332 ◾ Kweku-Muata Osei-Bryson
1. Having equal weights for all candidate classiers, such as in bagging.
2. Using the relative performance of the base classier on its training dataset
to determine its weight. us, a classier with lower error (e.g., minimum
square error and error rate) would have a higher weight.
3. Applying data envelopment analysis (e.g., Sohn and Choi 2001).
Ueda (2000) noted that many of the proposed methods involve approach type(b),
often involving the minimum square error, and that while this approach type might
appear to be promising, it has some signicant problems that were also noted earlier
by Breiman (1996):
this approach poses two serious problems in practice: 1) e data is used
both in the training of each ...