Improved Adversarial Learning for Fair Classification