Abstract
Automated Facial Expression Recognition (FER), involves two critical aspects: data and model design. Both significantly influence bias and fairness in FER tasks. However, issues related to bias and fairness in FER datasets and models remain underexplored. This study investigates bias and fairness in FER datasets and models. The bias of four common in-the-wild FER datasets, including AffectNet, ExpW, Fer2013, and RAF-DB, is studied. Additionally, this research evaluates the bias and fairness of seven deep models, including three generic CNN models: MobileNet, ResNet, XceptionNet, as well as two popular Transformer-based models: ViT and CLIP, plus two FER-specific state-of-the-art models: POSTER and CEPrompt. Unlike prior studies that examine only limited aspects of bias, our work introduces a unified evaluation framework for FER that integrates five existing and two newly proposed dataset metrics with four fairness criteria for model analysis. We further introduce two new metrics, Conditional-Entropy Bias Index and Concentration Index, designed to quantify conditional dependencies and intra-group data imbalance that existing measures fail to capture. Our results show that all four datasets carry significant demographic bias, most notably in race, with AffectNet exhibiting the highest overall bias and Fer2013 the lowest. At the model level, we find that residual-based CNN architectures (ResNet and XceptionNet) exhibit the lowest overall bias, whereas Transformer-based models (ViT and CLIP) exhibit the highest, despite often achieving comparable or superior accuracy. These findings demonstrate that high predictive accuracy does not guarantee fairness, and that dataset-level and model-level bias must be addressed jointly rather than in isolation.