Sort of like #15222 , but for model selection .

For any public object that accepts a random_state parameter, we should document what parts of the algorithm are randomized. It's not always obvious what is and what isn't randomized. We should also always link to the glossary, where the different possible values of random_state are clearly explained.

For example for the random forest estimators, it would be helpful to indicate that random_state determines in particular the subsampling of the samples and the subsampling of the features. Something like:

random_state : int, np.random.RandomStateInstance or None, default=None
Controls the randomness of the estimator, in particular the subsampling
of the samples and the subsampling of the features. See
term:random_state for details.