Using Bayesian networks for selecting classifiers in GP ensembles

Ensemble techniques have been widely used to improve classification performance also in the case of GP-based systems. These techniques should improve classification accuracy by using voting strategies to combine the responses of different classifiers. However, even reducing the number of classifiers composing the ensemble, by selecting only those appropriately “diverse” according to a given measure, gives no guarantee of obtaining significant improvements in both classification accuracy and generalization capacity. This paper presents a novel approach for combining GP-based ensembles by means of a Bayesian Network. The proposed system is able to learn and combine decision tree ensembles effectively by using two different strategies: in the first, decision tree ensembles are learned by means of a boosted GP algorithm; in the second, the responses of the ensemble are combined using a Bayesian network, which also implements a selection strategy to reduce the number of classifiers. Experiments on several data sets show that the approach obtains comparable or better accuracy with respect to other methods proposed in the literature, considerably reducing the number of classifiers used. In addition, a comparison with similar approaches, confirmed the goodness of our method and its superiority with respect to other selection techniques based on diversity.