Improving the Traffic Prediction Capability of Neural Networks Using Sliding Window and Multi-task Learning Mechanisms

Abstract

Due to the diversity of network services and the unpredictability of their behaviors, there is an increasing need for tools that can aid in the global management of IP networks. Being able to predict network data can be very useful to anticipate network upgrading decisions or changes on the network functional operation. This paper proposes a practical approach, based on neural networks, that is able to predict network traffic in a specific network link. In order to improve the prediction capabilities of the different neural network models, sliding window and multi-task learning mechanisms are introduced and tested. By applying this prediction framework to different network links, it will be possible to predict the evolution of the global network traffic and use this information for network security, management and planning purposes. The results obtained by applying the proposed model to realistic network scenarios show that this concept can achieve excellent performance in the prediction of the network traffic on the selected links. The prediction is accurate even when there are significant changes in the number of users and their respective profiles. Moreover, the proposed prediction approach is generic and can be used to predict different network data with a very satisfactory accuracy, even with simple and small NN models.