Robust Classification, Regression and Clustering.

[TimotheeMathieu]

This update the robust module.

• Fix the PEP8 violation. Closes PEP8 violation #69
• Change from RobustWeightedEstimator to the three classes RobustWeightedClassifier, RobustWeightedRegressor, RobustWeightedKMeans . Each class does one task.
• Fix examples, doc and test in consequence

EDIT : I also added Huber loss because it is more robust than squared loss and it makes more sense in robust regression.

[rth]

Thanks @TimotheeMathieu and sorry for slow review. I'll do a pass now. If you don't mind I will merge upstream/master in, which should resolve the CI issues..

[rth]

Thanks @TimotheeMathieu ! This is a nice improvement, and the split of RobustWeightedEstimator into separate regressor, classifier, and clusterer classes, makes indeed them easier to understand. We can easily do that since the added RobustWeightedEstimator was not yet par of a release.

I have a few comments below. Also it would be good to makes sure following parts are run in unit tests (either by adding new tests or extending existing tests to cover them e.g. using pytest.mark.parametrize)

• weighting='mom' and k=None
• check that classifier with another loss than log raises an error

  def test_robust_no_proba():
      ...
      est = RobustWeightedClassifier(loss='hinge').fit(X, y)
      msg = "Probability estimates are not available"
      with pytest.raises(AttributeError, match=msg):
          est.predict_proba(X)

Finally a few minor fixes of types: typos_diff.txt. Otherwise changes look good I think.

[TimotheeMathieu]

I followed your recommendations @rth, thank you for the review, thanks to you there is a better test coverage.

[rth]

Thanks!