Introduction
Predicting aggregate values—such as the total travel time of a multi-segment route or the average demand across several locations—is a common task in logistics, urban planning, and public health. While Conformal Prediction (CP) offers robust, distribution-free uncertainty guarantees for individual predictions, applying it to sums or averages of multiple labels remains challenging. This paper introduces Conformal Interval Arithmetic (CIA), a framework designed to provide valid coverage for aggregate predictions even when individual labels are dependent or belong to overlapping groups.
The Challenge of Aggregate Uncertainty
Standard CP typically provides marginal coverage for a single label. When predicting the sum of several variables, two main issues arise:
- Statistical Dependency: If labels within a group are correlated, the joint probability of all labels falling within their respective individual intervals is difficult to bound.
- Efficiency Loss: A naive approach using Bonferroni correction to ensure joint coverage results in prediction intervals that are excessively wide and practically useless.
The core insight of this work is that the "sum of quantiles is not the quantile of the sum." To address this, the authors propose calibrating the uncertainty of the aggregate value directly.
Conformal Interval Arithmetic and Symmetric Calibration
The proposed CIA framework shifts the focus from individual data points to group-level residuals. By calculating a score based on the residual of the sum of labels within a group, the model inherently accounts for internal correlations.
A critical requirement for CP is exchangeability between calibration and test data. In aggregate settings, labels may appear in multiple groups, complicating this requirement. The authors introduce Symmetric Calibration, a technique that ensures residuals of sums share the same distribution by symmetrically assigning indices within a group to either the calibration or test set.
This framework is further enhanced by:
- Stratification: Grouping by size to improve interval efficiency.
- CQR Integration: Using Conformal Quantile Regression to handle heteroscedasticity.
- Overlapping Subsets: Providing theoretical bounds (Theorem 2) for cases where groups share members, such as different routes sharing road segments.
Empirical Performance
The authors evaluated CIA on diverse datasets, including bike-sharing demand, community crime rates, and medical expenditures.
Key results include:
- Reliable Coverage: CIA variants consistently met the target $1-\alpha$ coverage across all datasets. In contrast, "Normal" approximation methods frequently under-covered.
- Superior Efficiency: The CIA (CQR) Stratified method produced the narrowest intervals while maintaining validity, significantly outperforming the conservative Bonferroni baseline.
- Robustness in Networks: In traffic network experiments (Anaheim and Chicago), CIA maintained near-perfect coverage despite dependencies from overlapping paths, whereas baselines without symmetric calibration saw performance degradation as overlap increased.
Conclusion
Conformal Interval Arithmetic provides a principled way to quantify uncertainty for aggregate predictions. By treating the group residual as the fundamental unit of calibration and employing symmetric calibration, the method bypasses the need for complex joint distribution modeling. While the method's validity can degrade with very high overlap ratios, it remains a powerful tool for real-world applications requiring reliable group-level guarantees.
Summary updated