The basic idea of the ensemble learning scheme Stacking is to use the predictions
of base classifiers as attributes in a new training set that keeps the original class
labels. Stacking thus utilizes a meta classifier to combine the predictions from
several base classifiers which were estimated via crossvalidation on the training
data. Any classifier may be used as base and/or meta classifier. We shall refer
to the type of meta data consisting of base classifiers predictions as preds.
A straightforward extension of this approach is using class probability distributions
of the original classifiers1 which convey not only prediction information,
but also confidence for all classes. We shall call the meta data of this extension
class-probs. This approach was evaluated and found to be superior to Stacking
with predictions in (Ting & Witten, 1999), provided multi-response linear
regression (MLR) is used as meta classifier.