Learning Fair Representations via Rate-Distortion Maximization

Abstract Text representations learned by machine learning models often encode undesirable demographic information of the user. Predictive based on these can rely such information, resulting in biased decisions. We present a novel debiasing technique, Fairness-aware Rate Maximization (FaRM), that removes protected making instances belonging to same attribute class uncorrelated, using rate-distortion function. FaRM is able debias with or without target task at hand. also be adapted remove about multiple attributes simultaneously. Empirical evaluations show achieves state-of-the-art performance several datasets, and leak significantly less against an attack non-linear probing network.