QC-SMOTE Improves Imbalanced Classification by Generating Quality Samples

Class imbalance presents a significant hurdle in classification tasks, where existing oversampling methods like SMOTE often inadvertently generate low-quality synthetic samples, particularly in noisy regions or areas of class overlap. These poor-quality samples can degrade model performance rather than improve it. To counter this, a new framework called QC-SMOTE (Quality-Controlled SMOTE) has been proposed. QC-SMOTE is designed to estimate the reliability of minority samples using a composite neighborhood trustworthiness score. This score combines local density, safe-level status, and isolation from the majority class, providing a more nuanced understanding of where to generate synthetic data. Synthetic candidates are generated using an IPQ-guided "best-of-K" strategy that evaluates midpoint purity and, when necessary, majority clearance. The allocation of these samples is guided by their reliability and boundary informativeness. Crucially, QC-SMOTE's generation behavior adapts to different overlap-imbalance regimes, adjusting interpolation ranges and selection criteria to match the local data geometry. In severely noisy regions, low-quality synthetic samples are gracefully replaced with original minority duplicates, preventing further degradation. Experimental validation across 30 imbalanced datasets showed that QC-SMOTE achieved the strongest average AUC-ROC and Macro F1 scores among compared oversampling methods, with particularly notable gains under moderate and severe imbalance conditions.