A Difficulty-aware Approach to Fair Classification on Imbalanced Datasets

Class imbalance in real-world datasets often biases standard classifiers toward the majority class, degrading performance on the minority class. While existing methods like sample re-weighting can mitigate this, they may increase overall misclassification errors or fail to consider the difficulty of training instances. To address these shortcomings, we introduce a difficulty-aware classification framework based on multi-objective evolutionary optimization. Our approach uses a specialized fitness function to simultaneously optimize for minority-class recall and overall accuracy, guiding the selection of the most informative training samples. We quantify sample difficulty using a fuzzy approach, which then modulate class-specific weights to refine the classifier's decision boundary. Furthermore, we incorporate chaotic dynamic maps into the evolutionary operators to accelerate convergence and maintain population diversity. Evaluated on various UCI benchmark datasets with 10-fold cross-validation, our method improves minority-class performance on imbalanced data without compromising accuracy on balanced data. Comparative analysis using AUC, G-mean, and F-measure confirms our approach achieves a superior trade-off between minority-class detection and overall accuracy compared to state-of-the-art methods.