Drift-Aware Adaptive Classification for Imbalanced Data via Dynamic Class Reweighting and Structural Regularization

Abstract

This study proposes a dynamically adaptive classification framework to address the widely observed problems of class imbalance and distribution shift in real-world tasks. The framework incorporates distribution difference measurement, dynamic class weight allocation, and structural regularization to achieve coordinated adaptation across feature representation, loss modeling, and decision boundary optimization. This enables the model to perceive both class rarity and distribution changes. The method first estimates the degree of shift between the current data and the source distribution through statistical feature analysis, and then adjusts class weights and internal representations accordingly. This strengthens attention to minority classes and mitigates performance degradation caused by distribution drift. At the same time, the structural regularization mechanism constrains the feature space and preserves robust representations even under complex class structures or strong sample heterogeneity. The framework is highly scalable and can be applied to a variety of imbalanced data scenarios while maintaining stable performance under dynamic distribution changes. By integrating imbalance learning and distribution shift handling from a unified adaptive perspective, this study offers a new solution for building classification systems capable of long-term stable operation in real environments.