讲座题目：Concordance Measure-based Feature Screening and Variable  Selection

主讲人：西南财经大学 林华珍教授  

讲座时间：2017年12月25日（星期一）16:00-16:50

讲座地点：综合楼601会议室  

主讲人简介：林华珍，西南财经大学统计学院教授、博导，统计研究中心主任，美国华盛顿大学生物统计系博士后，四川大学博士。教育部长江学者特聘教授，国家杰出青年科学基金获得者，入选国家百千万人才工程，教育部新世纪优秀人才，第十一批四川省学术和技术带头人。 论文发表在JASA、Annals of  Statistics、JRSSB、Biometrika、Journal of  Econometrics及Biometrcs等国际统计学及计量经济学顶级期刊上，并先后担任国际统计学期刊《Biometrics》、《Scandinavian  Journal of Statistics》、《Statistics and Its Interface》、《Statistical Theory and  Related Fields》Associate Editor， 国内核心学术期刊《应用概率统计》、《系统科学与数学》、《数理统计与管理》编委。  研究领域：非参数理论和方法、转换模型、生存数据分析、函数型数据分析、时空数据分析。

摘要：The  $C$-statistic, measuring the rank concordance between predictors and outcomes,  has become a standard metric of predictive accuracy and  is therefore  a  natural criterion for  variable screening and selection. However, as  the $C$-statistic is a step function, its optimization requires brute-force  search, prohibiting its direct usage in the presence of high-dimensional  predictors. We develop a smoothed version of the $C$-statistic to facilitate  variable screening and selection. Specifically, we propose a smoothed  $C$-statistic sure screening (C-SS) method for screening ultrahigh-dimensional  data, and a penalized  $C$-statistic (PSC)  variable selection method for  regularized modeling  based on the screening results. We have shown that these  two coherent procedures form an integrated framework for screening and  variable selection: the C-SS possesses the sure screening property, and the PSC  possesses the oracle property. Specifically, the PSC achieves the oracle  property if $m_n = o(n^{1/4})$, where $m_n$ is the cardinality of the set of  predictors captured by the C-SS. Our extensive simulations reveal that, compared  to existing procedures, our proposal is more robust and efficient. Our procedure  has been applied to analyze a multiple myeloma study, and has identified several  novel genes that can predict patients response to treatment.