Using Time-Series Privileged Information for Provably Efficient Learning of Prediction Models