Dzeroski and Zenko (2002) investigate Stacking in the extension proposed
by Ting & Witten (1999). They conclude that, when comparing against other
ensemble learning schemes, Stacking with MLR as meta classifier is at best competitive
to selection by crossvalidation (X-Val) and not significantly better as
some papers claim, while their new variant sMM5 clearly beats X-Val. They
propose a comparative study to resolve these contradictions in the literature.
Seewald & Furnkranz ¨ (2001) propose a scheme called Grading that learns a
meta-level classifier for each base classifier. Grading trains a meta classifier for
each base classifier which tries to predict when its base classifier fails. This
decision is based on the dataset’s attributes. A weighted voting of the base
classifiers’ prediction gives the final class prediction. The voting weight is the
confidence for a correct prediction of a base classifier, which is estimated by its
associated meta classifier