Active Learning in the Predict-then-Optimize Framework: A Margin-Based Approach