Thanks for the report. Feel free to open a PR to improve the docstring.

@ArturoSbr are you working on this? Would be interested to contribute.

@ArturoSbr also suggested the following:

As an alternative, the resulting dictionary could have an additional key named n_features (or something along those lines) that states how many features each element in the dictionary represents.

I think this would further help usability of this attribute (e.g. by loading it in a dataframe to plot n_features vs mean_test_score with error bars.

We could then update the following example to make it simpler by using pd.DataFrame(rfecv.cv_results_) and not having to magically recompute the x values of the plot with `range(min_features_to_select, n_scores + min_features_to_select):

https://scikit-learn.org/stable/auto_examples/feature_selection/plot_rfe_with_cross_validation.html

@miguelcsilva would you be interested in working in separate PR to implement this?

@miguelcsilva would you be interested in working in separate PR to implement this?

Yes, thanks for that. I'll get started working on a separate PR for this.

@ogrisel while working on this I noticed the following unexpected behavior. If I initialize the RFECV with a min_features_to_select larger than the number of features that I pass to the fit method, I do not get an error (as I expected), but instead a result is returned. See the minimal example below:

from sklearn.feature_selection import RFECV
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import make_classification
from pandas import DataFrame

X, y = make_classification(n_samples=1000, n_features=20, n_redundant=0, n_classes=2, random_state=0)

rfecv = RFECV(
    estimator=LogisticRegression(random_state=0),
    min_features_to_select=21,
    cv=2,
    step=2,
    scoring="precision"
)
rfecv.fit(X=X, y=y)

print(DataFrame(rfecv.cv_results_))

Although I do not know the code base too well, my suspicion is that we start from an array with the value for min_features_to_select already present and append the remaining steps until we reach the maximum number of features. This however means that we get the unexpected behavior described above.

One option would be to raise an InvalidParameterError at the start of the fit method like this:

        if self.min_features_to_select > n_features:
            raise InvalidParameterError(
                "Minimum number of features to select cannot exceed maximum number of available features."
            )

Let me know if you think it is worth throwing an error here or if you have any other ideas? And also if you think this is better treated in the same PR I'm currently working on, or in a new one?

Let's open a dedicated PR for this!