Imbalanced learning tackles datasets with skewed class distributions. While Localized Random Affine Shadowsampling (LoRAS) augments minorities via affine combinations of shadow points, it neglects boundary regions and noise control. We propose Clear Boundary LoRAS (CBLoRAS), an oversampling method that first isolates boundary minority samples, applies LoRAS within this subset, and then prunes synthetic noise with reverse k-nearest-neighbor filtering. Evaluated on 12 highly imbalanced tabular sets, CBLoRAS raises F1-scores by 3–5 % over LoRAS; ablations confirm that boundary focus and RkNN denoising jointly yield cleaner, more representative synthetic data.