Lifting Interpretability-Performance Trade-off via Automated Feature Engineering