Learning Based Methods for Traffic Matrix Estimation From Link Measurements

Abstract

Network traffic matrix (TM) is a critical input for capacity planning, anomaly detection and many other network management related tasks. The TMs are often computed from link load measurements. The TM estimation problem is the determination of the TM from link load measurements. The relationship between the link loads and the TM that generated the link loads can be modeled as an under-determined linear system and has multiple feasible solutions. Therefore, prior knowledge of the traffic demand pattern has to be used in order to find a potentially feasible TM. In this paper, we consider the TM estimation problem with limited prior information. Unlike previous methods that require past measurements of complete TMs, which are hard to obtain or protected by regulations, our method works even if only the distribution of TMs is known. We develop an iterative projection based algorithm to solve this problem. If large number of past TMs can be measured, we propose a Generative Adversarial Network (GAN) based approach for solving the problem. We compare the strengths of the two approaches and evaluate their performance for several networks using varying amounts of past data.

Structured Nonnegative Matrix Factorization for Traffic Flow Estimation of Large Cloud Networks

“…It would be interesting to adapt our NMF based traffic matrix estimation model into a deep NMF model. Other directions of research include the use of other NMF models to perform the traffic flow estimation, such as [48,49], or to improve prediction using data coming from other sources than the traffic flows and then use multi-view techniques such as [21,50].…”

Generative Deep Learning Techniques for Traffic Matrix Estimation From Link Load Measurements

“…It involves constructing a solution to the underdetermined TME linear system by leveraging the range space of a properly trained generative model. However, contrary to our methods, the authors in [18] employ a Generative Adversarial Network (GAN) and not a probabilistic model allowing to capture uncertainty like VAE. In addition, they only utilize fully connected layers, they do not make use of any form of attention mechanism (neither convolutional "spatial" attention nor self-attention), and most importantly, they only consider the spatial dependencies among the OD flows of the traffic matrix, ignoring the temporal domain entirely.…”

“…Our work also adopts a data-driven approach based on deep neural networks (DNNs) but pertains to the unsupervised learning paradigm-particularly, generative deep learning. Taking this into consideration, the most relevant approach to our work has been presented in [18] because it employs the same problem formulation and enabler for TM estimation (i.e., linear inverse problem and generative deep learning). It involves constructing a solution to the underdetermined TME linear system by leveraging the range space of a properly trained generative model.…”

“…The window length in the displayed output shapes is set to two time steps. The source code of the examined generative models and the complete traffic matrix estimation method (including the approach presented in [18]) is published in https://gitlab.com/gkakkavas/gdl-tme, accompanied by detailed documentation.…”

“…• We extensively evaluate the performance of the proposed models over a comprehensive set of suitable metrics on two publicly available datasets of traffic matrices obtained from real backbone networks. Additionally, we compare our approach with a state-of-the-art method previously published in the literature [18].…”