MCROR: A Two-Step Framework for Class Imbalance and Overlap Reduction in Binary Classification

Class imbalance and overlap present significant challenges in binary classification tasks, particularly in domains such as medical diagnosis and fraud detection. The disproportionate representation of classes and the overlapping regions between them lead to biased classifiers that favor the majority class and exhibit poor generalization. In this paper, we introduce MCROR (Majority Class Reduction and Overlap Removal), a novel two-step framework designed to mitigate class imbalance and reduce overlaps in binary classification. The first step of the proposed approach reduces the number of majority class instances by partitioning the dataset into regions and selectively replacing instances with their mean values. In the second step, overlapping instances are removed by calculating the Euclidean distance between instances and eliminating those that fall within a critical distance threshold. We evaluate MCROR on synthetic imbalanced and overlapped datasets and compare its performance with existing resampling techniques, such as SMOTE, ADASYN, and Random Under-Sampling. The results demonstrate that MCROR achieves competitive accuracy (92%) while significantly improving recall (92.59%) and AUC-ROC (96.57%), outperforming other methods in terms of identifying minority class instances. The proposed method shows promise for applications in imbalanced datasets, particularly in medical and anomaly detection tasks.