Long-tailed classification based on dynamic class average loss

In real-world data distributions, class imbalance is a common issue. When training deep learning models on class-imbalanced data, the performance of classes with fewer samples tends to deteriorate. Numerous studies have addressed this problem, focusing on loss reweighting techniques based on the number of training samples per class. However, because some classes are inherently easier or harder to classify, having a larger number of samples in a particular class does not necessarily ensure lower loss or better learning for that class. Additionally, if the ratio of loss magnitudes differs substantially from the ratio of the number of training samples per class, reweighting based solely on sample size may be inappropriate. This study proposes a method to reweight losses based on dynamic class average losses rather than the number of training samples per class to address these issues. Specifically, this method evaluates the class average losses for each mini-batch, applies a nonlinear transformation to these values, and dynamically adjusts the class-wise loss weights within the loss function during training to better mitigate class imbalance. Experimental results from various types of datasets, including image and tabular data, demonstrate that the proposed method improves performance by 1%–8% across various datasets compared to existing methods.