Abstention-SMOTE
Cheng Zhang, Yufei Chen, Xianhui Liu, Xiaodong Zhao · 2017
In recent years, classification of imbalanced data has troubled most classification models because of the imbalanced class distribution. Synthetic Minority Oversampling Technique (SMOTE) is one of the solutions at data level, but this kind of method doesn't consider the distribution of the data set, thus the result is not satisfied. Based on the SMOTE method, this paper proposed an over-sampling method for imbalanced data classification, called Abstention-SMOTE. Firstly, we construct abstaining classifiers using ROC analysis. Then we use the abstaining classifiers to generate the abstention positive samples, which only includes the positive samples that are easily to be misclassified. Finally, we use these abstention positive samples to synthetize new positive samples to balance the data distribution. The experiment results indicate that our approach can achieve better results through comparing with three other over-sampling methods, i.e. RO-sampling, SMOTE and Borderline-SMOTE.