Make class_weights a constructor parameter for all classifier models about scikit-learn

[agramfort]

+1 on this. I've recently faced the pb.

[ogrisel]

Then I'll try to play the "Gael Manipulation Trick" (tm):

"If you write the patch, I will do the code review and the merge."

[agramfort]

I should have kept my mouth shut ... however i don't need it now. So you might need to find another volunteer.

[GaelVaroquaux]

+1 one for class parameter, -1 for auto as default, and +1 for 'Gael manipulation tricks'.

[amueller]

I'm not sure but wouldn't this be inconsistent?
The estimator does not know how many classes there are, right?
So if you initialize with (.2, .8) class weights and give the estimator a 3-class problem, it can only raise an error.
I think this is weird. As far as I understood the sklearn API, constructor parameters don't limit the kind of dataset one can use afterward. Is that not correct?

[GaelVaroquaux]

On Wed, Jan 04, 2012 at 11:41:21AM -0800, Andreas Mueller wrote:

So if you initialize with (.2, .8) class weights and give the estimator a 3-class problem, it can only raise an error.
I think this is weird. As far as I understood the sklearn API,
constructor parameters don't limit the kind of dataset one can use
afterward. Is that not correct?

Yes. However, 90% of the use cases of class weights are either to set it
to 'auto', or for people who know how many classes they are working for.

The best rule of thumb that I have for deciding if something needs to be
a fit parameter is: do I need to set it to different values at each
fold in a cross-validation. So, applying this rule, a sample weight can
only be a fit parameter, whereas a class weight should be a class
parameter.

The reason for this rule of thumb is practical: it enables to put models
in something that the 'GridSearchCV'.

Gael

[amueller]

Most often this parameter is called "class_weight". I think it should be "class_weights".
More API deprecation warnings?

[GaelVaroquaux]

Most often this parameter is called "class_weight". I think it should be "class_weights".

Is it worth breaking the API?

[amueller]

I thought it was sometimes called class_weights. But that was only in a test. If it is consistently 'class_weight' then it's not worth breaking the API.
My bad.

[agramfort]

I would stick to class_weight (no s). we usually avoid plural (e.g.
sklearn.linear_model rather than linear_models)

[amueller]

In my current PR, I moved the parameter from fit to __init__. Do you think it should be present in both?
This is currently the case for SGDClassifier (and Perceptron?). Shouldn't there be only one way ;)

This still has to be done for some linear_models.

[amueller]

Fixed :)