[FEA] huber & modified_huber loss

[jeremypoulain]

Hello everyone,

Are there some interests or plans to add the huber , and modified_huber as additional loss for the cuml.SGD class.

I use to work on text classification problems with linear models, and those 2 loss functions are really useful ,and usually outperforms SVM's hinge loss.

Here are the code implementation from scikit learn : (https://github.com/scikit-learn/scikit-learn/blob/7813f7efb5b2012412888b69e73d76f2df2b50b6/sklearn/linear_model/sgd_fast.pyx#L112) (https://github.com/scikit-learn/scikit-learn/blob/7813f7efb5b2012412888b69e73d76f2df2b50b6/sklearn/linear_model/sgd_fast.pyx#L243)

Thanks in advance !

[JohnZed]

Yes, this is a potentially-interesting algorithm and should fit in well with the existing optimizers. It may take another release or two before we get to it, but we'll definitely add it to the list.