Theoretical study of the relationship between diversity and single-class measures for class imbalance learning

This paper presents the theoretical research about the relationship between diversity of classification ensembles and single-class measures that are commonly used in class imbalance learning. Although there have been studies on diversity and its links to overall ensemble accuracy, little work has been done on the impact of diversity on single-class performance measures in class imbalance learning. The study of class imbalance learning is important, because many real-world problems, such as those in medical diagnosis, fraud detection, condition monitoring, etc., have imbalanced classes, where a minority class is usually more important and interesting than the majority class. In order to gain a deeper understanding of ensemble learning for imbalanced classes, this paper studies the impact of diversity on single-class performance measures theoretically and empirically. One of the main objectives of this paper is to find out if and when ensemble diversity can improve the classification performance on the important (minority) class. © 2009 IEEE.