Fig. 4
From: An approach for handling imbalanced datasets using borderline shifting
Illustration of the three majority-class regions identified during resampling. Each majority-class instance is categorized as safe, borderline, or noise based on the distribution of its neighboring minority and majority samples. Safe instances (dark blue) are surrounded by other majority samples, borderline instances (light blue) lie near the decision boundary, and noisy instances (white) are surrounded primarily by minority samples. Minority-class instances are shown in red. This classification is commonly used to guide noise handling and oversampling strategies in imbalanced learning46 proposed a neighbor-displacement-based oversampling approach to improve class representation in multiclass imbalance sce- narios.E, as well asSaglam (2025)47 introduced a noise module which relocates points for noise-tolerant synthetic point gen- eration including DatRel. In contrast to these works, our method is aimed at relocate (or reassign) majority class instances near the decision boundary to the minority class, thereby reshaping the boundary and improving classification accuracy, which provides more stable sample generation and reduces overlapping instances. Fig. 5. visually depicts the data distribution before and after the boundary shift process.
