A Fair Generative Model Using Total Variation Distance

Soobin Um, Changho Suh

Keywords: trustworthy AI, fairness, generative model, total variation distance

Abstract: We explore a fairness-related challenge that arises in generative models. The challenge is that biased training data with imbalanced representations of demographic groups may yield a high asymmetry in size of generated samples across distinct groups. We focus on practically-relevant scenarios wherein demographic labels are not available and therefore the design of a fair generative model is particularly challenging. In this paper, we propose an optimization framework that regulates such unfairness by employing one prominent statistical notion, total variation distance (TVD). We quantify the degree of unfairness via the TVD between the generated samples and balanced-yet-small reference samples. We take a variational optimization approach to faithfully implement the TVD-based measure. Experiments on benchmark real datasets demonstrate that the proposed framework can significantly improve the fairness performance while maintaining realistic sample quality for a wide range of the reference set size all the way down to 1% relative to training set.

One-sentence Summary: We propose a generative model framework that regulates imbalanced representations of demographics via total variation distance measure.