首页|期刊导航|安庆师范大学学报(自然科学版)|变系数部分函数型贝叶斯复合分位数回归

变系数部分函数型贝叶斯复合分位数回归OA

Bayesian Composite Quantile Regression for Varying-Coefficient Partial Functional Linear Models

中文摘要英文摘要

针对变系数部分函数型线性模型的参数函数估计,传统最小二乘方法或单一分位数回归方法对异常值与厚尾误差稳健性不足情况,文章构建了一种贝叶斯复合分位数回归估计方法,融合多分位数信息以提升复杂数据场景下的估计精度与稳健性.首先,通过采用B样条基逼近变系数函数,函数型主成分基谱分解斜率函数,将无限维函数估计转化为有限维参数估计;假设误差服从复合非对称拉普拉斯分布,建立贝叶斯复合分位数回归形式,结合参数先验分布以推导未知参数联合后验分布.其次,基于Gibbs抽样算法对后验样本进行采样,通过基函数重构,最终得到变系数函数与斜率函数的估计量.数值模拟表明,不同误差分布下所提贝叶斯复合分位数回归方法在根均方误差与标准差准则上均优于传统复合分位数回归估计,实际数据建模亦展现了更优的估计精度.

To address the parameter function estimation in varying-coefficient partial functional linear models,where con-ventional methods lack robustness against outliers and heavy-tailed errors,this paper develops a Bayesian composite quantile regression approach.By integrating multi-quantile information,the proposed method enhances estimation accuracy and robust-ness in complex data scenarios.First,B-spline basis functions are employed to approximate the varying-coefficient functions,and the slope function undergoes spectral decomposition via the functional principal component basis,thereby transforming in-finite-dimensional functional estimation into finite-dimensional parameter estimation.Assuming that the errors follow a com-posite asymmetric Laplace distribution,we establish the Bayesian composite quantile regression framework and derive the joint posterior distribution of unknown parameters by incorporating prior distributions.Second,A Gibbs sampling algorithm is employed to generate posterior samples,and functional estimates are reconstructed via basis expansions.Numerical simula-tions demonstrate that,under various error distributions,the proposed Bayesian composite quantile regression method outper-forms traditional composite quantile regression in terms of.both root mean squared error and standard deviation criteria.Fur-thermore,real data modeling also exhibits superior estimation accuracy.

赵庭;李荣;雷启贵;李明艳

贵州民族大学 数据科学与信息工程学院,贵州 贵阳 550025贵州民族大学 数据科学与信息工程学院,贵州 贵阳 550025贵州民族大学 数据科学与信息工程学院,贵州 贵阳 550025贵州民族大学 数据科学与信息工程学院,贵州 贵阳 550025

数理科学

变系数部分函数型线性模型贝叶斯复合分位数回归B样条基函数型主成分基Gibbs抽样算法

varying-coefficient partial functional linear modelsBayesian composite quantile regressionB-spline basisfunctional principal component basisGibbs sampling algorithm

《安庆师范大学学报(自然科学版)》 2026 (2)

17-27,11

贵州省教育厅自然科学研究项目(黔教技[2022]015号)和贵州省高等学校大数据分析与智能计算重点实验室项目(黔教技[2023]012号)

10.13757/j.cnki.cn34-1328/n.2026.02.003

评论