Abstract
In this paper we investigate the performance of pairwise (or round robin) classification, originally a technique for turning multi-class problems into two-class problems, as a general ensemble technique. In particular, we show that the use of round robin ensembles will also increase the classification performance of decision tree learners, even though they can directly handle multi-class problems. The performance gain is not as large as for bagging and boosting, but on the other hand round robin ensembles have a clearly defined semantics. Furthermore, we investigate whether confidence estimates can be used to improve the accuracy of the predictions of the ensemble. Finally, we show that the advantage of pairwise classification over direct multi-class classification and one-against-all binarization increases with the number of classes, and that round robin ensembles form an interesting alternative for problems with ordered class values.