Learning fair representations via an adversarial framework

Fairness has become a central issue for our research community as classification algorithms are adopted in societally critical domains such recidivism prediction and loan approval. In this work, we consider the potential bias based on protected attributes (e.g., race gender), tackle problem by learning latent representations of individuals that statistically indistinguishable between groups while sufficiently preserving other information classification.To do that, develop minimax adversarial framework with generator to capture data distribution generate representations, critic ensure distributions across different similar. Our provides theoretical guarantee respect statistical parity individual fairness. Empirical results four real-world datasets also show learned representation can effectively be used tasks credit risk obstructing related groups, especially when removing is not sufficient fair classification.