In fact, it looks like this PR is just a really small re-factoring to have two new coefficients accessible.
@jnothman: regarding what you've said in the previous PR (#8353), is it still relevant to benchmark it?

What do you think, @jnothman? :)

Hi @jnothman, this is a friendly reminder for this stalled PR follow-up.
Do you think that it is still worth benchmarking it?

Hi @jjerphan thanks for your pull request. There are linting issues and a conflict that should be solved before review. Also if you think that this is ready for review, do you mind changing the title from [WIP] to [MRG]? Thanks a lot for your patience!

@cmarmo: are you interested in reviewing this PR? 🙂

Hi @jjerphan, thanks for your patience! The documentation for the two new functions is not generated: this is because a reference to the function should be added in the doc/modules/classes.rst file.
To bring more attention to your pull request I believe it could be a good idea to propose the requested benchmarks: scikit-learn contains a lot of lines of code, adding more features adds maintenance costs. Nice benchmarks provide more motivation to feature addition. Thanks for your collaboration.

With this PR, r_regression and abs_r_regression would be the only functions in sklearn.feature_selection that does not return p-values. Given the computational cost of p-values is not high: #17169 (comment), I see the benefit of this PR is the convenience method for absolute value, which is is shorthand for:

from sklearn.datasets import make_regression
from sklearn.feature_selection import SelectKBest, f_regression
import numpy as np

X, y = make_regression(random_state=0)

f_abs_regression = lambda *args: np.abs(f_regression(*args)[0])
X_new = SelectKBest(f_abs_regression, k=10).fit_transform(X, y)

If we want to make abs easier to use, we can add a parameter to SelectKBest:

X_new = SelectKBest(f_regression, k=10, transform_func=np.abs).fit_transform(X, y)

I am wondering if adding 2 new functions worth the maintenance cost for making SelectKBest easier to use.

Thanks for the comprehensive review, @glemaitre. I should have adressed all your requests.

IMO, the value of this PR is the introduction of r_regression, and I see no problem to remove the originally proposed abs_r_regression.

I would go for @thomasjpfan's first suggestion, hence assuming that it is up to users to define the score_func they want SelectKBest to be used with.

r_regression is f_regression without the p value computation. For consistency, are we add going to add anova_classif which is f_classif without the p value computation?

There is also the option of adding a return_p_value to f_regression, which gives the option to compute the p value.

r_regression is f_regression without the p value computation.

f_regression relies on r_regression but return different statistics.

Still, it would make sense allowing a way to specify the computation of the p-values as you propose.
I don't think that I am the best to know what the most suitable choice here is.

I think that we could drop abs_f_regression. Then, I still think that we need r_regression and f_regression as proposed. Additionally, it might be nice to add a keyword return_p_values (True by default to be back-compatible) that allows bypassing the computation if p-values are not needed.

So we can limit the scope for this PR to introduce r_regression only.

Merged! Thanks @jjerphan.

'welcome!

Thanks to @DSLituiev for the original PR!