Abstarct: Analysis of high-dimensional data in all fields of science, industry and commerce is always faced with the challenge of choosing important variables. Common methods for selecting variables are baxsed on linear structures, but in many cases and in the above, the structural relationship between the response variable and the explanatory variables may not be. Also, these methods ignore variables that alone have little effect on the variable, but in combination with other variables for prediction. In this research, by providing a multi-stage algorithm baxsed on linear local regression, important variables are identified and selected in the form of subsets, and the prediction model is fitted.