부산대학교 도서관

This paper deals with the problem of selecting (aka pruning) components for multilabel classifier ensembles. Three methods, one based on a genetic algorithm (GA) and the others based on ranking, are applied for this purpose. In particular, we investigate the adoption of different multilabel classification measures to play the role of the objective function of these methods, paying special attention to how the choice of one measure affects the accuracy of the pruned ensembles as gauged by other measures. A preliminary empirical analysis is conducted on five problems and the results achieved indicate that the classification performance of pruned ensembles as estimated by a given multilabel measure may significantly vary according to the measure adopted for selecting the components. Moreover, although the pruning methods used can yield significant gains in terms of ensemble size reduction, the ensemble models they generate usually lag behind when contrasted to those equipped with the whole sets of components.