![[2602.20151] Conformal Risk Control for Non-Monotonic Losses](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.20151] Conformal Risk Control for Non-Monotonic Losses
Summary
This article presents a novel approach to conformal risk control for non-monotonic losses, extending traditional methods to multidimensional parameters and offering applications in various fields such as image classification and recidivism prediction.
Why It Matters
The research addresses limitations in existing risk control methods by providing guarantees for non-monotonic losses, which are common in real-world applications. This advancement can enhance decision-making processes in critical areas like healthcare and criminal justice, where accurate predictions are essential.
Key Takeaways
- Introduces conformal risk control for non-monotonic losses.
- Provides stability-based guarantees for algorithms in multidimensional contexts.
- Demonstrates applications in selective image classification and recidivism predictions.
Statistics > Methodology arXiv:2602.20151 (stat) [Submitted on 23 Feb 2026] Title:Conformal Risk Control for Non-Monotonic Losses Authors:Anastasios N. Angelopoulos View a PDF of the paper titled Conformal Risk Control for Non-Monotonic Losses, by Anastasios N. Angelopoulos View PDF HTML (experimental) Abstract:Conformal risk control is an extension of conformal prediction for controlling risk functions beyond miscoverage. The original algorithm controls the expected value of a loss that is monotonic in a one-dimensional parameter. Here, we present risk control guarantees for generic algorithms applied to possibly non-monotonic losses with multidimensional parameters. The guarantees depend on the stability of the algorithm -- unstable algorithms have looser guarantees. We give applications of this technique to selective image classification, FDR and IOU control of tumor segmentations, and multigroup debiasing of recidivism predictions across overlapping race and sex groups using empirical risk minimization. Subjects: Methodology (stat.ME); Machine Learning (cs.LG); Statistics Theory (math.ST); Machine Learning (stat.ML) Cite as: arXiv:2602.20151 [stat.ME] (or arXiv:2602.20151v1 [stat.ME] for this version) https://doi.org/10.48550/arXiv.2602.20151 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Anastasios Angelopoulos [view email] [v1] Mon, 23 Feb 2026 18:58:54 UTC (523 KB) Full-text links: Access Paper: View a PDF of th...
