The paper introduces the concept of sure screening and proposes a method called Sure Independence Screening (SIS) to reduce the dimensionality of high-dimensional data from a large scale to a moderate scale below the sample size. SIS is based on correlation learning, which ranks features by their marginal correlations with the response variable and filters out those with weak correlations. The authors show that SIS has the sure screening property, meaning that all important variables are likely to survive after the screening process. This property is crucial for improving the efficiency and accuracy of variable selection in high-dimensional settings. The paper also discusses the iterative SIS (ISIS) and its extensions, and reviews several well-developed model selection techniques such as SCAD, Dantzig selector, Lasso, and adaptive Lasso. Numerical studies and a real data example from leukemia classification demonstrate the effectiveness of SIS-based methods in reducing computational burden and improving estimation accuracy. The authors conclude that SIS makes ultra-high-dimensional variable selection feasible and efficient, and can be combined with other model selection techniques to achieve better results.The paper introduces the concept of sure screening and proposes a method called Sure Independence Screening (SIS) to reduce the dimensionality of high-dimensional data from a large scale to a moderate scale below the sample size. SIS is based on correlation learning, which ranks features by their marginal correlations with the response variable and filters out those with weak correlations. The authors show that SIS has the sure screening property, meaning that all important variables are likely to survive after the screening process. This property is crucial for improving the efficiency and accuracy of variable selection in high-dimensional settings. The paper also discusses the iterative SIS (ISIS) and its extensions, and reviews several well-developed model selection techniques such as SCAD, Dantzig selector, Lasso, and adaptive Lasso. Numerical studies and a real data example from leukemia classification demonstrate the effectiveness of SIS-based methods in reducing computational burden and improving estimation accuracy. The authors conclude that SIS makes ultra-high-dimensional variable selection feasible and efficient, and can be combined with other model selection techniques to achieve better results.