# Conformal Risk Control

**Authors:** Anastasios N. Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, Tal Schuster  
**arXiv:** [2208.02814](https://arxiv.org/abs/2208.02814)  
**Date:** 2022-08-04  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2208.02814v1/

## Key points

- The paper introduces an extension of conformal prediction for controlling the expected value of any monotone loss function
- The algorithm builds upon split conformal prediction and includes a coverage guarantee
- Similar to traditional conformal prediction, the proposed procedure is accurate up to an $\mathcal{O}(1/n)$ factor
- Worked examples in computer vision and natural language processing demonstrate the effectiveness of the algorithm
- The approach can be used to bound metrics such as false negative rate, graph distance, and token-level F1-score
- The paper expands on conformal prediction techniques and introduces a method for controlling different loss functions' expected values
- Practical applications in computer vision and natural language processing highlight the potential usefulness of this approach in real-world scenarios
- Suitable for use with large datasets due to its accuracy up to an $\mathcal{O}(1/n)$ factor

## Summary

The paper titled "Conformal Risk Control" introduces an extension of conformal prediction that allows for the control of the expected value of any monotone loss function. The algorithm presented in the paper builds upon split conformal prediction and includes a coverage guarantee. Similar to traditional conformal prediction, the proposed conformal risk control procedure is accurate up to an $\mathcal{O}(1/n)$ factor. The authors demonstrate the effectiveness of their algorithm through worked examples in computer vision and natural language processing. Specifically, they showcase how their approach can be used to bound various metrics such as the false negative rate, graph distance, and token-level F1-score. Overall, this paper provides a valuable contribution by expanding on conformal prediction techniques and introducing a method for controlling the expected value of different loss functions. The practical applications demonstrated in computer vision and natural language processing highlight the potential usefulness of this approach in real-world scenarios. Moreover, its accuracy up to an $\mathcal{O}(1/n)$ factor makes it suitable for use with large datasets.