Integrating complementary techniques for promoting diversity in classifier ensembles: A systematic study

Various studies have provided theoretical and empirical evidence that diversity is a key factor for yielding satisfactory accuracy-generalization performance with classifier ensembles. As a consequence, in the last years, several approaches for boosting reasonable levels of diversity have been investigated, ranging from the use of data resampling techniques to the use of different types of classifiers as ensemble components. However, little work has been pursued on the combination of diversity-promoting techniques into a single conceptual framework. The aim of this paper is thus to empirically assess the impact of using, in a sequential manner, three complementary approaches for enhancing diversity in classifier ensembles. For this purpose, simulations were conducted on 15 well-known classification problems with ensemble models composed of up to 10 different types of classifiers. Overall, the results evidence the usefulness of the proposed integrative strategy in incrementing the levels of diversity progressively. •Sequential combination of three diversity-promoting techniques for classifier ensembles.•Thorough assessment on a range of datasets using several types of classifiers and three well-known diversity measures.•Results evidence the usefulness of the proposed strategy in incrementing the levels of diversity progressively.•Generalization error is also significantly improved in some circumstances.