<![CDATA[Class imbalance remains a major challenge in multi-class classification, where existing hybrid resampling methods often combine oversampling and undersampling in a loosely coupled manner, without explicitly coordinating minority enrichment and majority reduction. In this experimental study, we propose a novel hybrid resampling method, Adaptive Weighted Oversampling and Instance Hardness-Based Undersampling (AWO-IHU), which differs from existing hybrid approaches by explicitly aligning boundary-aware minority oversampling with instance hardness-based majority undersampling. Rather than independently applying oversampling and undersampling, the proposed method integrates both processes through a coordinated design guided by classification difficulty to improve decision boundary quality. Methodologically, AWO-IHU first applies adaptive weighted oversampling to emphasize informative minority instances near class boundaries, followed by instance hardness-based undersampling that selectively removes redundant majority samples using an ensemble-based difficulty estimation. The experimental evaluation is conducted using multiple benchmark datasets with varying numbers of instances, attributes, and classes. Classification performance is evaluated using Accuracy, Precision, Recall, and Cohen’s Kappa, enabling a comprehensive assessment of overall correctness, minority sensitivity, and agreement beyond chance under class imbalance. Experimental results show that AWO-IHU consistently outperforms SMOTE, Random Undersampling, and conventional hybrid sampling methods. In particular, the proposed method achieves perfect or near-perfect Recall values up to 1.0, while maintaining high Precision values above 0.89 and producing the highest Cohen’s Kappa values up to 0.86. These findings demonstrate that explicitly coordinating minority enrichment with difficulty-aware majority reduction yields more reliable decision boundary learning and improved generalization in imbalanced multi-class classification. ]]>
