Counterfactual Adversarial Learning for Recommendation