Bayesian Tensor Completion for Network Traffic Data Prediction

Zecan Yang, Laurence T. Yang, Huaimin Wang, Bocheng Ren, Xiangli Yang

Abstract: Network traffic data prediction plays a significant role in various network applications and is a fundamental task in network traffic engineering. However, with the development of the Internet of Things and mobile computing, network traffic data presents the characteristics of high volume, variety, and velocity, which brings unprecedented challenges to network traffic data prediction. In this article, a Bayesian tensor completion model is proposed to predict network traffic data. More specifically, we represent network traffic data as a third-order tensor, which better preserves the underlying relationships inside network traffic data. Furthermore, based on tensor factorization and Bayesian learning, reasonable priors are set for the network traffic tensor and factor matrices. Finally, two approximate estimation algorithms, Gibbs sampling and variational Bayesian inference (VB) are developed to estimate the posterior of factor matrices, which are the latent information of the predicted network traffic data. In addition, we also analyze the difference between Gibbs sampling and VB, and discuss that these methods are more suitable for those scenarios of network traffic data prediction respectively. Experiments on two real network traffic data shows that the performance of our model is better than the traditional tensor completion models.