学院动态

学术报告信息发布

当前位置: 首页 >> 学术报告信息发布 >> 正文

A locally sequentially reweighted gradient descent estimator to enhance statistical efficiency for decentralized federated learning

发布日期:2026-09-06    作者:     点击:

报告题目:A locally sequentially reweighted gradient descent estimator to enhance statistical efficiency for decentralized federated learning

报告时间:202696日下午16:30

报告地点:南湖校区办公楼五楼会议室1561

主办单位:数学与统计学部、大数据科学研究院

报告人:张宝学

报告人简介:首都经济贸易大学统计与数据科学学院二级教授、博士生导师。担任全国应用统计专业学位研究生教育指导委员会委员兼案例组组长,全国工业统计学教学研究会副会长,中国现场统计研究会副理事长。《数理统计与管理》、《统计与信息论坛》、《统计与决策》以及《运筹与模糊数学》等期刊的编委。主要从事高维数据的统计理论方法及应用的研究。先后主持国家自然科学基金项目5项,教育部博士点基金1项,吉林省应用统计学研究创新团队1项,参加国家自然科学基金委重点项目1项,教育部创新团队项目1项。在 《Genome Research》、《Bioinformatics》、《Pattern Recognition 》、《Statistica Sinica》、《Scandinavian Journal of Statistics》和《Science China Mathematics》等国内外著名杂志发表或接受论文90余篇。其中,SCI/SSCI索引论文70余篇。获北京市教育系统“教书育人榜样”,北京市宣传思想文化系统“四个一批”人才等荣誉称号。曾入选教育部新世纪优秀人才支持计划,曾荣获教育部高等学校自然科学奖二等奖吉林省科学技术奖自然科学奖二等奖。

摘要:While many studies have considered the numerical convergence of federated learning algorithms, far less attention has been given to their statistical convergence.  In this paper,  to enhance statistical efficiency, we propose a novel Locally Sequentially Re-weighted Gradient Descent (LSRGD) estimator for decentralized federated learning. Furthermore, we prove that the LSRGD estimator is asymptotically normal and achieves optimal statistical efficiency. Moreover, we also propose a parallel version of the LSRGD algorithm, referred to as LSRGD-P. Finally, extensive experiments demonstrate that LSRGD and LSRGD-P estimators exhibit superior statistical efficiency compared to existing competitors. This advantage is particularly pronounced in scenarios where the data across different clients are imbalanced.


下一条:供应链数字化与供应链配置模式选择:集中化还是多元化