Balancing Competing Objectives for Welfare-Aware Machine Learning with Imperfect Data