Bias Mitigation through Proxy Sensitive Attribute Label Generation

Existing bias mitigation algorithms are limited in their applicability to real-world scenarios because they require access to sensitive attributes to achieve fairness. In practice, these attributes are unavailable due to legal and policy regulations or data unavailability for a given demographic. This paper aims to overcome this obstacle by introducing an unsupervised technique for generating proxy-sensitive attribute labels. Our approach involves two stages: unsupervised embedding generation and clustering, resulting in universally applicable proxies for sensitive labels that can be used with existing bias mitigation techniques. We believe that bias is present in non-sensitive attributes correlated to the sensitive attributes, and clusters representing different demographic groups can be identified by mapping them to a high-dimensional latent space. Experimental results demonstrate that the proxies generated by our proposed method can be effectively utilized with existing bias mitigation techniques and produce similar outcomes compared to using true sensitive attributes, making it a promising solution to combat bias in machine learning models.