A Fuzzy Adjustment Mechanism to Tune Focal Loss Hyperparameters in Imbalanced Deep Learning

To address the problem of data imbalance in deep learning, this paper proposes a fuzzy adjustment mechanism that dynamically adjusts the focal loss focusing parameter according to the class-specific learning status. Instead of using a single fixed focusing parameter for all classes, the proposed method adjusts the class-wise parameter based on three factors: class size, class-specific focal loss, and changes in focal loss. This enables the model to adjust the focus on different classes during training according to their learning conditions. The experimental results show that the proposed approach achieves a competitive performance compared with the evaluated fixed focal loss baselines. Based on paired t-tests across five trials, the proposed method shows statistically significant improvements over cross-entropy loss on both CIFAR-10 and SVHN datasets. For the fixed focal loss baselines, some differences are statistically significant, while others are not. Given the small number of paired trials (n = 5), these statistical results should be interpreted with caution. Overall, the proposed fuzzy dynamic mechanism provides a practical alternative to manually specifying a fixed focusing parameter by adaptively adjusting the parameter according to class-level learning conditions.