Research And Application Of The Time-varying Coefficient Model Based On Lasso Penalty

The advent of the era of big data makes multivariate time series an important research object in many fields.How to extract the effective information of time series is one of the problems that need to be solved in many fields.When analyzing time series,since the traditional time series analysis method ignores the time-varying relationship between variables,it is a new direction of current research to introduce a dynamic model that can consider the time-varying characteristics of data.In this paper,based on the idea of time-varying learning model coefficients,and using the Lasso regularization method which can estimate the model coefficients while implementing variable selection,the following researches are carried out on time-varying regression model and time-varying graph model respectively.Firstly,for the problem that the traditional regression model requires the sample data to be stable and independent,a time-varying streaming Lasso regression model is introduced.The model uses the adaptive filtering principle to achieve the purpose of dynamically adjusting the model coefficients by adding the forgetting factor,and realizes the linear regression model to analyze the correlated and non-stationary sample data,and improve the training mode of the traditional model.The model was used to analyze the diabetes dataset and the Beijing-Tianjin-Hebei air quality index.The experimental results show the superiority of the streaming Lasso regression model in terms of variable selection and prediction accuracy.Secondly,considering the time-varying characteristics of non-stationary multivariate time series,according to the idea of discarding data by forgetting index,the Lasso penalty vector autoregressive model is improved,and the online adaptive Lasso penalty vector autoregressive model with time-varying coefficients is introduced.The method solves the model parameters.The model is applied to the prediction of wind power and the selection of characteristics of EEG signals.The experimental results show that it is superior to the traditional model in interpretability and predictability.Finally,in order to study the dynamic correlation of variables in multiple time series,a time-varying Lasso model is introduced.Based on the theory of graph theory and the principle of Lasso penalty,the model solves the time-varying sparse covariance matrix by alternating direction multiplier method and reveals the dynamic correlation between variables.The model is used for the analysis of simulation data and stock data.The experimental results show that the accuracy and scalability are better than the static graph Lasso model.