Generating diverse and accurate classifier ensembles using multi-objective optimization

Accuracy and diversity are two vital requirements for constructing classifier ensembles. Previous work has achieved this by sequentially selecting accurate ensemble members while maximizing the diversity. As a result, the final diversity of the members in the ensemble will change. In addition, little work has been reported on discussing the trade-off between accuracy and diversity of classifier ensembles. This paper proposes a method for generating ensembles by explicitly maximizing classification accuracy and diversity of the ensemble together using a multi-objective evolutionary algorithm. We analyze the Pareto optimal solutions achieved by the proposed algorithm and compare them with the accuracy of single classifiers. Our results show that by explicitly maximizing diversity together with accuracy, we can find multiple classifier ensembles that outperform single classifiers. Our results also indicate that a combination of proper methods for creating and measuring diversity may be critical for generating ensembles that reliably outperform single classifiers.