`_check_feature_names` raises UserWarning when accessing bagged estimators

Describe the bug

Using the estimators inside a BaggingRegressor to predict data raises a

UserWarning: X has feature names, but DecisionTreeRegressor was fitted without feature names

coming from _check_feature_names, even if the BaggingRegressor or its base estimator (in this case DecisionTreeRegressor) are able to take the feature names into account while fitting.

Steps/Code to Reproduce

from sklearn.ensemble import BaggingRegressor
import pandas as pd
import numpy as np

df = pd.DataFrame(
    {
        "feature_name": [-12.32, 1.43, 30.01, 22.17],
        "target": [72, 55, 32, 43],
    }
)
X = df[["feature_name"]]
y = df["target"]

bagged_trees = BaggingRegressor()
bagged_trees.fit(X, y)
bagged_trees_predictions = bagged_trees.predict(X) # rises no warning
bagged_trees.estimators_[0].predict(X) # rises UserWarning

Expected Results

No warning should be thrown

Actual Results

/home/arturoamor/miniforge3/envs/scikit-learn-course/lib/python3.9/site-packages/sklearn/base.py:438: UserWarning: X has feature names, but DecisionTreeRegressor was fitted without feature names warnings.warn( array([72., 55., 32., 32.])

Versions

System: python: 3.9.5 | packaged by conda-forge | (default, Jun 19 2021, 00:32:32) [GCC 9.3.0] executable: /home/arturoamor/miniforge3/envs/scikit-learn-course/bin/python machine: Linux-5.13.0-1017-oem-x86_64-with-glibc2.31

Python dependencies: pip: 21.1.3 setuptools: 49.6.0.post20210108 sklearn: 1.0.1 numpy: 1.21.0 scipy: 1.7.0 Cython: None pandas: 1.3.0 matplotlib: 3.4.2 joblib: 1.0.1 threadpoolctl: 2.1.0

[jeremiedbb]

And now to 1.3 😃

[glemaitre]

Mentioning the problem to @ogrisel IRL, he was mentioning that this is still a bit annoying. This might not be a blocker for 1.0.2 thought but we could try to fix it.

We have a couple of things to have in mind here: the RandomForest (or other ensemble methods) does not want to validate the data for non-finite values for each underlying estimator because it will be too costly. The data validation thus make sense in the ensemble estimator indeed. However, we need to find a mechanism to go around the issue. One possibility would be to attach the metadata such as (n_features_in_, feature_names_, etc.) to each of the trees. We potentially need to be careful to the bootstrap sampling to have something that make sense.