클래스 불균형 데이터의 효과적인 분류를 위한 k-NN과 생성적 적대 신경망 기반의 오버 샘플링

Abstract

Class imbalanced dataset is common in real world and may degrade the performance of the classifier. To address this, oversampling method that artificially creates the samples of minority class is adopted but is known to be ineffective for the high-dimensional dataset because it generates the samples whose distribution is far different from that of existing samples. Novel oversampling methods based on the generative adversarial networks (GAN) have been recently developed, but generated samples may have different degrees of impact on the performance of the classifier. Therefore, more efficient method that can capture the characteristics of the generated samples and select those samples that will be used to train and improve the performance of the classifier is necessary. This study proposes a GAN-based new oversampling method which generates artificial samples based on the distribution of existing minority class samples and extracts only those which are effective to expand the realm of samples of minority class using the k-nearest neighbor. We show the proposed method outperforms existing methods with respect to F1-measure by several illustrative datasets.