Group-Conditional Conformal Risk Control: Exact Guarantees, Limits, and Overlapping Extensions
Organizations: Department of Mathematical Sciences New Jersey Institute of Technology, Newark, NJ, USA
Abstract
We study group-conditional conformal risk control (CRC) for disjoint and overlapping groups under bounded monotone losses. For disjoint groups, we establish exact finite-sample conditional risk control, two-sided bounds on the group-restricted risk, and fundamental limitations arising from local support and calibration-rank granularity. For overlapping groups, we study two exactly valid approaches: group-restricted calibration followed by union deployment, and direct calibration of the deployed prediction-set family. We characterize the union's limiting reconciliation gap, identify conditions for its disappearance, and establish sharp lower bounds for structured losses in specified product-form models. We also demonstrate a strict size advantage for direct calibration in an ordinal model. Experiments show that direct calibration can reduce prediction-set size, but its efficiency depends on group structure, loss type, and shape fitting. Results on historical diabetic-retinopathy scores further illustrate these tradeoffs. Overall, local support, reconciliation slack, and threshold heterogeneity determine efficiency, with neither approach uniformly dominating.
Figures & tables
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Severity group | Marg. risk | Mond. risk | Restricted risk | Marg. size | Mond. size | Fallback | |
|---|---|---|---|---|---|---|---|
| No DR | 594.4 | 0.040 | 0.057 | 0.098 | 1.47 | 2.19 | 0.00 |
| Mild/mod. | 507.3 | 0.143 | 0.055 | 0.098 | 1.68 | 2.51 | 0.00 |
| Severe/PDR | 112.3 | 0.208 | 0.036 | 0.079 | 1.42 | 2.05 | 0.00 |
| Clinical group | Marg. risk | Grade risk | Union risk | Restricted risk | Union size | ||
|---|---|---|---|---|---|---|---|
| No DR | 594.4 | 0.040 | 0.031 | 0.058 | 0.098 | 0.041 | 2.18 |
| NPDR | 530.8 | 0.148 | 0.023 | 0.056 | 0.100 | 0.044 | 2.48 |
| Referable | 557.2 | 0.135 | 0.023 | 0.038 | 0.098 | 0.060 | 2.34 |
| Severe/PDR | 112.3 | 0.208 | 0.017 | 0.036 | 0.079 | 0.044 | 2.05 |
| Grade | 23.5 | 0.248 | 0.000 | 0.041 | – | – | 2.36 |
Conformal Risk Control
Conformal risk control uses calibration data to choose prediction or action thresholds with finite-sample guarantees on expected error or other risks.
See all 72 papers