Hybrid sampling with bagging for class imbalance learning

Abstract

For class imbalance problem, the integration of sampling and ensemble methods has shown great success among various methods. Nevertheless, as the representatives of sampling methods, undersampling and oversampling cannot outperform each other. That is, undersampling fits some data sets while oversampling fits some other. Besides, the sampling rate also significantly influences the performance of a classifier, while existing methods usually adopt full sampling rate to produce balanced training set. In this paper, we propose a new algorithm that utilizes a new hybrid scheme of undersampling and oversampling with sampling rate selection to preprocess the data in each ensemble iteration. Bagging is adopted as the ensemble framework because the sampling rate selection can benefit from the Out-Of-Bag estimate in bagging. The proposed method features both of undersampling and oversampling, and the specifically selected sampling rate for each data set. The experiments are conducted on 26 data sets from the UCI data repository, in which the proposed method in comparison with the existing counterparts is evaluated by three evaluation metrics. Experiments show that, combined with bagging, the proposed hybrid sampling method significantly outperforms the other state-of-the-art bagging-based methods for class imbalance problem. Meanwhile, the superiority of sampling rate selection is also demonstrated.