Fair Transfer Learning with Factor Variational Auto-Encoder

Recently, in the field of fair machine learning, a large number of studies have considered how to remove discriminatory information from the data and achieve fairness in downstream tasks. Fair representation learning considers removing sensitive information (e.g. race, gender, etc) in the latent space, and the learned representations can prevent machine learning systems from being biased by discriminatory information. In this paper, we study the problems of existing methods and propose a novel fair representation learning method for the fair transfer learning where the labels of the downstream tasks are unknown. Specifically, we bring a new training model with information-theoretically motivated objective which avoids the problem of alignment for learning disentangled fair representations. Empirical results in various settings demonstrate the broad applicability and utility of our approach.