On Shapley Credit Allocation for Interpretability