报告人: 王晓飞副教授（东北师范大学）

报告时间：2023年11月18日，14:20-15:10

报告地点：金融工程研究中心105报告厅

报告摘要：Feature selection in linear models is widely applied to diverse fields. Many traditional approaches can behave well based on the penalized likelihood. However, the selection of hyperparameters or tuning parameters is often critical and challenging. From the oracle viewpoint for attaining an ideal risk, we design a global adaptive generative adjustment algorithm, which can adaptively learn multiple tuning parameters in the personalized Tikhonov regularization and further select features by a personalized thresholding strategy. We prove that the output of our algorithm has the consistency for both the model selection and the signal estimate. Specially, for the signal estimate, our method provides a lower error bound than traditional penalized likelihood methods. Finally, in the numerical experiment, we compared our algorithms to the adaptive LASSO, the SCAD, and the MCP. The experiment results affifirmed the effificiency of our algorithms for feature selection and demonstrated the superiority of our algorithms over other methods.

个人简介：王晓飞，东北师范大学，数学与统计学院，统计系，副教授，博士生导师。主要研究方向：图算法，图模型和机器学习。在国际期刊上发表多篇学术论文。主持国家自然科学基金委项目天元，青年，面上项目各一项，参与国家自然科学基金委项目多项，主持中央高校青年教师科研发展项目一项，主持吉林省优秀青年人才基金项目一项。

邀请人：刘芳，徐礼柏