[Paper Review] Assessing Demographic Bias Transfer from Dataset to Model: A Case Study in Facial Expression Recognition

Iris Dominguez-Catena, Daniel Paternain+1 more|arXiv (Cornell University)|May 20, 2022

Emotion and Mood Recognition4 citations

TL;DR

This paper proposes a novel metric framework to assess demographic bias transfer from datasets to models in facial expression recognition (FER), focusing on representational imbalance and stereotypical bias in race and gender. Using the AffectNet dataset, it reveals significant racial underrepresentation and gender stereotyping, showing that gender bias strongly transfers to models—while racial bias does not significantly impact model performance even after dataset balancing.

ABSTRACT

The increasing amount of applications of Artificial Intelligence (AI) has led researchers to study the social impact of these technologies and evaluate their fairness. Unfortunately, current fairness metrics are hard to apply in multi-class multi-demographic classification problems, such as Facial Expression Recognition (FER). We propose a new set of metrics to approach these problems. Of the three metrics proposed, two focus on the representational and stereotypical bias of the dataset, and the third one on the residual bias of the trained model. These metrics combined can potentially be used to study and compare diverse bias mitigation methods. We demonstrate the usefulness of the metrics by applying them to a FER problem based on the popular Affectnet dataset. Like many other datasets for FER, Affectnet is a large Internet-sourced dataset with 291,651 labeled images. Obtaining images from the Internet raises some concerns over the fairness of any system trained on this data and its ability to generalize properly to diverse populations. We first analyze the dataset and some variants, finding substantial racial bias and gender stereotypes. We then extract several subsets with different demographic properties and train a model on each one, observing the amount of residual bias in the different setups. We also provide a second analysis on a different dataset, FER+.

Motivation & Objective

• To address the lack of fairness metrics for multi-class, multi-demographic problems like FER, where existing metrics fail to isolate dataset vs. model bias.
• To develop a systematic methodology to measure bias transfer from training data to trained models, enabling evaluation of mitigation strategies.
• To analyze real-world FER datasets—particularly AffectNet—for demographic bias, including representational imbalance and stereotypical associations.
• To evaluate the impact of dataset-level bias mitigation (e.g., balancing by race or gender) on the residual bias of trained models.
• To extend analysis to a second dataset (FER+) to validate findings across different data distributions.

Proposed method

• Proposes three new metrics: two for dataset bias (representational imbalance and stereotypical bias via NMI/NPMI), and one for model residual bias.
• Uses FairFace, a pretrained model, to infer apparent race and gender from face images in AffectNet and FER+ when ground-truth demographics are unavailable.
• Applies normalized mutual information (NMI) and normalized pointwise mutual information (NPMI) to quantify stereotypical associations between demographic groups and expression classes.
• Measures model bias via a single-value metric (OD) that captures recall variation across demographic groups per expression class.
• Trains models on multiple dataset variants: stratified, race-balanced, and gender-balanced subsets to compare bias transfer.
• Employs quantitative analysis across training sizes to assess the impact of data quantity and quality on bias propagation.

Experimental results

Research questions

• RQ1 To what extent does representational imbalance in the AffectNet dataset affect model fairness in facial expression recognition?
• RQ2 How do stereotypical biases—where certain demographic groups are over- or under-associated with specific expressions—affect model predictions?
• RQ3 Does balancing the dataset by race or gender reduce residual bias in the trained model, and if so, which demographic factor has a stronger impact?
• RQ4 How does training data size influence the transfer of bias from dataset to model, independent of representational or stereotypical bias?
• RQ5 Are the proposed metrics effective in detecting and quantifying bias transfer across different FER datasets and model configurations?

Key findings

• The AffectNet dataset exhibits substantial representational imbalance, with significant underrepresentation of certain racial groups and overrepresentation of others.
• Stereotypical bias is pronounced in gender associations: females are disproportionately linked to certain expressions (e.g., 'happy' or 'sad'), while males are underrepresented in emotional categories.
• Despite balancing the dataset for race, the trained model’s racial bias remains largely unchanged, suggesting that racial bias transfer is not easily mitigated by simple balancing.
• Gender-balanced datasets significantly reduce model bias, indicating that gender stereotyping in the data strongly influences model predictions.
• Model bias decreases with increasing training size, even when representational and stereotypical biases are held constant, suggesting data quantity helps mitigate bias.
• The proposed metrics successfully quantify bias transfer, revealing that gender bias has a more substantial impact on model fairness than racial bias in this FER context.