Oversampling Methods to Handle the Class Imbalance Problem: A Review

Abstract

Many real-world applications, such as medical diagnostics, fraud detection etc., have a class imbalance problem where one class has less instances than the other. The non-uniform distribution of the dataset has a significant impact on the classification models’ performance because they fail to detect a minority class instance. In this paper, we empirically reviewed five oversampling methods to address the class imbalance problem (CIP), including SMOTE (Synthetic Minority Oversampling Technique), Safe level SMOTE, SMOTE Tomek Links, Borderline SMOTE1 and Adaptive SMOTE (ADASYN), using four classification models: Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF) and K Nearest Neighbor (KNN). Different performance metrics such as accuracy, precision, recall, f1 score, and area under the curve (AUC) were also used. The experimental results showed that SMOTE Tomek Links technique outperformed the other methods for most of the datasets.