Change default solver in LogisticRegression and replace multi_class with multinomial #11903
Conversation
This reverts commit d06bad1.
…={False,'multiclass',True}
|
I've realised that this still needs a FutureWarning when multi_class='deprecated' and multinomial='compat' and the data is multiclass, to say that default behaviour will change to multinomial logistic... I'm not entirely sure that deprecating |
|
Overall, I think I'm inclining towards @agramfort's position that we should just have
While there's a problem of discoverability, as @amueller suggests, once the change of defaults is completed, it will be simple for most users, whereas the transition proposed here is quite confusing for users, and involves deprecating a parameter with a standard (if problematic name) to be replaced with one where there should be some debate that we do not have time for before release. If we want to deprecate |
|
ok let's do auto. |
This continues the work in #11476, but provides more flexibility to the user in whether or not softmax regression is used, by deprecating
multi_classand adding a setting multinomial in {False, True, 'multiclass'}.Some points of note:
multi_classparammultinom 8000 ialbut could have called itsoftmaxor called the inversebinarizedmulti-class??Closes #11476
Closes #10001
Closes #9997