Problems selection under dynamic selection of the best base classifier in one versus one: PSEUDOVO

Abstract

Class binarization techniques are used to decompose multi-class problems into several easier-to-solve binary sub-problems. One of the most popular binarization techniques is One versus One (OVO), which creates a sub-problem for each pair of classes of the original problem. Different versions of OVO have been developed to try to solve some of its problems, such as DYNOVO, which dynamically tries to select the best classifier for each sub-problem. In this paper, a new extension that has been made for DYNOVO, called PSEUDOVO, is presented. This extension also tries to avoid the non-competent sub-problems. An empirical study has been carried out over several UCI data sets, as well as a new data set of musical pieces of well-known classical composers. Promising results have been obtained, from which can be concluded that the PSEUDOVO extension improves the performance of DYNOVO.