Learning Global Transparent Models consistent with Local Contrastive Explanations