불균형 데이터의 분류 성능 향상을 위한 샘플링 기법 비교 연구

초록

Imbalanced data pose a significant challenge in classification tasks across various industries, often leading to substantial performance degradation, particularly in critical applications such as fraud detection, medical diagnosis, and customer churn prediction. To address this issue, this study systematically evaluates the effectiveness of oversampling, undersampling, and hybrid sampling techniques. Specifically, Random Oversampling, SMOTE, and ADASYN were employed as oversampling methods, while Random Undersampling, NearMiss, Tomek Links, Condensed Nearest Neighbors, and One-Sided Selection were used as undersampling approaches. In addition, hybrid methods combining oversampling and undersampling—such as Random Oversampling with Random Undersampling, SMOTE with Random Undersampling, and SMOTE with Tomek Links—were also investigated. The performance of these sampling techniques was assessed across multiple imbalanced datasets using various classification models. Evaluation metrics included minority class recall, specificity, and F-scores to capture both sensitivity and class balance. The results provide a comprehensive comparison of sampling strategies and reveal how their effectiveness varies depending on the underlying methodology and data characteristics. This study offers practical insights into selecting appropriate sampling techniques for handling imbalanced data and provides a useful guideline for researchers and practitioners aiming to improve classification performance in real-world applications.

키워드