A new approach for traffic matrix estimation in high load computer networks based on graph embedding and convolutional neural network

Abstract

In computer networks, transmitted traffic between origin‐destination nodes has been considered a crucial factor in traffic engineering applications. To date, measuring the traffic directly in high load networks is difficult due to high computational costs. On the other hand, accurate estimation of network traffic by means of link load measurements and routing information is currently a challenging problem. In this paper, we propose a new approach to estimate end‐to‐end traffic, inspired by graph embedding. In the proposed approach, we consider a computer network as a time‐varying graph. Our model provides explicit routing information to a convolutional neural network estimator via link load measurements and network topological structure. When explicit routing information is provided, the learning model is only expected to embed the relations between link loads into a traffic estimation vector, instead of figuring out the routing paths. The experimental results showed that the proposed approach outperforms similar estimators in terms of lower estimation error and better tracking the fluctuations.

Estimation of Traffic Matrix from Links Load using Genetic Algorithm

“…The Recurrent Neural Network takes real world data for training of the model to extract spatio-temporal features of TM. Convolutional neural network is another finding for spatial relationship between link loads and Origin-Destination flows [29]. These artificial neural networks approach are good in finding hidden features of TM, although it requires large training data.…”

Design and Implementation of a Robust Convolutional Neural Network‐Based Traffic Matrix Estimator for Cloud Networks

“…Jiang et al [11] fuse the neural network approach with time frequency analysis for network traffic matrix estimation. Emami et al [13] have devised a convolutional neural network-(CNN-) based traffic matrix estimation architecture in which one part of the training data (link loads) is transformed into one higher dimension by considering the network topology. This enables exploitation of convolutional neural network-based techniques which are optimized for image (2D) datasets and are considered more robust than other types of neural networks.…”

“…1 st generation research works using the assumption of stationary routing matrix A have investigated range of spatiotemporal methods as explained in the previous section. More recent works, e.g., [13] have devised strategies in which the topology information is embedded into the model using flexibility, e. g., by using the link load measurements and embedding them into link adjacency matrix. This relaxes the spatial constraints in the original problem while still enabling accurate estimation through an efficient learning process of a neural network.…”

“…Recent work by Emami et al [13] incorporating graph embedding and convolutional neural network proposed an idea to fuse the topology of network into the neural network-based estimation framework by using the graph embedding approach. This enables to add one dimension to the training data.…”

“…A consequence of the violation of any of the above assumptions is that if we employ CNN using standard architectures, they may yield suboptimal performance, e.g., if a small volume of dataset is available for training or if we use an adaptive learning-rate optimizer such as Adagrad as used in [13], and it will yield inferior performance because of the incorrect assumption of the availability of large clean dataset; it will adaptively lower the learning rate as the training would progress over time.…”