Building similarity metrics reflecting utility in case-based reasoning