随机性假设下的 conformal 预测普遍性
机器学习
2025-06-10 v2 统计理论
统计理论
摘要
conformal 预测提供的集合或函数预测在随机性假设（即独立同分布数据假设）下具有有效性。本文提出的问题是：是否存在在相同随机性假设下有效且比conformal 预测更高效的预测器？答案表明，conformal 预测器的类在仅限于在预测效率方面获得有限提升的情况下具有普遍性。此前的研究依赖于随机性的算法理论，涉及未指定的常数，而本文的结果更为实用。它们在某些方面也被证明是最优的。
引用
@article{arxiv.2502.19254,
title = {Universality of conformal prediction under the assumption of randomness},
author = {Vladimir Vovk},
journal= {arXiv preprint arXiv:2502.19254},
year = {2025}
}
备注
24 pages, 2 figures; changes since v1: exposition simplified, applications to classification extended and new optimality results added, applications to regression (Sect. 5 of v1) removed (the results in that section were correct but weak and less interesting than the new results)