A Comparison of Oversampling Effects on Imbalanced Topic Classification of Korean Texts

Abstract

Imbalanced data is a widely-acknowledged problem in supervised learning classification tasks. Oversampling is one way to overcome the problem and there are many methods of oversampling that have been discovered. While researches on the effect of oversampling on other languages have been widely conducted, studies comparing oversampling methods on Korean texts are scarce. This study compares the effect of oversampling methods on the task of classifying Korean internet news articles. This study finds that support vector machines (SVM) and logistic regression reacted with stability and performed best when paired with borderline-SMOTE2 in imbalanced conditions.