Multi-Layer Selector(MLS): Dynamic selection based on filtering some competence measures

Abstract

Multi-Classifier Systems (MCSs) have been widely studied as committee methods for increasing accuracy in pattern recognition. Recently, dynamic selection (DS) techniques have attracted undivided attention as effective types of MCSs in application. In DS techniques, the selection process is accomplished in two steps including measurement and selection. Normally, most DS methods in the literature have tried to focus on approaches adopted based on only one of the two options that are to offer a new measurement or a new selection method. On the other hand, some other papers have proposed frameworks in which both options were applied in combination with each other making use of previously existing measurement methods. For the first time, the idea of combining DS methods via multi-layer selectors is offered in this paper. In specific, two main tasks are performed in each layer of the multi-layer selector. First, the competence-level of classifiers are calculated. Then, the classifiers with high competence-level are passed to the next layer, and this way, best classifiers will be passed to the output in the last layer. During this procedure, some competence criteria are combined to measure the competence-level of classifiers. Furthermore, this paper introduces a novel taxonomy for dynamic selection methods and a new technique, called Probability Based (PB) method, to find the competence level of classifier. Experimental results show that the suggested framework improves the classification accuracy when compared against current state-of-the-art dynamic ensemble selection techniques. The Quade non-parametric statistical test confirms the capability of our proposed method to deal with classification problems.