An asymptotic analysis of some expert fusion methods

We study the asymptotic behavior of three classifier combination methods fo r two-class classification: average, median, and majority vote. Assuming th at the estimates of the posterior probability given by individual classifie rs constitute a sample from a distribution, we show that as the number of i ndividual classifiers becomes large, median and majority will produce the s ame result but average may yield a completely different decision if the dis tribution is not symmetric. (C) 2001 Elsevier Science B,V. All rights reser ved.