DEP parameter penalty in LogisticRegression and LogisticRegressionCV #32659
Conversation
ogrisel
left a comment
There was a problem hiding this comment.
Thanks for getting this is in. I just have a concern with the warning strategy for l1_ratios in LogisticRegressionCV (see below).
sklearn/linear_model/_logistic.py
Outdated
| _raise_for_params(params, self, "fit") | ||
|
|
||
| solver = _check_solver(self.solver, self.penalty, self.dual) | ||
| if isinstance(self.l1_ratios, str) and self.l1_ratios == "warn": |
There was a problem hiding this comment.
Are we sure we want to warn by default? I would expect LogisticRegressionCV(Cs=np.logspace(-6, 6, 13)).fit(X_train, y_train) to behave the same in 1.10 and 1.8, and so I don't see the point in warning the user when using the default. The only behavioral change would be sklearn.get_params()["l1_ratios"] which would change from None to (0.0,). I don't know we if we should consider a worse breaking change than from None to "warn".
I think we should only warn when the user explicitly passes LogisticRegressionCV(Cs=np.logspace(-6, 6, 13), l1_ratios=None).fit(X_train, y_train) instead.
There was a problem hiding this comment.
In that case, same as l1_ratio for LogisticRegression. I will add the changes accordingly.
There was a problem hiding this comment.
I'm easy either way, but generally speaking, we don't have the policy of things not warning when using default values. In many cases, the user has to set something to silence the warning.
There was a problem hiding this comment.
At first, I thought so, too. In this particular case, the behavioral change ist almost none such that a „hard“ change of the default value without warning period seems justified.
As I demonstrated, I can implement either way (simply reverting eab936a eould be enough to switch).
I am rather indifferent to the possibilities, just say which one:
In any case, |
doc/whats_new/upcoming_changes/sklearn.linear_model/32659.api.rst
Outdated
Show resolved
Hide resolved
sklearn/linear_model/_logistic.py
Outdated
| l1_ratio : float, default=0.0 | ||
| The Elastic-Net mixing parameter, with ``0 <= l1_ratio <= 1``. Only | ||
| used if ``penalty='elasticnet'``. Setting ``l1_ratio=0`` is equivalent |
There was a problem hiding this comment.
I got a bit surprised with the "only used if penalty='elasticnet'".
Could we move it at the end of the paragraph instead and mention that l1_ratio has an effect with penalty is let as its default or that one force penalty="elasticnet" (but it is deprecated).
| is_logreg_l1_penalized = est_name == "LogisticRegression" and ( | ||
| hasattr(estimator, "l1_ratio") and np.isclose(estimator.l1_ratio, 1.0) | ||
| ) | ||
| is_logregcv_l1_penalized = est_name == "LogisticRegressionCV" and ( | ||
| hasattr(estimator, "l1_ratio_") | ||
| and np.all(np.isclose(estimator.l1_ratio_, 1.0)) | ||
| ) |
Reference Issues/PRs
Partially solves #28711.
This is the the no-brainer of #32042 (comment).
What does this implement/fix? Explain your changes.
This PR
deprecates the parameters penalty
changes the default of l1_ratio from None to 0
deprecates the parameters penalty
changes the default of l1_ratios from None to (0,)
Any other comments?