A Comparative Analysis of Resampling Techniques to Address Imbalanced Datasets

A typical issue in machine learning is imbalanced datasets, which occur when there are significantly more samples in one class than another. Algorithms have a tendency to prioritize the classes with more examples and ignore the classes with fewer examples when dealing with this issue, which can lead to poor model performance. Many data-level approaches, including oversampling, undersampling, and data augmentation, have been presented as potential solutions to this problem. The effectiveness of imbalanced learning has also been enhanced by the use of algorithmic methods including cost-sensitive learning, one-class learning, and anomaly detection. However, there are benefits and drawbacks to every strategy; choosing the right one requires considering the specifics of the data and the problem at hand. Therefore, it is crucial to have an in-depth familiarity with the various methods in order to construct efficient and trustworthy models from imbalanced datasets. The impact of imbalanced datasets on machine learning models and the techniques used to address the challenges have been analyzed in this article (Shorten and Khoshgoftaar in J Big Data 6:1–48, 2019 [1]). These five resampling techniques, cluster sampling, random oversampling, near miss undersampling, random undersampling, and SMOTE, were put to the test in this study. In order to compare the four approaches, we chose four publicly accessible imbalanced datasets “Stroke Prediction”, “Glass Classification”, “Classification of Silhouettes of Vehicles”, and “Thyroid Prediction” from Kaggle. Performance evaluation on the k-nearest neighbors (KNNs) algorithm shows that oversampling outperforms undersampling for KNN and achieves superior ratings in several evaluation metrics.