Comparative Analysis of Oversampling Techniques on Small and Imbalanced Datasets Using Deep Learning

In imbalanced datasets, certain classes have a larger number of samples compared to others, leading to an unequal distribution of samples across the classes. Since many crucial real-world classification problems, like medical diagnosis, involve imbalanced data, the research community places a high priority on understanding how to use this data. If machine learning is performed directly on the imbalanced data, the disparity between the majority and minority classes will cause bias towards the majority class and lead to inaccurate results. There is growing interest in this field of study, and several algorithms have been created. This study aims to evaluate the effectiveness of five oversampling strategies that are intended to address data imbalance, namely random oversampling, SMOTE, borderline SMOTE, ADASYN, and Deep SMOTE. A comparative analysis is carried out, and the effectiveness of each strategy is examined in terms of evaluation metrics. Experimental results demonstrated that DeepSMOTE outperformed all other oversampling techniques on small and imbalanced datasets.