A0650
Title: Enhancing classification performance under class imbalance using improved synthetic oversampling methods
Authors: Saman Muthukumarana - University of Manitoba (Canada) [presenting]
Abstract: Class imbalance poses a significant challenge in classification tasks, especially in domains such as medical diagnostics, fraud detection, and anomaly identification, where instances of interest are often rare. Standard classification algorithms tend to favor the majority class, resulting in poor detection of minority class instances and reduced overall model utility. This issue is addressed by presenting four novel extensions of the synthetic minority oversampling technique (SMOTE): Distance ExtSMOTE, Dirichlet ExtSMOTE, FCRP SMOTE, and BGMM SMOTE. These approaches improve the generation of synthetic minority instances by incorporating local neighborhood structures and probabilistic modeling to reduce noise sensitivity and mitigate the influence of outliers. Extensive experiments on both simulated and real-world datasets reveal that these enhanced SMOTE methods consistently outperform the original SMOTE and several of its strong variants. Notably, Dirichlet ExtSMOTE achieves superior results across multiple evaluation metrics, including F1 score, Matthews Correlation Coefficient (MCC), and Precision-Recall AUC, underscoring its robustness and effectiveness in handling imbalanced classification problems.
