Conformalized Interval Arithmetic with Symmetric Calibration

Rui Luo, Zhixin Zhou

Abstract

Conformal Prediction (CP) provides distribution-free uncertainty quantification for single targets, but extending it to aggregate metrics like sums or averages is non-trivial. This paper introduces Conformal Interval Arithmetic (CIA), a framework that uses a novel "Symmetric Calibration" technique to generate tight, mathematically guaranteed prediction intervals for aggregate labels. CIA outperforms standard interval arithmetic and Bonferroni corrections, maintaining valid coverage while significantly reducing interval width across various real-world datasets including bike sharing, crime rates, and traffic networks.

Introduction

Conformal Prediction is a popular tool for uncertainty quantification because it provides "valid" coverage—meaning if you want a 95% confidence interval, the true value is guaranteed to fall within that interval 95% of the time, regardless of the underlying data distribution.

However, traditional CP is designed for single targets. In many real-world scenarios, we care about the aggregate: for example, the total travel time for a route consisting of multiple road segments, or the average crime rate across a specific set of counties. Simply adding up the individual prediction intervals for each component (standard interval arithmetic) often fails; the resulting sum might be uselessly wide or fail to meet the coverage guarantee if errors are correlated. CIA fills this gap by calibrating the uncertainty of the sum directly.

Methods

CIA calculates conformity scores based on the absolute residual of the sum of labels within a group. The core innovation is Symmetric Calibration. Standard CP assumes data points are exchangeable. In aggregate tasks, calibration and test samples might belong to the same network or group. Symmetric calibration treats the assignment of indices to the calibration or test set as a random coin flip, ensuring that the residuals of the sums share the same distribution.

Schematic of Symmetric Calibration (Theorem 1) vs standard split-conformal prediction (Proposition 1).

To improve efficiency, the authors also introduce Stratified CIA, which groups test sets by the number of unknown labels they contain. This prevents a "one-size-fits-all" interval that might be too wide for small groups and too narrow for large ones. Additionally, the method can be combined with Conformal Quantile Regression (CQR) to handle heteroscedasticity.

Key Results

The researchers evaluated CIA on diverse datasets: Bike Sharing demand, Community Crime rates, Medical Expenditure (MEPS), and Road Traffic networks (Anaheim & Chicago).

In all experiments, CIA consistently hit the target coverage rate almost perfectly. Compared to the "Group" baseline and "Bonferroni" correction, CIA produced much smaller interval sizes. For instance, in the Bike Sharing and MEPS datasets, CIA's intervals were significantly tighter while maintaining the same reliability.

Performance comparison showing coverage and interval size for CIA against baselines on bike, community, and meps21 datasets.

On the Anaheim and Chicago road datasets, CIA outperformed "Normal Confidence Intervals." Because real-world traffic data is rarely perfectly Gaussian, the Normal method often failed to provide the promised coverage, whereas CIA remained valid.

Discussion and Future Work

CIA bypasses the need to calculate complex convolutions of density functions by focusing on the exchangeability of the groups themselves. While the paper provides a theoretical bound for overlapping subsets, the validity of the interval degrades as the degree of overlap increases. Future research could explore non-symmetric functions or more complex aggregate statistics in decision-making pipelines. Symmetric calibration remains a key tool for ensuring coverage even in transductive settings where traditional split-conformal methods struggle.

Summary updated

Abstract

Uncertainty quantification is essential in decision-making, especially when joint distributions of random variables are involved. While conformal prediction provides distribution-free prediction sets with valid coverage guarantees, it traditionally focuses on single predictions. This paper introduces novel conformal prediction methods for estimating the sum or average of unknown labels over specific index sets. We develop conformal prediction intervals for single target to the prediction interval for sum of multiple targets. Under permutation invariant assumptions, we prove the validity of our proposed method. We also apply our algorithms on class average estimation and path cost prediction tasks, and we show that our method outperforms existing conformalized approaches as well as non-conformal approaches. Copyright © 2025, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.
BibTeX
@inproceedings{luo2025conformalized,
  author = {Rui Luo and Zhixin Zhou},
  title = {{Conformalized Interval Arithmetic with Symmetric Calibration}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i18.34114},
}
Figure