Model-Free Feahture Screening With Exposure Variable

As the Age of Big Data is coming,more and more ultrahigh-dimensional data has been applied in various fields of scientific studies.Feature screening for ultrahigh-dimensional data becomes a popular project for statisticians.Based on the previous studies,in this paper,we propose a new model-free feature screening method,which is applicable when the influence of per.The proposed procedure takes the condition-al correlation between the predictors and the indicative function of the response as a marginal score function to measure the importance of predictors.We use the Kernel smoothing method to estimate the marginal utility and get the criteria in sample ver-sion.After ranking the criteria of all the predictors,we choose the largest part into the sub-model.This paper also introduces and proves the sure independence property and the ranking consistency property of the proposed method,which imply that all the truly important predictors are ranked in the top and selected into the sub-model.Taking four kinds of model frameworks as examples,this paper conducts Monte Carlo simulations to examine the properties of this procedure,and compares its performance with another three feature screening methods.At last,we conduct an empirical analysis using the transcription profiling of human breast cancer samples,and select the most correlative genes with lymph node metastasis.We finally draw a conclusion that the new screening procedure without specific model framework can be used in ultrahigh-dimensional data with exposure variable,which has not been achieved by the existed feature screening methods.