Network Traffic Modeling and Prediction Using Graph Gaussian Processes

Traffic modeling and prediction is a vital task for designing efficient resource allocation
strategies in telecommunication networks. This is challenging because network traffic data exhibits complex
nonlinear spatiotemporal interactions. Moreover, the data can have missing values when traffic statistic
collection is unavailable in certain nodes. In this paper, we introduce a graph Gaussian Process (GP) model
for this challenging problem. The GP is a Bayesian non-parametric model and highly flexible in capturing
complex patterns in the data. Additionally, it provides uncertainty information which can be exploited for
robust resource allocation problems. The developed graph GP model is almost free of hyper-parameter
tuning, can accurately capture short-term and long-term temporal patterns and can infer missing values
by learning spatiotemporal interactions among the nodes in the network. Subsequently, we approximate
the intractable posterior distribution using Variational Bayes (VB) algorithm which can be efficiently
implemented. Finally, we evaluate the accuracy of the proposed model for predicting the data traffic using
two real-world network datasets. Our simulation results shows that the proposed model can achieve better
prediction accuracy with respect to the state-of-the-art approaches