Towards Knowledge-driven Distillation and Explanation of Black-box Models