Application Research Of Group Lasso Penalty Regression Model And Algorithm

Globally,the use of big data to promote economic development has led to a geometric increase in the dimensionality of data features.The characteristics of data in many fields have prompted feature selection to become one of the effective means to solve the dimensional disaster and improve the generalization ability of algorithms.At the same time,the characteristics of the multi-category data exist in the form of group structure.The method of feature selection based on traditional method of adding a relationship terms to group structure is becoming more and more mature.In this paper,the limitations of the group Lasso regression model,where all variables in a group are selected or discarded at the same time,The following studies were conducted respectively from the Overlapping Group Lasso and the Hierarchical Group Lasso.model parameters was solved by Group coordinate descent method.Firstly,considering the overlapping of features,the overlapping group Lasso logic regression model is introduced.According to the overlap coefficient decomposition mechanism,The overlapping variables in the group were decomposed into potential variables to form a new feature vector.Simulation experiment and disease feature gene selection experiment were conducted to verify the superiority of Lasso in the overlapping group over the control group.Secondly,a Hierarchical Vector AutoRegression Moving Average model is introduced to solve the problem of the feature sharing a lag selection in the Group Lasso Vector AutoRegressive Model.While realizing the sparsity pattern of the coefficients,accounting serial correlation in the errors.Hierarchial group lasso's the parameters estimation could be solved by the two-stage estimation and the proximal gradient descent method.Through the experiments of troposcatter communication transmission level prediction and EEG signal feature classification,it is proved that this model has certain advantages in prediction and feature selection,and features are more sparse.Finally,in order to solve the problem of quadratic growth of parameter space in vector autoregressive model with exogenous variables,a Hierarchical Vector AutoRegression With Exogenous Variable was designed.The model uses the proximal gradient descent method to solve the parameters.It can effectively describe the relationship between features,and learn the different lag selection of endogenous and exogenous variables,and reduce the parameter dimension.Through the experiment of atmospheric pollutant prediction and EEG signal feature selection data set,it is proved that the model can obtain lower normalized mean square error and more easily recognize multi-channel EEG signal features in the prediction of high-dimensional time series.