mapping predict_proba columns to class labels

[philrhinehart]

When building a model (e.g. SuperLearner) with proba=True on the meta estimator, how can I access the class labels for the model.predict() output columns.

sklearn classifiers generally expose a classes_ property with the class labels, however this doesn't appear to be available on the mlens ensemble classes.

What is the preferred way to map the prediction output columns to class labels?

[flennerhag]

Currently, ml-ensemble only handles numerical class labels. In this case, the class label corresponds to the column index, so there shouldn't be any need to access a classes_ attribute. For instance, an ensemble of two estimators with 2 class labels would have the following column structure:

col=0	col=1	col=2	col=3
est=0	est=0	est=1	est=1
label=0	label=1	label=0	label=1

If you really want to access the classes_ attribute of an estimator, you can with

sl.layer_1.learners[0]._learner_[0].estimator.classes_

Adjust for the layer and estimator you want to look at. Note that the backend structure may change so this way of access the classes_ attribute can break without warning.

[philrhinehart]

Ok, thanks! I had a feeling this was the case but wasn't sure if it was a guarantee.