406

On training locally adaptive CP

International Symposium on Conformal and Probabilistic Prediction with Applications (ISCPPA), 2023
Abstract

We address the problem of making Conformal Prediction (CP) intervals locally adaptive. Most existing methods focus on approximating the object-conditional validity of the intervals by partitioning or re-weighting the calibration set. Our strategy is new and conceptually different. Instead of re-weighting the calibration data, we redefine the conformity measure through a trainable change of variables, AϕX(A)A \to \phi_X(A), that depends explicitly on the object attributes, XX. Under certain conditions and if ϕX\phi_X is monotonic in AA for any XX, the transformations produce prediction intervals that are guaranteed to be marginally valid and have XX-dependent sizes. We describe how to parameterize and train ϕX\phi_X to maximize the interval efficiency. Contrary to other CP-aware training methods, the objective function is smooth and can be minimized through standard gradient methods without approximations.

View on arXiv
Comments on this paper