SC-SMOTE: Stability-constrained SMOTE Based on Multi-model Consistency Constraints

Class imbalance is a prevalent and challenging issue in machine learning that severely degrades the performance of standard classifiers. Although the Synthetic Minority Over-sampling Technique (SMOTE) is widely used to address this problem by interpolating new data points, its standard formulation tends to blindly generate synthetic samples based solely on geometric Euclidean distances. This blind interpolation often creates noisy or unrealistic samples in overlapping regions, which destroys the original data distribution and introduces decision ambiguity. To overcome these limitations, this paper proposes a novel over-sampling framework named Stability-constrained SMOTE (SC-SMOTE).

The core idea of SC-SMOTE is to leverage the collective intelligence of an ensemble system to define a "secure and stable region" in the feature space. The proposed method first trains multiple heterogeneous base classifiers (such as Logistic Regression, Random Forest, and Support Vector Machines) on the original imbalanced dataset. A Multi-model Consistency Index (CI) is then introduced to quantify the prediction confidence of the ensemble. During the sample generation phase, a candidate synthetic sample is accepted if and only if its CI meets a predefined consistency threshold, ensuring that new samples are strictly restricted to the stable minority regions agreed upon by multiple models. By incorporating this prior filtering mechanism, SC-SMOTE effectively prevents the over-generalization of artificial samples into majority class territories or overlapping noise areas. The proposed approach enhances the quality of synthetic data, preserves the integrity of the true decision boundaries, and provides a more robust solution for handling complex imbalanced data classification tasks.