Classification by pairwise coupling

We discuss a strategy for polychotomous classification that involves estima ting class probabilities for each pair of classes, and then coupling the es timates together. The coupling model is similar to the Bradley-Terry method for paired comparisons. We study the nature of the class probability estim ates that arise, and examine the performance of the procedure in real and s imulated data sets. Classifiers used include Linear discriminants, nearest neighbors, adaptive nonlinear methods and the support vector machine.