Long-term prediction of multiple types of time-varying network traffic using chunk-based ensemble learning

With the development of networking technologies, global Internet traffic is constantly increasing. Moreover, various traffic types associated with a variety of network services and applications co-exist in the network, having diverse volume and seasonality patterns. The knowledge of future patterns of these heterogeneous traffic types is beneficial for the network operators, as it enables proper adjustment of available network resources in near real-time, and allow performing periodical network reconfigurations. In this paper, we propose a chunk-based ensemble learning method for a long-term prediction of multiple network traffic types. Our developed prediction method combines two popular machine learning (ML) approaches: chunk-based learning on data streams and ensemble learning. The proposed method does not require large volumes of training data and is resilient to changes in traffic characteristics. Finally, we create a custom metric called allocation outside blocking threshold (AOBT), which links the problem of network traffic prediction to bandwidth blocking probability in dynamic traffic routing. We compare the performance of our online method to a number of baselines, using the root mean square percentage error (RMSPE) and the proposed AOBT metrics. According to conducted experiments, the proposed streaming approach outperforms the reference ones in RMSPE and AOBT. Depending on the metric and model, chunk-based ensemble learning achieves 5%-34% lower prediction errors across traffic types. However, the choice of a specific model depends on the investigated traffic type and applied metric.(c) 2022 Elsevier B.V. All rights reserved.