Enhancing techniques for learning decision trees from imbalanced data

Several machine learning techniques assume that the number of objects in considered classes is approximately similar. Nevertheless, in real-world applications, the class of interest to be studied is generally scarce. The data imbalance status may allow high global accuracy through most standard learning algorithms, but it poses a real challenge when considering the minority class accuracy. To deal with this issue, we introduce in this paper a novel adaptation of the decision tree algorithm to imbalanced data situations. A new asymmetric entropy measure is proposed. It adjusts the most uncertain class distribution to the a priori class distribution and involves it in the node splitting-process. Unlike most competitive split criteria, which include only the maximum uncertainty vector in their formula, the proposed entropy is customizable with an adjustable concavity to better comply with the system expectations. The experimental results across thirty-five differently class-imbalanced data-sets show significant improvements over various split criteria adapted for imbalanced situations. Furthermore, being combined with sampling strategies and based-ensemble methods, our entropy proves significant enhancements on the minority class prediction, along with a good handling of the data difficulties related to the class imbalance problem.