Making the Explainer work with both classification and regression models

[tomaszmf]

The explainer used model.predict() which for classification returns the predicted class label, which is then cast to 0 or 1. The correct approach is to use model.predict_proba() instead.

This fix also adds some unit tests to ensure code quality. Minor PEP related refactoring has also been done.

By Tomasz Rudny, Alistair Garfoot @ MindFoundry (http://mindfoundry.ai)

[bondyra]

I understand that due to lack of explicitly defined interfaces we must call hasattr to check if classifier has probability prediction method.
This check should not be a part of actual algorithm though. I suggest moving this conditional expression to short property, like this:

@property  
def _prediction_function(self):
    if hasattr(self.clf, 'predict_proba'):
        return self.clf.predict_proba
    else:
         return self.clf.predict

This property that returns function could be then called in the explain procedures to get predictions, e.g.:

predictions = self._prediction_function(data)[:, 0]

Doing it this way, we would no longer need to duplicate code in the class.

[pbiecek]

@tomaszmf what about comment from @ bondyra?
Should I merge this PR?

[tomaszmf]

Hi, Yes, I need to address it. Will do next week.

Could you please wait with the merge till then? Thanks