Improving Users' Mental Model with Attention-directed Counterfactual Edits