BCGAN-based over-sampling scheme for imbalanced data

Classification is a process of identifying the class to which input data belong. One of the most popular methods to do this is to construct a classification model by training a machine learning algorithm using a given set of data. For better classification performance, the dataset should have a balanced data distribution by class. If the dataset is imbalanced, that is, one class (minority class) has very fewer data than the other class (majority class); a model has little chance to learn about the minority class, and training is biased to the majority class. As a result, the model tends to classify any input to the majority class and does not handle data of the minority class properly. To overcome this data imbalance problem, we propose a novel over-sampling scheme based on Borderline-Conditional Generative Adversarial Networks (BCGAN). Our BCGAN generates data for the minority class, particularly along the borderline between majority class and minority class. Through various experiments on actual imbalanced datasets, we show the performance of our scheme.