ReCIRC: Rectified Conformal Risk Control

Imagine you're developing a medical imaging tool that needs to accurately identify lesions in scans. The challenge lies in ensuring that the model doesn't miss critical areas while also not overwhelming users with false positives. Currently, many models use a single threshold for decision-making, which can lead to problems: it might be too lenient for easy cases, allowing errors, or too strict for difficult cases, resulting in missed detections. This is what's called a marginal guarantee, where the model's performance varies significantly based on the input's complexity. To address this, the authors propose a method called ReCIRC, which tailors the risk threshold for each input based on its estimated local risk. By reparameterizing the threshold as a risk budget, ReCIRC maintains the benefits of existing conformal risk control while allowing for more precise adjustments based on the input's characteristics. This means that when the risk estimates are accurate, the model can achieve nearly exact control over conditional risks, improving performance across various tasks like segmentation and classification. Compared to previous methods, ReCIRC not only retains the original guarantees but also enhances the model's ability to adapt to different scenarios, making it a more reliable tool for real-world applications.

ReCIRC introduces a new way to manage conditional risk in predictive models.

The method is validated across multiple datasets and settings, showing consistent performance.