Seewald (2002a) investigates Stacking in the extension proposed by Ting & Witten
(1999). He claims a weakness of this extension which is not apparent in the
original version of Stacking and introduces a new variant, StackingC, in order to
compensate for this weakness. Empirical evidence is given and supports these
claims. An analysis into the reasons for improvement yields some interesting
insights, most notably that the reason for this improvement is not mainly the
dimensionality reduction of the meta dataset, but also the higher diversity of
the class models