Research On Feature Selection Based On Sparse Group Lasso Related Penalty Items

With the explosive growth of data in recent years and the continuous rise of the concept of big data,data dimension reduction technology has drawn more and more attention.Feature selection can effectively reduce the dimensionality of the data.There arc also a variety of methods to select features.Many scholars use the method of regression to carry out the selection of features,one of the most widely used is lasso.At present,scholars at home and abroad have made a lot of research achievements in this field,and some methods including lasso,group lasso,fused lasso and graph lasso have been developed.In this paper,based on the shortcomings of sparse group lasso algorithm in practical application,this paper proposes three improvements in algorithm:Firstly,the loss function is improved,and the correlation between group and group is added to the loss function to reduce collinearity between non-zero fea-tures in order to achieve further sparse coefficient.Secondly,it adds data pre-processing steps and enriches the data set used to fit the model.It is hoped that not only the influence of each group's featues on the result can be obtained,but also the common influence they have on each other.Thirdly,the algorithm is ex-tended to other linear models,and the improved loss function is given by taking logistic regression as an example.Finally,we use simulated data to validate the algorithm and apply the algorithm to a movie-scoring dataset.