Bi-SMOTE: bi-directional oversampling for reducing class overlap and synthetic noise
EXPERT SYSTEMS WITH APPLICATIONS, cilt.334, 2027 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 334
- Basım Tarihi: 2027
- Doi Numarası: 10.1016/j.eswa.2026.134648
- Dergi Adı: EXPERT SYSTEMS WITH APPLICATIONS
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Applied Science & Technology Source, Compendex, INSPEC, Public Affairs Index, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Ondokuz Mayıs Üniversitesi Adresli: Evet
Özet
Class imbalance remains a fundamental challenge in supervised classification, often leading to biased decision boundaries and poor minority class recognition. Synthetic oversampling methods such as SMOTE address this issue by interpolating new minority samples; however, unconstrained generation may increase class overlap and degrade generalization performance. In this study, we introduce BiSMOTE, a geometrically constrained oversampling technique that restricts synthetic sample generation to locally bounded interpolation regions defined by minority neighborhood structure and proximity to opposing classes. By enforcing a class-aware geometric constraint, the proposed method aims to expand the minority class while preserving boundary integrity. Extensive experiments were conducted across 68 benchmark datasets, nine classifiers, and seventeen competing resampling methods. Performance was evaluated using four imbalance-robust metrics, and statistical comparisons were performed via Friedman and post-hoc tests, complemented by a Bayesian pairwise comparison based on a Bradley-Terry model. The method's robustness to class overlap - its central design objective - was further examined using two established overlap measures and a controlled simulation with an exactly known decision boundary. The results demonstrate that BiSMOTE consistently achieves superior average ranks on AUC, GMEAN, and BACC, and largely consistent, competitive ranks on MCC, compared to widely used oversampling methods, with its advantage on overlap-sensitive metrics maintained or strengthened as class overlap becomes more severe. Additional analyses reveal stable behavior across varying dataset characteristics and limited sensitivity to hyperparameter choices. Overall, the findings indicate that boundary-aware constrained interpolation provides an effective balance between minority expansion and structural preservation, offering a reliable alternative to conventional SMOTE-based approaches.