@NicolasHug I could give it a try. Furthermore, should _check_sample_weight also guarantee non-negativeness and sum(sw) > 0 ?

I think for the above mentioned estimators @NicolasHug intended this as an easier refactoring issues for new contributors, but if you want to look into it feel free to open PRs.

(I left-out the linear_model module because it seems more involved there)

@lorentzenchr Your expertise would certainly be appreciated there. As you mention #15438 there is definitely work to be done on improving sample_weight handling consistency in linear models.

Furthermore, should _check_sample_weight also guarantee non-negativeness and sum(sw) > 0 ?

There are some use cases when it is useful, see #12464 (comment) but in most cases it would indeed make sense to error on them. In the linked issues it was suggested maybe to enable this check but then allow it to be disabled with a global flag in sklearn.set_config. So adding that global config flag and adding the corresponding check in _check_sample_weight could be a separate PR.

intended this as an easier refactoring issues for new contributors

In that case, I will focus on _check_sample_weight and on linear models. So new contributors are still welcome 😃

Thanks @lorentzenchr .

BTW @rth maybe we should add a return_ones parameter to _check_sample_weights. It would be more convenient than protecting the call with if sample_weight is not None:...

Should sample_weights be made an array in all cases? I feel like we have shortcuts if it's None often and I don't see why we would introduce a multiplication with a constant array.

Also: negative sample weights used to be allowed in tree-based models, not sure if they still are.

Should sample_weights be made an array in all cases?

No hence my comment above

Oh sorry didn't see that ;)

@fbchow and I will pick up the fix for BaseDecisionTree for the scikitlearn sprint

@fbchow we will do DBSCAN for the wmlds scikitlearn sprint (pair programming @akeshavan)

working on BaseBagging for the wmlds sprint with Honglu Zhang (@ritalulu)

@MDouriez and I will work on the GaussianNB one

Working on BaseGradientBoosting for wimlds sprint (pair programming @akeshavan)

Working on BaseForest for wimlds sprint (pair programming @lakrish)

often the check is within a if sample_weights is not None: so we wouldn't need to add an argument

@fbchow and I will move onto DummyClassifier.

[MRG] Fixed DummyRegressor
test_dummy.py passed
PR here

Working on the IsotonicRegression one (notice that the isotonic_regression function doesn't have a validating step for sample_weight, so no need to work on that one)

Picking up KernelDensity with @fbchow

In case people are wondering the ones left up for grabs are

CalibratedClassifierCV
KernelRidge
BaseDiscreteNB

We now need a github feature to prevent an issue from being closed by PRs..

@NicolasHug I've updated the list of the estimators already generalized: I think that the issue will be more readable if the list is on the top of the page. Do you think you could find a minute to update it? Thanks a lot!

Done, thanks @cmarmo

can take KernelRidge

We'd like to try to tackle this one for BaseDiscreteNB. @gelavizh1 @lschwetlick #ScikitLearnSpint

Now working with @gelavizh1 and @lschwetlick on IsotonicRegression.

@jeremiedbb it looks like everything has been addressed? We can open a specific issue for the linear models. BTW looks like there was already an open PR for IsotonicRegression...

BaseDiscreteNB has not been addressed yet. It's done actually (I reviewed it :/)

BTW looks like there was already an open PR for IsotonicRegression...

Yes I guess it has been overtook a bit quick. I'm closing it