Make RidgeCV, LogisticRegressionCV, ... warn when found optimal regularization parameter lies at the boundary of the range

[ogrisel]

I think the RidgeCV().fit(X_train, y_train) should warn the user if the found value for alpha_ is either alphas.min() or alphas.max(). E.g.

StatisticalWarning: the optimal value for the regularization parameter 'alpha' was 0.01 which lies at a boundary of the explored range (between 0.01 and 1.). Consider setting the 'alphas' parameter to explore a wider range. 

We could add a new boundary_warning=True constructor parameter to make it possible to silence the warning.

BTW, the default ranges could probably be extended between 1e-6 and 1e6 with 13 levels on the logspace whenever it is cheap to do so (e.g. for RidgeCV whos current default range much too narrow [0.1, 1., 10.]).

[divyaprabha123]

Hi @ogrisel can I raise a pull request for this?
I also think it would be more appropriate if we extend the boundaries then warn the users.

[rth]

Please do @divyaprabha123 ! I think it would be better to make 2 separate PRs: one to warn for such cases (which makes sense no matter the boundaries) and one to increase the default boundaries in a few models.

[divyaprabha123]

Please do @divyaprabha123 ! I think it would be better to make 2 separate PRs: one to warn for such cases (which makes sense no matter the boundaries) and one to increase the default boundaries in a few models.

Sure! I will do for both

[divyaprabha123]

take

[Reksbril]

@divyaprabha123 are you going to work on the other classes too, or only RidgeCV? In the first case, I've listed the ones that I think would also benefit from the change (in #16783).

[divyaprabha123]

Yes I am! : ). @Reksbril

[bmreiniger]

I think I can finish off the work by @divyaprabha123 toward RidgeCV in #16408. Depending on how that goes I may have a look at the others listed by @Reksbril at #16783.