Performance Analysis of SMOTE-based Oversampling Techniques When Dealing with Data Imbalance

Building classification models on imbalanced data proves to be a challenging task despite the multitude of available classifiers. The classifier bias towards the majority class can be ameliorated through various manners and with varying degrees of success. Oversampling minority or undersampling majority instances are prominent amongst these due to both their simplicity and effectiveness. Probably the most popular approach to oversampling is the well-known SMOTE algorithm, based on which numerous enhancement attempts were made. This paper aims to compare the performance of these, more complex, oversampling techniques with regard to the original on a wide array of problems. Additionally, it attempts to give insight into the behavior of different interpretations of the original algorithm apparent in the literature. In that regard, some interesting findings were made.