Diversity Measure for Multiple Classifier Systems

Abstract

Multiple classifier systems have become a popular classification paradigm for strong generalization performance. Diversity measures play an important role in constructing and explaining multiple classifier systems. A diversity measure based on relation entropy is proposed in this paper. The entropy will increase with diversity in ensembles. We introduce a technique to build rough decision forests, which selectively combine some decision trees trained with multiple reducts of the original data based on the simple genetic algorithm. Experiments show that selective multiple classifier systems with genetic algorithms get greater entropy than those of the top-classifier systems. Accordingly, good performance is consistently derived from the GA based multiple classifier systems although accuracies of individuals are weak relative to top-classifier systems, which shows the proposed relation entropy is a consistent diversity measure for multiple classifier systems.