Feature Selection by the Principle of Correlation for High-dimensional Models with Complex Structures

报告人：陈泽华
报告地点：数学与统计学院415室
报告时间：2018年06月06日星期三10:00-11:30
邀请人：
报告摘要：

The intrinsic mechanism of feature selection is correlation. For instance, in sequential approaches, the features are selected according to their Pearson's correlation with the residual of a current model, in penalised likelihood approaches, at a _xed value of the penalty parameter, the active set is indeed the set of features whose Pearson's correlations with the response exceed a certain threshold. In this talk, we discuss the principle of correlation search and consider its application to feature selection
for high-dimensional models with complex structures. Speci_cally, we consider two such models: (i) a multi-response model where both the response variables and covariates have group structures, and (ii) an uni-response interaction model. For the _rst model, we develop a sequential canonical correlation search method. For the second model, we develop a sequential interaction group selection method. The asymptotic properties of these methods as well as simulation studies will be presented. These sequential methods can achieve selection consistency under meld conditions. The simulation studies demonstrate that they have an edge over other existing methods across a comprehensive simulation settings.

主讲人简介：
在国际统计杂志发表文章60篇左右，由 Springer 和 Chapman&Hall 分别出版专著各一本，曾任 JSPI 和 journal of nonparametric statistics Associate editor. 现任 annals of institute of statistical mathematics Associate editor.曾任一届泛华统计协会 board of directors 成员。