Compatability issue with TransformeredTargetRegressor and RFECV

Description

Based on this SO post.

TransformedTargetRegressor is not working with RFECV because coef or feature_importance is not exposed directly in TransformedTargetRegressor.

Steps/Code to Reproduce

from sklearn.datasets import make_friedman1
from sklearn.feature_selection import RFECV
from sklearn.linear_model import LinearRegression
from sklearn.compose import TransformedTargetRegressor

X, y = make_friedman1(n_samples=50, n_features=10, random_state=0)
log_estimator = TransformedTargetRegressor(regressor=LinearRegression(),
                                                func=np.log,
                                                inverse_func=np.exp)

log_selector = RFECV(log_estimator, step=1, cv=5, scoring='r2')
log_seletor = log_selector.fit(X,y)

Expected Results

No error is thrown

Actual Results

RuntimeError: The classifier does not expose “coef_” or “feature_importances_” attributes

[jnothman]

Firstly note that there is a parallel issue open for SelectFromModel (#15154) the solution might be #9606.

What I’m proposing as a simple solution is that RFE and SelectFromModel add a param importance_getter which can be a callable ((fitted_est,) -> importances), and otherwise defaults to the current logic.

[venkyyuvy]

What I’m proposing as a simple solution is that RFE and SelectFromModel add a param importance_getter which can be a callable ((fitted_est,) -> importances), and otherwise defaults to the current logic.

I like this idea. Can I create a PR for this?