Conformalised imprecise inference for robust extrapolation under limited data
2026-05-25 • Machine Learning
Machine Learning
AI summaryⓘ
The authors focus on improving how machine learning models handle uncertainty, especially when making predictions on data they haven't seen before. They introduce a new method that adds a layer of cautiousness by giving predictions as ranges (called probability boxes) instead of single values. This method is flexible and works with different models, helping it stay reliable even when the data changes. Tests show it performs better at knowing when it might be wrong, especially with less data to learn from.
uncertainty quantificationaleatory uncertaintyepistemic uncertaintydistributional shiftconformal inferenceimprecise probabilitiesprobability boxesextrapolationmodel-agnostic methodscoverage guarantee
Authors
Yu Chen, Scott Ferson
Abstract
Recent advances in uncertainty quantification increasingly emphasise the distinction between aleatory and epistemic uncertainty in machine learning, motivating the need for more unified frameworks. However, despite much progress in producing reliable predictions, existing methods often lack rigorous guarantees when generalising beyond the training domain. We propose a conformalised imprecise inference framework for robust extrapolation, which is model-agnostic and augments predictive models with imprecision and distance awareness. The proposed approach yields imprecise predictions (probability boxes) that remain valid under distributional shift, maintaining coverage while adaptively expanding uncertainty in extrapolation regimes. Experiments on synthetic and benchmark datasets demonstrate improved robustness and reliable coverage compared to standard probabilistic approaches, particularly under limited data.