Research on Unbalanced Data Classification Algorithm Based on Boundary Simulation SMOTE

Classifying imbalanced datasets poses a significant challenge in machine learning, as traditional classifiers tend to favor the majority class, leading to poor recognition rates for the minority class. Although the Synthetic Minority Over-sampling Technique (SMOTE) alleviates this issue by interpolating between minority instances, its blind sampling mechanism often creates synthetic samples in overlapping areas or majority class regions, resulting in blurred decision boundaries and noise amplification. To address these limitations, this paper proposes a novel over-sampling method named Boundary Simulation SMOTE (BS-SMOTE). The core idea of BS-SMOTE is to identify and simulate the behavioral boundary of classifiers to guide the generation of synthetic instances. By analyzing the local neighborhood structure of both majority and minority classes, the algorithm filters out isolated noise and precisely extracts critical "boundary minority samples." During the sample generation phase, a boundary constraint mechanism based on the nearest majority neighbors is introduced. This mechanism adjusts the interpolation direction to simulate the decision boundary, ensuring that synthetic samples reinforce the boundary area without invading the majority class space. Performance is evaluated using comprehensive metrics tailored for imbalanced data, including F1-Score, G-Mean, and AUC. This approach effectively refines the decision boundary, prevents class overlapping, and significantly enhances the classifier's generalization capabilities on imbalanced data.