Decision-Theoretic Approximations for Machine Learning