Towards Explainability for AI Fairness