Methods of Learning Classifier Competence Applied to the Dynamic Ensemble Selection

Abstract

In the design of dynamic ensemble selection system (DES) based on competence measure of base classifiers, two steps can be distinguished. The first step consists in calculating the set of competences of a classifier at all points of validation set. In the second step this set is generalized to the whole feature space or - in other words - the procedure of learning competence function using the competence set is performed. In the paper different methods of learning competence function are developed using the concept of randomized reference classifier as a basis for determining the competence set. The methods developed are: potential function method, linear regression, neural network, radial basis neural network, generalized regression neural network and 1-Nearest Neighbor method. Performance of DES systems for different learning methods were experimentally investigated using 9 benchmark data sets from the UCI Machine Repository and results of comparative analysis are presented.