Conditional Coverage Estimation for High-Quality Prediction Intervals

作     者：Ziyi Huang Henry Lam Haofeng Zhang Ziyi Huang;Henry Lam;Haofeng Zhang

作者机构：School of Engineering and Applied ScienceColumbia UniversityNew YorkNY 10027USA 

出 版 物：《Journal of Systems Science and Systems Engineering》 (系统科学与系统工程学报（英文版）)

年 卷 期：2023年第32卷第3期

页      面：289-319页

核心收录：

学科分类：0710[理学-生物学] 02[经济学] 0202[经济学-应用经济学] 020208[经济学-统计学] 07[理学] 0714[理学-统计学（可授理学、经济学学位）] 070103[理学-概率论与数理统计] 0701[理学-数学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：the National Science Foundation under grants CAREER CMMI-1834710 and IS-1849280 The research of Ziyi Huang and Haofeng Zhang is supported in part by the Cheung-Kong Innovation Doctoral Fellowship 

主　　题：Uncertainty quantification prediction intervals conditional coverage neural networks calibrationerror 

摘      要：Deep learning has been recently studied to generate high-quality prediction intervals(PIs)for uncertainty quantification in regression tasks,including recent applications in simulation *** high-quality criterion requires PIs to be as narrow as possible,whilst maintaining a pre-specified level of data(marginal)***,most existing works for high-quality PIs lack accurate information on conditional coverage,which may cause unreliable predictions if it is significantly smaller than the marginal *** address this problem,we propose an end-to-end framework which could output high-quality PIs and simultaneously provide their conditional coverage *** doing so,we design a new loss function that is both easy-to-implement and theoretically justified via an exponential concentration *** evaluation on real-world benchmark datasets and synthetic examples shows that our approach not only achieves competitive results on high-quality PIs in terms of average PI width,but also accurately estimates conditional coverage information that is useful in assessing model uncertainty.