In-Depth Performance Analysis of SMOTE-Based Oversampling Algorithms in Binary Classification

In the field of machine learning, the problem of class imbalance considerably impairs the performance of classification algorithms. Various techniques have been proposed that seek to mitigate classifier bias with respect to the majority class, with simple oversampling approaches being one of the most effective. Their main representative is the well-known SMOTE algorithm, which introduces a synthetic instances creation mechanism as an interpolation procedure between minority instances. To date, an abundance of SMOTE-based extensions that intend to improve the original algorithm has been proposed. This paper aims to compare the performance of several such extensions. In addition to comparing the overall performance, the impact of the selected oversamplers on the per-class performance is also evaluated. Finally, this paper tries to interpret the obtained performance results with respect to the internal procedures of oversampling algorithms. Some interesting findings have been made in this regard