報(bào)告人簡(jiǎn)介
李婷，上海財(cái)經(jīng)大學(xué)統(tǒng)計(jì)與管理學(xué)院副教授，本科畢業(yè)于華東師范大學(xué)統(tǒng)計(jì)系，博士畢業(yè)于復(fù)旦大學(xué)管理學(xué)院統(tǒng)計(jì)學(xué)系。曾赴美國(guó)德州大學(xué)MD安德森癌癥中心、香港中文大學(xué)統(tǒng)計(jì)系，北卡羅來(lái)納大學(xué)教堂山分校訪問(wèn)學(xué)習(xí)。研究方向包括函數(shù)型數(shù)據(jù)、醫(yī)學(xué)基因影像數(shù)據(jù)分析、分位數(shù)回歸和因果推斷。在Journal of the American Statistical Association， Annals of Applied Statistics, Statistica Sinica，Biometrics等統(tǒng)計(jì)學(xué)期刊發(fā)表以及人工智能頂會(huì)Neurips，ICML上發(fā)表過(guò)多篇論文。
內(nèi)容簡(jiǎn)介
Conformal prediction, a powerful framework that constructs a prediction band for the response variable using any regression function estimators, often faces the challenge of producing overly broad bands with limited target data. In this paper, we study the transfer learning problem in conformal prediction, aiming to improve the precision of the prediction interval of the target data with insufficient data by leveraging information from related auxiliary source datasets. Allowing for non-exchangeability between source datasets and the target dataset, two transfer conformal prediction algorithms have been proposed for scenarios with and without knowledge about the informative source data. Utilizing the conditional Kullback-Leibler divergence, the proposed algorithm effectively identifies relevant source data for transfer. We present a comprehensive analysis of the non-asymptotic theoretical properties of the proposed algorithms, including lower and upper bounds as well as the bounds on the width of the prediction bands, demonstrating the potential for efficiency gains through narrower intervals while preserving coverage accuracy. Empirical evidence, drawn from extensive simulations and real data analysis, further validates the effectiveness of our algorithms in improving prediction interval quality by leveraging source data, and achieving narrower intervals while maintaining desired coverage levels.
