Adaptive penalty for improved minority class detection

Abstract

Imbalanced datasets present significant challenges across a variety of real-world applications, including medical diagnostics, fraud detection, and computer vision. In many of these scenarios, accurately identifying instances of the minority class is critically more important than correctly classifying those of the majority class. This paper introduces a novel method specifically designed to enhance the accuracy of minority class predictions. Our approach employs a custom loss function that initially assigns a high penalty to the misclassification of minority instances and dynamically reduces this penalty as the model’s accuracy for the minority class improves. This adaptive weighting strategy effectively achieves high accuracy for the minority class while maintaining robust overall classification accuracy. Experimental results on synthetic and benchmark datasets demonstrate the efficacy of the proposed method, highlighting its potential for addressing imbalanced learning problems in diverse application domains.