score are delegated.

Reproduction

import warnings
import numpy as np, pandas as pd
from sklearn.base import is_classifier
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import GridSearchCV, cross_val_score
from pretab import RepresentationSearchCV

rng = np.random.default_rng(0)
X = pd.DataFrame({"x": rng.uniform(-3, 3, 300)})
y = (np.sin(X["x"]) + rng.normal(0, 0.3, 300) > 0).astype(int)
search = RepresentationSearchCV(LogisticRegression(), ["minmax", "ple"], cv=3).fit(X, y)
grid = GridSearchCV(LogisticRegression(), {"C": [0.1, 1.0]}, cv=3).fit(X, y)
for name, est in (("GridSearchCV", grid), ("RepresentationSearchCV", search)):
    print(f"{name:23s} is_classifier={is_classifier(est)}, predict_proba={hasattr(est, 'predict_proba')}, "
          f"classes_={getattr(est, 'classes_', None)}")
with warnings.catch_warnings():
    warnings.simplefilter("ignore")
    print("nested CV roc_auc:", cross_val_score(RepresentationSearchCV(LogisticRegression(), ["minmax", "ple"], cv=3),
                                                 X, y, scoring="roc_auc", cv=3))

Output (PreTab 1.0.0 (6bacaba), Python 3.12, numpy 2.5.3, pandas 2.3.3, scikit-learn 1.9.1, scipy 1.18.1):

GridSearchCV            is_classifier=True, predict_proba=True, classes_=[0 1]
RepresentationSearchCV  is_classifier=False, predict_proba=False, classes_=None
nested CV roc_auc: [nan nan nan]

Expected behavior

The same surface as GridSearchCV: predict_proba / predict_log_proba / decision_function / classes_ delegated to best_estimator_ when the wrapped estimator has them, and the estimator type taken from the wrapped estimator.

Actual behavior

The attributes are missing, is_classifier is False, and probability-based scorers fail on every fold.

Impact

Nested evaluation of a representation search, a natural next step after using it, silently produces NaN scores for any probability-based metric. Probability predictions of the selected model are only reachable by transforming with best_preprocessor_ and calling best_estimator_.predict_proba by hand.

Root cause

The class does not delegate the classifier API and does not take its estimator type from the wrapped estimator.

Suggested fix direction

• Delegate with sklearn.utils.metaestimators.available_if, as BaseSearchCV does.
• Expose classes_ after fit.
• Derive the estimator type in __sklearn_tags__ from self.estimator.

Related issues

• RepresentationSearchCV with a classifier uses task="regression" for placement; string labels crash #77: RepresentationSearchCV with a classifier also never passes the classification task to its Preprocessors.
• RepresentationSearchCV cannot use group-based CV splitters and accepts no groups or fit parameters #91: another part of the scikit-learn search API that RepresentationSearchCV does not provide (groups and fit parameters).