-
活动&公告
活动&公告
报告时间： 2022-11-04 14:00——2022-11-04 16:00
报告地点： 腾讯会议： 745-406-998
报告题目：When Transelliptical Graphical Models meet Transfer Learning.
报告人：梁万丰 博士
主办单位：数据科学与人工智能学院 数理统计研究中心
【专家简介】
梁万丰，南开大学统计与数据科学学院博士生。主要研究高维数据背景下的协方差矩阵和精度矩阵的估计方法。发表Statistics and Probability Letters一篇, Stat三篇。在校期间多次获得多种奖学金等。
【讲座摘要】
This talk considers estimating a transelliptical graphical model in the context of transfer learning where auxiliary samples (typically, the large majority) from different but possibly related distributions are available in addition to observed target data. We propose a three-stage method called Trans-TE-CLIME to achieve modeling flexibility and estimation robustness simultaneously based on Kendall's tau statistic by utilizing information from auxiliary samples after defining a similarity measure. Theoretically, we derive the convergence rate under some mild conditions. The results illustrate that sufficiently related supplementary samples can help improve the rate of convergence and estimation accuracy. Computationally, the optimization procedure in each stage can be decomposed into some linear programming problems, and our method is scalable to large datasets. Empirically, simulation studies show that our proposal has outstanding numerical performance. We also apply our proposal to an S&P 500 stock dataset.
撰稿：王云龙 审核：富宇 单位：数据科学与人工智能学院
