Counterfactual Fairness for Facial Expression Recognition

Abstract

Given the increasing prevalence of facial analysis technology, the problem of bias in these tools is becoming an even greater source of concern. Causality has been proposed as a method to address the problem of bias, giving rise to the popularity of using counterfactuals as a bias mitigation tool. In this paper, we undertake a systematic investigation of the usage of counterfactuals to achieve both statistical and causal-based fairness in facial expression recognition. We explore bias mitigation strategies with counterfactual data augmentation at the preprocessing, in-processing, and post-processing stages as well as a stacked approach that combines all three methods. At the in-processing stage, we propose using Siamese Networks to suppress the differences between the predictions on the original and the counterfactual images. Our experimental results on RAF-DB with counterfactuals added show that: (1) The in-processing method outperforms at the pre-processing and postprocessing stages, in terms of accuracy, F1 score, statistical fairness and counterfactual fairness, and (2) stacking the pre-processing, in-processing and post-processing stages provides the best performance.