Online pruning of base classifiers for Dynamic Ensemble Selection

Abstract Dynamic Ensemble Selection (DES) techniques aim to select only the most competent classifiers for the classification of each test sample. The key issue in DES is how to estimate the competence of classifiers for the classification of each new test sample. Most DES techniques estimate the competence of classifiers using a given criterion over the set of nearest neighbors of the test sample in the validation set, these nearest neighbors compose the region of competence. However, using local accuracy criteria alone on the region of competence is not sufficient to accurately estimate the competence of classifiers for the classification of all test samples. When the test sample is located in a region with borderline samples of different classes (indecision region), DES techniques can select classifiers with decision boundaries that do not cross the region of competence, assigning all samples in the region of competence to the same class. In this paper, we propose a dynamic selection framework for two-class problems that detects if a test sample is located in an indecision region and, if so, prunes the pool of classifiers, pre-selecting classifiers with decision boundaries crossing the region of competence of the test sample (if such classifiers exist). After that, the proposed framework uses a DES technique to select the most competent classifiers from the set of pre-selected classifiers. Experiments are conducted using the proposed framework with 9 different dynamic selection approaches on 40 classification datasets. Experimental results show that for all DES techniques used in the framework, the proposed framework outperforms DES in classification accuracy, demonstrating that our proposal significantly improves the classification performance of DES techniques, achieving statistically equivalent classification performance to the current state-of-the-art DES frameworks.