Integrative Analysis Based On Sparse Group Lasso Penalty

The so-called big data often come from different sources and tend to have high dimen-sionality and sparsity.How to reasonably and effectively mine and analyze the correlation in-formation and difference between such datasets and simultaneously reducing dimension and denoising is a problem worth pondering and studying.Integrative analysis is different from the previous single-dataset analysis,it combines multiple independent data sets and analyzes them at the same time,which provides an effective way of pooling information from the original data.The integrative analysis based on penalty function combines the integrative analysis and the idea of variable selection by penalty function together,so as to single out the important characteristic variables and reduce dimensions.Integrative analysis based on penalty function is different from group punishment of a single dataset,the regression coefficients of each ex-planatory variable in all data sets are considered as a group.The integrative analysis based on bi-level variable selection method can not only screen out significant characteristic variables,but also identify which data sets have significant selected important variables,so as to study the correlation and difference between different data sets.This paper applies the bi-level variable selection penalty function namely Sparse Group Lasso to the integrative analysis with similar sparse structure,and adopts a penalty to promotes such sparse structure,then creates a corresponding algorithm namely block coordinate descent for solving the problem and also proposes several evaluation indexes and a parameter tuning method for the model.This model method successfully solved the integrative analysis modeling problem when the structure in the data sets is unknown in advance but there are certain priori information shows that it has similar sparse structure.In this paper,different simulation results are analyzed under some model evaluation criteria,that shows the feasibility of the model and its excellent performance which is proved to be as good as the existing methods.Besides,the practicability of the model is also proved in two real examples.In the process of simulation,the integrative analysis based on Sparse Group Lasso penalty is almost as good as the existing integrative analysis based on Composite MCP penalty,and each has its own advantages in different situations.However,in terms of time cost,the model in this paper has more advantages.In the case analysis,the modeling solution based on the model presented in this paper are interpretable,and the correlation and differences between data sets was analyzed accordingly.