Dealing with Class Imbalances with Synthetic Examples

Benjamin X. Wang and Nathalie Japkowicz

A dataset is imbalanced if the classification categories are not approximately equally represented. This paper shows that a combination of our method of adaptive over-sampling of the minority (abnormal) class and under-sampling of the majority (normal) class can achieve a good classification performance. The method is evaluated using the Receiver Operating Characteristic (ROC) technique. Our method is an extension of a recently proposed state-of-the-art approach, SMOTE, and the result shows that it outperforms SMOTE, on three imbalanced datasets.