In The Case Of Ultra-high Dimensional Data, The Variable Filtering Of The Model Can Be Added

The development of science and technology and the continuous improvement of computer computing power complement each other,so that nowadays ultra-high-dimensional data appear in many of our research fields.In natural science,human genes and other research problems,the huge amount of data and the corresponding increase in the difficulty of calculation have brought difficulties to researchers.In the case of hyperdimensional data,only a few covariables are correlated with response variables.The basis of ultra-high dimensional data analysis is to identify the important variables in the data,so that the ultra-high dimensional data can be modeled.After years of research,current strategy is to create a two stage feature selection process,the first phase using high calculation efficiency variable selection method will be able to keep all the important features will be under the premise of to reduce the number of dimensions to the proper size under the sample size,the second stage to adopt some effective method for dimension reduction after data are important variables selection.The main content of this paper is the variable screening method of additive model in the case of super-high dimensional data.A variable screening method based on martingale difference correlation coefficient of additive model is proposed to reduce the variable dimension to the appropriate number below the sample size,and realize the first stage of variable selection in the case of ultra-high-dimensional data.Under certain conditions,we prove the filter consistency of the method.The results of simulated data experiment and real data analysis show that the proposed method has a good screening effect.