A framework for the analysis of majority voting

Abstract

Majority voting is a very popular combination scheme both because of its simplicity and its performance on real data. A number of earlier studies have attempted a theoretical analysis of majority voting . Many of them assume independence of the classifiers while deriving analytical expressions. We propose a framework which does not incorporate any assumptions. For a binary classification problem, given the accuracies of the classifiers in the team, the theoretical upper and lower bounds for performance obtained by combining them through majority voting are shown to be solutions of a linear programming problem. The framework is general and could provide insight into majority voting. © Springer-Verlag 2003.