Informational and Computational Efficiency of Set Predictors

There are two methods of set prediction that are provably valid under the assumption of randomness: transductive conformal prediction and inductive conformal prediction. The former method is informationally efficient but often lacks computational efficiency. The latter method is, vice versa, computationally efficient but less efficient informationally. This talk discusses a new method, which we call cross-conformal prediction, that combines informational efficiency of transductive conformal prediction with computational efficiency of inductive conformal prediction. The downside of the new method is that its validity is an empirical rather than mathematical fact.