Bagging Folds using Synthetic Majority Oversampling for Imbalance Classification

Despite the impressive strides made in machine learning, imbalanced data remains a persistent challenge for contemporary models. Misclassifying minority class instances can cause biased models and inaccurate predictions. However, resampling techniques have shown promise as an effective strategy for dealing with imbalanced data, and it is essential to be mindful of their potential drawbacks. To tackle the persistent challenge of imbalanced data in machine learning, we present a novel solution: BaggFold. This innovative meta-framework leverages majority class oversampling and a bagging architecture to generate homogeneous classifier ensembles. Its model-agnostic nature sets BaggFold apart, making it compatible with any classifier. To derive the final classification results, BaggFold employs a voting mechanism with a threshold optimization based on Youden's J statistic and the receiver operating characteristic. Our research involved a rigorous evaluation of BaggFold's performance, stability, and speed on 11 imbalanced datasets, three metrics (F1-score, G-mean, H-measure), and 156 models. The results demonstrate that BaggFold has surpassed several state-of-the-art models in numerous tests and has also maintained a high level of comparability in other tests.