Comparison of fusion algorithms based on logistic model of correlated classifiers
Sunghwan Sohn, Youna Kim · 2004
Abstract- This study compares the classification ability of various fusion algorithms (average, majority vote, median, max/min) when individual classifiers are potentially correlated. A logistic transformation of multivariate normal distribution (MVN) is used to generate the posterior probability estimates, assuring that the probability exists between 0 and 1. With varying parameters of MVN and number of classifiers, we measure the relative performance of the fusion algorithms to that of single classifier. Our results can be utilized for the selection of the most effective fusion method for given situation.