The authors propose a new method for conformal risk control that provides formal guarantees when the data is not exchangeable, while also achieving the same guarantees as existing methods if the data is exchangeable. This framework is flexible and allows weighting the data based on its statistical similarity with the test examples. The method has been tested using both synthetic and real-world data, demonstrating its effectiveness and utility in various scenarios.
Publication date: 2 Oct 2023
Project Page: https://arxiv.org/abs/2310.01262
Paper: https://arxiv.org/pdf/2310.01262
