EHSBoost: Enhancing ensembles for imbalanced data-sets by evolutionary hybrid-sampling

To binary problems, class-imbalanced data are defined as datasets in which the instance number of two classes is extremely different. Classifiers based on undersampling and oversampling were proposed to solve the class imbalanced problems. Techniques based on ensemble of classifiers has been successful for class imbalance problems. In this paper, we proposed an ensemble based approach EHSBoost. EHSBoost is based on EUSBoost, which combines the EUS with Boosting algorithm. The goal of our proposed algorithm is to take advantage of undersampling and oversampling. We consider the usage of hybrid sampling in a supervised manner to improve the accuracy of the weak learner. We have conducted experiments on 16 datasets. These datasets are from the KELL data repository. Three evaluation metrics are used to evaluate the proposed method and the state of art counterparts. Experiments show that, the proposed ensembles algorithm based on evolution hybrid sampling method has significant advantage over other methods for classification of class imbalance datasets.