A scalable federated learning-based approach for accurate traffic prediction in edge computing-enable metro optical networks

Abstract

Accurate network traffic prediction is indispensable for efficient load-aware resource management and performance optimization in metro optical networks (MONs). Existing machine learning (ML)-based methods for network node traffic prediction typically adopt a centralized approach, in which a centralized learning model is trained with traffic data collected from all nodes. However, the requirement for transferring massive amount of data to a centralized server in such an approach may increase the communication delays and raise privacy concerns. Although these problems can be solved by using federated learning (FL) technique, the generalization ability and prediction accuracy of a global model in a federated setting are still affected by the diverse traffic patterns and scale of the MON, which are also the limitations of traditional centralized approaches for traffic prediction. To further improve the traffic prediction accuracy in edge computing-enabled MONs, we propose an FL-based traffic prediction framework, namely FedMON, that consists of a clustering strategy based on traffic patterns, and a two-stage model aggregation scheme, in which a scalable global model is collaboratively trained by edge nodes. By exploiting the traffic patterns-based clustering, obtained by applying seasonal and trend decomposition using locally estimated scatterplot smoothing (Loess) followed by the K-means clustering algorithm, the proposed FedMON framework performs intra- and inter-cluster model aggregation in a hierarchical manner to deal with the huge amount and heterogeneity of traffic data issues experienced in real MONs. We conducted detailed experiments using synthetic traffic data to evaluate the performance of FedMON for MONs of varying scales. The results demonstrate that FedMON outperforms baseline FL methods as well as a state-of-the-art centralized learning model. More importantly, the strong scalability of FedMON reduces the traffic prediction errors by up to 14% for large-scale MONs.