New variants of pairwise classification

Abstract

Abstract Pairwise classification is the technique that deals with multi-class problems by converting them into a series of binary problems, one for each pair of classes. Typically, K-class classification rules tend to be easier to learn for K = 2 than for K > 2 – only one decision boundary requires attention. This paper presents new methods for obtaining class membership probability estimates for multi-class classification problems by coupling the probability estimates created by binary classifiers. Classifiers used include linear Bayes normal classifier, Parzen density based classifier, naive Bayes classifier, binary decision tree classifier and random neural net classifier. The accuracy of new pairwise classifiers is examined on some real data sets. The classification errors were estimated by stratified version of 10-fold cross-validation technique, i.e. the training examples were partitioned into 10 equal-sized blocks with similar class distributions as in the original set. The validation technique was repeated 10 times for each data set.