Spectral measure for multi-class problems

Abstract

Diversity among base classifiers is known to be a necessary condition for improved performance of a classifier ensemble. However, there is an inevitable trade-off between accuracy and diversity, which is known as the accuracy/diversity dilemma. In this paper, accuracy and diversity are incorporated into a single measure, that is based on a spectral representation and computed between pairs of patterns of different class. Although the technique is only applicable to two-class problems, it is extended here to multi-class through Output Coding, and a comparison made between various weighted decoding schemes. © Springer-Verlag 2004.