Enhancing Model Performance in Hybrid Class Imbalance Techniques
Kashika Wadhwa, Ritika Kumari, Anjana Gosain · Procedia Computer Science · 2025
Class imbalance is a crucial issue in real world scenarios. Traditional classifiers trained with imbalanced datasets become biased towards the majority class (class with more instances), resulting in misclassification. Therefore, it is essential to have uniform class distribution within the dataset. Ensemble methods have gained more attention from the researchers for managing the issue of imbalance distribution of classes. In this paper, three hybrid approaches are implemented using Over-sampling technique (Adaptive Synthetic Sampling- ADASYN) and Ensemble methods: Bagging, Boosting and Stacking. The performance of these hybrid approaches (ADASYN-Bagging, ADASYN-Boosting and ADASYN-Stacking) is evaluated using four performance metrics Accuracy (ACC), F 1 -score, Geometric-mean (GM) and Receiver operator characteristics Area under the ROC curve (ROC-AUC) on twelve imbalanced datasets taken from KEEL repository. The study suggest that ADASYN-Stacking outperforms all other approaches with ROC-AUC value 99.93%.