Boundary-avoid SMOTE: Explicit Majority Boundary Detection for Overlapping-free Oversampling in Imbalanced Classification

Class imbalance presents a severe challenge for traditional machine learning algorithms, as classification boundaries tend to shift toward the minority class, leading to a significant drop in its recognition rate. Although the Synthetic Minority Over-sampling Technique (SMOTE) is widely used to mitigate this issue, its inherent blindness in synthesizing samples often causes severe class overlapping and concept drift, especially in complex boundary regions where majority and minority classes intertwine. To resolve this deficiency, this paper proposes a novel oversampling framework named Boundary-avoid SMOTE (BA-SMOTE). Unlike existing edge-centric variants that only detect boundaries from the minority perspective, BA-SMOTE explicitly quantifies and outlines the external defense zone of the majority class. By calculating the minority infringement degree of majority instances, the algorithm establishes a series of hyperspherical forbidden zones along the majority boundaries. During the subsequent synthetic interpolation phase, a rigorous boundary-projection blocking mechanism is introduced to discard any candidate synthetic points that fall into these forbidden zones. Theoretical analysis demonstrates that the proposed BA-SMOTE effectively prevents the generation of ambiguous samples, preserves the clarity of the topological decision boundary, and increases the classification margin. The boundary-avoidance mechanism provides a robust and mathematically sound solution for handling highly overlapped and imbalanced datasets.