Fast, Accurate, and Simple Models for Tabular Data via Augmented Distillation

Neural Information Processing Systems 

While highly accurate, the resulting predictors are large, slow, and opaque as compared to their constituents.