大数据分位数回归下基于信息阵的最优子抽样  

Information matrix based optimal subsampling for big data quntile regression

作　　者：袁晓惠 郭伟 王纯杰 YUAN Xiao-hui;GUO Wei;WANG Chun-jie(School of Mathematics and Statistics,Changchun University of Technology,Changchun 130012,China)

机构地区：[1]长春工业大学数学与统计学院,吉林长春130012

出　　处：《东北师大学报（自然科学版）》2023年第3期30-36,共7页Journal of Northeast Normal University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(12271060);吉林省教育厅科学研究项目(JJKH20230749KJ).

摘　　要：基于D-optimality准则构造了分位数回归中基于信息阵的最优子抽样方法.从理论角度证明了此方法相较于以往方法存在优势,即所含信息量会随总样本量的增加而增加.通过模拟实验比较了新方法和均匀子抽样、基于L-optimality准则的子抽样在不同情况下的性能.实证分析表明了新方法在实践中表现出色.The optimal subsampling strategy based on the information matrix in the quantile regression model under the D-optimality criterion is constructed.From a theoretical perspective,the advantage of this sampling method over previous subsampling methods is that the information contained in this method increases as the total sample size increases.The performance of the new method is compared with uniform subsampling and subsampling based on L-optimality criterion in different cases by simulation experiments.Finally,real data analysis shows that the new method works well in practice.

关 键 词：大数据 分位数回归 D-optimality准则 信息矩阵 

分 类 号：O212.1[理学—概率论与数理统计]