Domain adaptation for bias mitigation in affective computing: use cases for facial emotion recognition and sentiment analysis systems

This study presents a novel approach using the Demographic Bias Mitigation Framework (DBMF), which leverages Domain Adaptation (DA) to mitigate demographic biases in affective computing AI systems. Biases in technologies like sentiment analysis (SA) and facial emotion recognition (FER) can result in adverse societal impacts and diminished trust. Unlike typical DA strategies, which adapt data from sources to targets, the DBMF innovatively adapts less biased target data to biased source domains to address gender bias in Natural Language Processing (NLP-SA) and racial bias in Computer Vision (CV-FER) tasks. Statistical methods and fairness metrics confirm the framework’s effectiveness in reducing bias while preserving task performance. Notably, for the CV-FER task, the DBMF achieves state-of-the-art accuracy for facial emotion recognition on a widely used dataset—SFEW2.0—marking a significant advancement in this domain. The framework’s ability to handle diverse tasks, domains, and bias types highlights its potential as a unified solution for bias mitigation. Additionally, the integration of Elastic Weight Consolidation (EWC) ensures the retention of task performance across domains, further reinforcing the framework’s robustness. These findings emphasize the DBMF’s and DA’s critical role in fostering fairness, reliability, and trustworthiness in AI-driven affective computing systems.