A new oversampling approach based differential evolution on the safe set for highly imbalanced datasets

Oversampling method is used to solve the class imbalanced issues. Some existing oversampling methods do not well remove noisy samples and avoid synthesizing noisy samples. Therefore, we propose a new oversampling approach based differential evolution on the safe set for highly imbalanced datasets (SS_DEBOHID). SS_DEBOHID firstly uses k-nearest neighbors (kNN) method to learn the safe area of minority; then the DEBOHID oversampling method is used to synthesize new minority samples in the safe area. The advantages of SS_DEBOHID include that (a) it generates samples in the safe area to reduce generation of noisy samples and reduce synthetic samples falling into the classification boundary and majority area; (b) it uses the DEBOHID method to synthesize samples and increase the diversity of samples; (c) the method is suitable for highly imbalanced datasets. The proposed method is compared with 10 methods on 43 highly imbalanced datasets and evaluated on AUC and G_Mean metrics. The experimental results show that SS_DEBOHID obtains more than 30 best performing datasets on KNN, SVM, and DT classifiers in terms of AUC and G_mean, respectively. The proposed method outperforms other methods by 8.07% to 24.34% on average AUC metric and by at least 6.96% and up to 45.37% on average G_mean metric. In addition, we validate the efficiency of SS_DEBOHID on 8 high-dimensional and large sample size datasets. The experimental results show that SS_DEBOHID has better classification performance and robustness.