Region-SMOTE: A New Over-Sampling Method in Imbalanced Data Sets Learning

Learning from imbalanced data sets usually produces biased classifiers that have a higher predictive accuracy over the majority class, but poorer predictive accuracy over the minority class. In this paper, we present a new over-sampling method, Region-SMOTE, to improve class imbalance and classification accuracy. In order to combine the boundary distribution information, the vote rule is formulated, which divides the minority samples into safety samples, intermediate samples and dangerous samples, and adopts different processing strategies for different samples. We evaluate Region-SMOTE method from two aspects, F-value and G-mean. We conduct extensive experiments on four data sets, including Diabetes, which is extracted from real-world physical examination database, and another three data sets, which are extracted from UCI, to verify the method. The evaluation results demonstrate that our proposed over-sampling method has better performance.