Learning Temporal Preferences