Prediction of Network Traffic through Light-Weight Machine Learning

Abstract

The evolving nature of network traffic challenges existing learning-based models, as frequently re-training the hyper-parameters is required to adaptively learn and predict its behavior. Benefiting from discovering the sparsity and self-similarity of network traffic, it becomes possible to develop compact algorithms with high accuracy and low computational complexity. However, how to properly make use of these properties still remains to be explored. In this article, we establish a light-weight learning framework for network traffic prediction based on sparse representation, and try to take the full advantage of these properties to enhance the capability of tracking its highly evolving characteristics. Specifically, 1). The strict causality constraint makes it difficult to equip the conventional sparse representation with predictive capability. To solve this issue, we make use of the self-similarity and train the representative/predictive dictionaries in a joint manner, such that the query point is embedded in terms of a sparse combination of dictionary atoms, and jointly coded with its T + 1 time slot behind counterpart, which proves to be optimal in the concatenated representative/predictive feature space. Then, the query point can be estimated through iterative projection method. 2). The consideration of the sparsity constraint loosens the upper bound of the time averaged prediction error. To address this problem, we slightly modify the sparse representation-based prediction by adopting Lyapunov optimization, and try to minimize the time averaged prediction error. Finally, the simulation results verify the performance improvements.