Best Predictive Generalized Linear Mixed Model with Predictive Lasso for High-Speed Network Data Analysis

Abstract

Optimizing network usage is important to maximize the network performance. When the network usage grows rapidly, it is important to build an accurate predictive model. We present a new predictive algorithm which can analyze the network performance in various network conditions and traffic patterns. Our approach is based on the best predictive generalized linear mixed model (GLMM). The parameters of the best predictive GLMM are estimated by minimizing the mean squared prediction error (MSPE). To expedite the parameter learning with the big data collected through the network, our algorithm introduced regularization, LASSO, and an innovative bootstrap. The merits of our new approach validated through data and simulation are that (1) the highest prediction accuracy even under a model misspecification; and (2) the least computation time compared to the Estimationoriented GLMM with Lasso and Stepwise Selection GLMM. A major computational advantage of our method is that, unlike some of the current approaches, our method does not require the EM (Expectation-Maximization algorithm) procedure.

Joint Selection in Mixed Models using Regularized PQL

“…Indeed, using regularized PQL for fast model selection only mirrors other works in the GLM context, where penalized likelihood approaches have been proposed purely as a means of computationally efficient model selection (e.g., Schelldorfer et al, 2014;Hui et al, 2015). Of course, we acknowledge further simulations are required to fully assess the robustness of regularized PQL selection e.g., how it performs when the truly non-zero coefficients and hence signal to noise ratio is small, and that there are other methods of joint selection in GLMMs which were not included in our study e.g., the predictive shrinkage selection method of Hu et al (2015) designed specifically for Poisson mixture models in the context of network analysis.…”