A hybrid oversampling for imbalanced data using incremental SMOTE-KMeans and cluster-structured positive class-SVM

Abstract

Class imbalance remains a significant challenge in classification tasks, particularly when the minority class exhibits complex internal structures. This study proposes a unified framework that integrates Incremental SMOTE-KMeans with a cluster-structured positive class-SVM (CS-PC-SVM) to jointly address data imbalance and structural heterogeneity. The proposed method introduces structural alignment between oversampling and classification by generating synthetic samples only within reliable clusters and modeling the minority class as multiple subgroups. This design reduces noise, preserves local data structure, and enables more adaptive decision boundaries. Experimental results on five benchmark datasets demonstrate that the proposed approach achieves consistently strong and balanced performance, with Accuracy up to 0.981, G-Mean 0.969, Precision 0.967, and Recall 0.962, outperforming or remaining competitive with existing methods. These findings highlight the effectiveness of integrating structure-guided data generation with structure-aware classification for improving robustness in imbalanced learning scenarios.