Paper Overview
Field: Machine Learning Authors: Ivan Petej, Vladimir Vovk Posted: 2026-05-07 arXiv: 2605.06646
Summary
Venn-Abers predictors are probabilistic predictors with attractive properties, most notably validity (calibration guarantees). Their main limitation has been that they were originally applicable only to binary classification; they have recently been extended to bounded regression.
This paper generalizes Venn-Abers predictors to the setting of unbounded regression, which requires incorporating elements of conformal prediction.
In simulation and empirical studies, the authors investigate the predictive efficiency of point regressors derived from Venn-Abers regressors. They argue that these derived regressors improve, to some extent, on the predictive efficiency of standard regressors when working with larger training sets.
Key Contributions
- Extends Venn-Abers predictors beyond binary classification and bounded regression to unbounded regression
- Introduces conformal prediction elements to enable this generalization
- Empirically evaluates point regressors derived from Venn-Abers regressors, showing improved predictive efficiency over standard regressors on larger training sets
*Auto-collected on 2026-05-10.*