An online local pool generation method for dynamic classifier selection

Abstract

Dynamic Classifier Selection (DCS) techniques have difficulty in selecting the most competent classifier in a pool, even when its presence is assured. Since the DCS techniques rely only on local data to estimate a classifier’s competence, the manner in which the pool is generated could affect the choice of the best classifier for a given instance. That is, the global perspective in which pools are generated may not help the DCS techniques in selecting a competent classifier for instances that are likely to be misclassified. Thus, it is proposed in this work an online pool generation method that produces a locally accurate pool for test samples in overlap regions of the feature space. That way, by using classifiers that were generated in a local scope, it could be easier for the DCS techniques to select the best one for those instances they would most probably misclassify. For the instances that are far from the class borders, a simple nearest neighbors rule is used in the proposed method. In this dissertation, an overview of the area of Multiple Classifier Systems is presented, with focus on Dynamic Selection schemes. The most relevant DCS techniques are also introduced, and an analysis on their effectiveness in selecting the most competent classifier for a given instance in a globally generated pool is presented. Based on that analysis, an online local pool generation scheme is proposed and analyzed step-by-step. The proposed method is then evaluated over 20 classification problems, and the effect of its parameters on performance are analyzed. Moreover, a comparative study with other related methods is performed and the experimental results show that the DCS techniques were more able to select the best classifier for a given sample when using the proposed locally generated pool than when using a globally generated pool. Furthermore, the proposed method obtained a greater accuracy rate in comparison with the related methods for all DCS techniques, on average, and presented a considerable improvement for problems with a high proportion of borderline instances. It also yielded a significant increase in performance compared to most related methods evaluated in this work.