Learning Global Transparent Models Consistent with Local Contrastive Explanations