Hybrid resampling for enhanced multiclass classification

Multiclass classification has become increasingly important in areas such as healthcare, fraud detection, agriculture, and network security. Yet, its effectiveness is often reduced when datasets are imbalanced, as conventional algorithms tend to favour majority classes while overlooking minority ones. Approaches like One-vs-One (OVO) and One-vs-All (OVA) attempt to manage this issue but often lead to information loss. Likewise, traditional resampling techniques either risk overfitting or remove useful data, limiting their impact in multiclass settings. The objective of this study is to identify a more effective strategy for handling multiclass imbalance by comparing a range of resampling methods and classifiers. We evaluated 14 well-known classifiers on 28 benchmark datasets from the UCI and KEEL repositories. Eight resampling methods, covering oversampling, undersampling, and hybrid techniques, were applied, with performance measured through cross-validation accuracy, accuracy, f1-score, and roc_auc. Our findings show that the hybrid resampling SmoteEnn combined with a Stacking Classifier consistently outperforms other methods, achieving up to 97.9% accuracy and f1-score, and 99.6% roc_auc. When benchmarked against the recently proposed Ensemble Partition Sampling (EPS), the proposed approach demonstrated clear superiority. These results confirm that hybrid resampling, particularly SmoteEnn, offers a robust and reliable solution for multiclass imbalanced learning.