Balanced Random Hyperboxes for Class Imbalanced Problems

摘要

A Random Hyperboxes (RH) classifier is a simple but powerful randomization-based ensemble model, including hyperbox-based classifiers used as base learners. Individual learners in this ensemble model are trained on random sub-spaces of both instance and feature spaces. This facet results in a flexible mechanism to form a high-performing classifier competitive with other ensemble models in the literature. Like other machine learning models, however, the RH classifier also faces inefficiency when dealing with class-imbalanced datasets. Meanwhile, data containing highly imbalanced class distributions are prevalent in practical applications. Hence, this paper proposes a new variant of the original RH model, namely Balance Random Hyperboxes (BRH), to bypass this drawback effectively. The proposed method uses an under-sampling strategy to build individual learners instead of the random sampling method employed in the original RH model. The experiment conducted on software fault datasets, which show a highly class-imbalanced property, indicated the proposed method's efficiency compared to the original RH model and other ensemble models.