Predictive Fair Representation Learning with Variational Autoencoders

Abstract Fairness is a critical concern in machine learning, as biases in prediction outcomes with respect to sensitive attributes such as gender and race have raised ethical and societal issues in real-world applications. To address this, fair representation learning aims to extract essential information from data as representations that are useful but independent of sensitive attributes. However, existing methods often suffer from limited predictive performance when their learned representations are applied to downstream tasks. In this study, we propose predictive FairDisCo (PdFairDisCo), an extension of FairDisCo to learn fair representations with variational autoencoders. PdFairDisCo enhances predictive performance by incorporating contrastive losses into the objective function. In addition, we introduce two oversampling methods for PdFairDisCo to mitigate bias by balancing the proportion of the sensitive attribute in training data. We demonstrate the effectiveness of PdFairDisCo through experiments. The experiments also show that the oversampling methods can further improve the performance of fairness.