Presenting an efficient selection method for dynamic classifiers

Abstract

Accurate classification of multiclass data with complex distribution is still a challenge. One of the effective ways to deal with complex data is to design a dynamic classifier in which for a given test sample, the selected ensembles of classifiers (EoCs) are changed to provide the maximum accuracy. Nevertheless, the performance of dynamic classifiers is highly dependent to its selection mechanism, which sometimes selects improper EoCs. This study aims to present an efficient selection procedure to enhance the performance of the dynamic classification by assignment of competence value to the labeling process. Prior to the classification phase, the probability distribution of samples for each class is separately estimated and an importance value (weight) is assigned to each sample according to its probability on the distribution. Next, the weight of samples is incorporated to the training phase and to find a better set of EoCs, multi-objective genetic algorithm is used to select those with higher accuracy simultaneous with diminishing the complexity. To evaluate the proposed scheme, nine datasets which coverall different aspects were driven from UCI database. Empirical results imply the superiority of the proposed method compared to the conventional dynamic selection methods.