Efficient Policy Learning from Surrogate-Loss Classifications