Meta-constraint SMOTE: A Robust Over-sampling Algorithm for Imbalanced Data Based on the Dynamics of Constraint Variation

Class imbalance remains a fundamental challenge in machine learning and data mining, where standard classifiers frequently exhibit bias toward the majority class at the expense of minority class recognition. While the Synthetic Minority Over-sampling Technique (SMOTE) is widely used to mitigate this issue, its blind interpolation mechanism often introduces noise and blurs decision boundaries, especially within complex data distributions characterized by overlapping regions and non-convex manifolds. Existing advanced variants typically impose rigid, hand-crafted spatial or density constraints on generated data, which inherently lack environmental adaptability and are prone to overfitting specific datasets. To address these limitations, this paper proposes a novel over-sampling framework termed Meta-constraint SMOTE. Moving away from traditional approaches that directly restrict synthetic data, our core innovation lies in imposing constraints on the "variations of the constraints themselves." By incorporating a meta-learning mechanism coupled with a dynamic feedback loop, the proposed algorithm evaluates the generalization performance of lower-level constraint conditions in real time during the data generation process. It then dynamically calibrates the constraint weights and boundaries based on meta-gradients driven by validation losses. This higher-level governance, or "constraint on constraints," empowers the algorithm with exceptional self-adaptation, enabling it to precisely capture and protect intricate minority class manifolds while inherently neutralizing the adversarial effects of noise. Theoretical analysis demonstrates that Meta-constraint SMOTE guarantees convergence under a bi-level optimization framework. Empirical evaluations across various complex imbalanced datasets confirm that our algorithm consistently and significantly outperforms traditional SMOTE and its state-of-the-art derivatives across key metrics, including G-mean, F1-score, and AUC.