Performance and Preferences: Interactive Refinement of Machine Learning Procedures