Research On Feature Selection Methods Of Two Types Of Ultra-high Dimension Right Censored Data

Ultrahigh dimensional right censored data are increasingly found in various fields of modern scientific research,such as biomedicine,economics and so on.Feature screening methods are essential tools for analyzing such data.This paper studies two types of feature screening methods for ultrahigh dimensional right censored data.The second chapter studies a kind of conditional feature screening method.It is well known that most of the existing feature screening methods are marginal methods,either based on marginal correlation coefficient.or based on marginal regression coefficient esti-mation.The marginal screening method has some important disadvantages.For example,they can not select covariates that have an effect on response variables in conjunction with other covariates but have no marginal effect on response variables;it is easy to misselect covariates that have a higher marginal correlation with response variables but actually have no effect on response variables.In addition,marginal screening methods are also affected by confounding variables,because confounding variables can lead to the problem of false correlation.In fact,in many practical problems,researchers know beforehand that a cer-tain(some)covariate has an important effect on the response variable.Prior information should be considered in the process of variable screening.Based on this consideration,a new conditional feature screening method for ultrahigh dimensional right censored data is proposed in the second chapter.This method is based on the technique of conditional typical correlation coefficient and inverse probability weighting.The new method is robust to both response and covariates and can be used to deal with cases in which they obey the heavy tail distribution.This chapter also establishes the theoretical nature of the new approach.This chapter also evaluates the performance of the new method under finite samples by numerical simulations.Finally,this chapter applies the new method to the actual data analysis.The third chapter studies a class of robust feature screening methods.Robustness has been an important issue in statistical research.Robust statistical methods have better adaptability and wider applicability.This chapter extends the Copula correlation(CC)under complete data to ultrahigh dimensional right censored data quality redistribution techniques.Since the test function in quantile is used in the construction,the new method has good robustness.Moreover,the method of compound quantile is used to further improve the efficiency.This chapter also evaluates the performance of the new method under the limited sample by numerical simulation,and applies the new method to the actual data analysis.The fourth chapter summarizes and discusses paper.