An Experimental Study of the Joint Effects of Class Imbalance and Class Overlap

It has been pointed out that the class imbalance problem is one of the critical areas in classification. Furthermore, existing literatures show that other factors such as class overlap, small disjuncts, and noises will aggravate classification performance when they are combined with class imbalance. In this work, we focus on the joint effects of class imbalance and class overlap, and study binary classification performances of six algorithms under different combination of imbalance ratios and overlap degrees. The experiments corroborate that different types of classifiers show distinct robustness to class imbalance and overlap degree. We arrive the conclusion that essentially the densities of different regions of data space affect the classification performance. In addition, based on observations from our experiments, we infer to changing the densities of different regions of data space should be a good way to address problem of class imbalance and class overlap.