Reducing Class Bias In Data-Balanced Datasets Through Hardness-Based Resampling

Class-bias, that is class-wise performance disparities, is typically attributed to data imbalance and addressed through frequency-based resampling. However, we demonstrate that substantial bias persists even in perfectly balanced datasets, proving that class frequency alone cannot explain unequal model performance. We investigate these disparities through the lens of class-level learning difficulty and propose Hardness-Based Resampling (HBR), a strategy that leverages hardness estimates to guide data selection. To better capture these effects, we introduce an evaluation protocol that complements global metrics with gap- and dispersion-based measures. Our experiments show that HBR significantly reduces recall gaps, by up to 32% on CIFAR-10 and 16% on CIFAR-100, outperforming standard frequency-based resampling. We further show that we can improve fairness outcomes by selectively using the hardest samples from a state-of-the-art diffusion model, rather than randomly selecting them. These findings demonstrate that data balance alone is insufficient to mitigate class bias, necessitating a shift toward hardness-aware approaches.