Learning Fair Representations via an Adversarial Framework