Solving the class imbalance problems using RUSMultiBoost ensemble

Abstract

A data set is considered imbalanced when its class representation is substantially different. Examples of rare class are infrequent and cost more than common class examples in binary class imbalance data sets. Common learners usually incline toward common class and rare class examples are missed due to class imbalance. Ensemble learning approach combined with data resampling gains popularity to solve class imbalance problem, recently. RUSBoost and SMOTEBoost are two such methods that combine data resampling techniques with boosting procedure. We propose RUSMultiBoost, a hybrid method that is constituent of MultiBoost ensemble and random undersampling (RUS) to solve the class imbalance problem. Our new method is as simple as RUSBoost but more efficient and effective. We test our method on twelve data sets for class imbalance problem and compare the performance with simple and advanced hybrid ensemble methods. Experimental results show that our hybrid ensemble method performs significantly better than other methods on benchmark data sets using G-mean, Sensitivity and F1-measure. In addition, our method is also suitable for parallel execution as contrast to other boosting methods.