Mapping Bias: Visualizing Valence-Arousal Distributions to Reveal Affective Gaps in Face Datasets

Abstract

Facial expression datasets play a foundational role in affective computing and emotion recognition. However, many rely on categorical emotion labels that may not reflect the true affective state captured in facial cues. The Valence–Arousal (VA) framework provides a dimensional alternative, mapping emotions along axes of polarity and activation. Despite its theoretical strengths, VA-based validation is rarely applied to large-scale facial datasets, raising questions about the fidelity of categorical annotations. This paper presents a systematic framework for auditing affective distributions in facial expression datasets through VA analysis. We evaluate two widely used datasets, DAISEE and AffectNet, by estimating VA scores with two state-of-the-art deep learning models, HSEmotion and EmoNet. Kernel density heatmaps and statistical measures, including mean valence, mean arousal, standard deviation, and 90% coverage ellipses, are used to assess alignment with Russell’s Circumplex Model of Affect. Our findings reveal significant inconsistencies: DAISEE’s engagement class and AffectNet’s fear class deviate substantially from their expected VA regions, while boredom and anger demonstrate more reliable clustering. These results expose systematic biases in dataset labeling that may compromise downstream affective models. By grounding dataset validation in continuous affective space, the proposed framework enhances transparency, supports ethical dataset design, and promotes the development of more robust and interpretable emotion recognition systems.