Realizable H-Consistent and Bayes-Consistent Loss Functions for Learning to Defer