Conformal sets in neural network regression

Abstract

This paper is concerned with predictive regions in regression models, especially neural networks. We use the concept of conformal prediction (CP) to construct regions which satisfy given confidence level. Conformal prediction outputs regions, which are automatically valid, but their width and therefore usefulness depends on the used nonconformity measure. A nonconformity measure should tell us how different a given example is with respect to other examples. We de ne nonconformity measures based on some reliability estimates such as variance of a bagged model or local modeling of prediction error. We also present results of testing CP based on different nonconformity measures showing their usefulness and comparing them to traditional confidence intervals.