Automatically weighted focal loss for imbalance learning

In the context of class-imbalanced learning, most CNN-based classification algorithms encounter the problem of majority class gradient dominance, which makes them susceptible to bias toward the majority class. It is the consequence of the underlying assumption that the costs of misclassification are equivalent and that the class distribution is reasonably balanced. A standard way for dealing with the data imbalance is by using a loss function that favors the minority class. A constant weight is considered for each class in the loss function, which is generally inversely related to the number of instances in the class. In this paper, for multi-class imbalanced learning, we proposed Dynamic Weighted Focal Loss (DWFL), which determined the weight of each class for the next epoch automatically based on the classifier’s results in all previous epochs. In this way, the better-learned class weighed less, and the less-learned class weighed more. Additionally, the application of focus loss reduced the importance of simple samples while increasing the importance of hard ones compared to cross-entropy. The experiments on several imbalanced datasets revealed that the performance of our suggested strategy (DWFL) is superior to that of previous resampling or reweighting methods with imbalanced classes. The experimental findings demonstrate the efficacy and convergence of the proposed methodology.