[FEA] Add option to accept sample weight vectors to fit methods

[beckernick]

Is your feature request related to a problem? Please describe.
In sklearn, estimator.fit can (almost always?) accept a sample_weight parameter (defaulting to None) that allows users to pass in a weights vector that determines how much weight each sample should receive (with length equal to the number of samples).

This would be a useful feature for cuML estimators, too. As an example, see the sklearn KMeans documentation

sample_weight : array-like, shape (n_samples,), optional
The weights for each observation in X. If None, all observations are assigned equal weight (default: None)

[JohnZed]

Agreed this will be useful for most estimators. It will be an estimator-by-estimator process to add it, but we could start with linear models and get some commonality there. Not going to make it to 0.9 given current load there, but we'll keep it for a near future release.

[JohnZed]

Priority is for KMeans based on requests

[Denisevi4]

Linear models pretty please?

[JohnZed]

Sorry, this didn't make it to the current release, but we'll add it to the list for an upcoming release.

[JohnZed]

Removing from 0.13 as we've added the k-means specific: #1625

[beckernick]

I think it may be worth re-opening this issue for tracking purposes.

A variety of issues exist requesting the ability to specify observation-level weights for various estimators and primitives. As the implementation may need to vary across estimators, it may make sense to keep these issues separate but linked together like an epic. Perhaps this issue can serve as that link, as it's the most broad and the oldest.

Estimators

• Logistic Regression ([FEA] Support for Weights in Nearest Neighbors #3006 )
• KMeans ([FEA] Support sample weights for KMeans #1625 ) (done single GPU)
• SVM ([FEA] sample_weight for SVM #2222 ) (done single GPU)
• KNN Classifier ([FEA] Support for Weights in Nearest Neighbors #3006 (comment))

Primitives

• contingency_matrix ([FEA] Support for sample weights for contingency_matrix prim #2142).

Additionally, as these are implemented, it will also unblock using the respective estimators inside the sklearn AdaBoostClassifer meta-estimator API (#2401 (comment))

[JohnZed]

Long term definitely viable. We will evaluate in more detail whether it can make it into 0.19 and mark it as P1 or P0 if so.