基于分位数方法的超高维删失数据的特征筛选

Quantile Feature Screening for Ultra High Dimensional Censored Data

利用条件分位数分别在响应变量非随机删失和随机删失的情形下研究超高维数据的特征筛选,提出相应的特征筛选方法,理论证明和模拟验证都证明该方法筛选出的变量集满足确定筛选性质和排序一致性。与已有方法相比,本文方法在协变量与删失变量相关时具有相对优势。

In this paper,conditional quantiles are used to study the feature screening problem of ultra-high dimensional data when response variables are randomly censored and nonrandomly censored,respeclively.Then,a corresponding feature screening method is proposed.Through theoretical and simulation study,it is verified that the variable set created by this method satisfies the sure screening and ranking consistency properties.Compared with the existing methods,this method has some advantages when covariates and censored variables are correlative.