MCSMOTE: A Transition Matrix-Driven Oversampling Technique for Class Imbalance

Abstract

Class imbalance presents a challenge in machine learning, often skewing predictive performance toward the majority class and undermining the accuracy of minority class predictions. To address this, we introduce MCSMOTE, a novel resampling method that employs a transition matrix-based Monte Carlo mechanism for generating synthetic samples. MCSMOTE differentiates itself by modeling the relationships among features and leveraging probabilistic transitions to generate synthetic data points that effectively capture the underlying data structure. This approach ensures enhanced representativeness of the minority class while approximating the local structure of the minority class and thereby generating samples that reflect the underlying data patterns. Comprehensive experiments across 63 diverse imbalanced datasets demonstrate that MCSMOTE consistently outperforms nine widely used resampling techniques-NORES, ROS, SMOTE, ADASYN, BLSMOTE, RWO, SMOTEWB, DeepSMOTE, RWO, and GQEO-when evaluated using multiple classifiers and six key performance metrics: balanced accuracy, F1-score, G-mean, MCC, ROCAUC, and ROCAUC. Results show that MCSMOTE achieves the highest average performance across all metrics. Friedman and Nemenyi tests confirm that these improvements are statistically significant. An ablation study further highlights the stability and effectiveness of MCSMOTE's hyperparameter choices across different data characteristics. These findings establish MCSMOTE as a powerful and reliable solution for addressing class imbalance in machine learning applications.