Oversampling for Imbalanced Data Classification Using Adversarial Network

Abstract

The imbalanced data classification problem occurs when the number of samples for one class is much lower than for the other class. In most classification algorithms, the class imbalance is key reason of performance degradation. One way to address the imbalancing issue is to balance them, either by oversampling instances of the minority class or undersampling instances of the majority class. In this paper, we propose an oversampling method for imbalanced data classification using an adversarial network. Firstly, a synthetic minority dataset is generated with a black box oversampler and refined using the refiner network. To bridge a gap between synthetic and real dataset, we train the refiner network using an adversarial loss. The adversarial loss fools a discriminator network that classifies a dataset as real or refined. Experimental results show that the proposed method has high performance comparing with the most common oversampling method.