Hybrid Methods for Class Imbalance Learning Employing Bagging with Sampling Techniques

Abstract

Class imbalance classification has become a dominant problem in supervised learning. The bias of majority class instances dominates in quantity over minority class instances in imbalanced datasets, which produce the suboptimal classification results for classifying the minority class instances. In the last decade, several methods including sampling techniques, cost-sensitive learning, and ensemble methods have been introduced for dealing with class imbalance classification. Among all the methods, the ensemble method performs better in compare with sampling and cost-sensitive learning. The ensemble learning uses sampling technique (either under-sampling or oversampling) with bagging or boosting algorithms. However, which sampling techniques will work better with ensemble learning to improve class imbalance is extremely depend on problem domains. In this paper, we propose two bagging based methods: (a) ADASYNBagging, and (b) RSYNBagging for dealing with imbalanced classification. The ADASYNBagging uses ADASYN based over-sample technique with bagging algorithm. On the contrary, the RSYNBagging uses random under-sampling and ADASYN based over-sample technique with bagging algorithm. RSYNBagging utilizes both under-sampling and over-sampling in alternate iterations and thus incorporates the advantages of both techniques without introducing any extra parameter to tune or increasing time complexity. We have tested the performance of our proposed ADASYNBagging and RSYNBagging methods against existing best performing methods Underbagging, SMOTEBagging on 11 benchmark imbalanced datasets and the initial results are strongly encouraging.