Premium
Spatiotemporal graph convolutional recurrent networks for traffic matrix prediction
Author(s) -
Zhao Jianlong,
Qu Hua,
Zhao Jihong,
Dai Huijun,
Jiang Dingchao
Publication year - 2020
Publication title -
transactions on emerging telecommunications technologies
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.366
H-Index - 47
ISSN - 2161-3915
DOI - 10.1002/ett.4056
Subject(s) - computer science , provisioning , traffic generation model , network traffic simulation , robustness (evolution) , data mining , traffic classification , backbone network , graph , traffic engineering , network traffic control , distributed computing , quality of service , computer network , theoretical computer science , biochemistry , chemistry , network packet , gene
Summary Through accurate network‐wide traffic prediction, network operators can agilely manage resources and improve robustness by proactively adapting to new traffic patterns, especially for traffic engineering, capacity planning and quality of service provisioning. However, due to the proliferation of backbone network traffic as well as the complexity and dynamics of network communication behavior, accurate and effective network‐wide traffic prediction is challenging. To address the challenges, this paper focuses on short‐term traffic matrix (TM) prediction in large‐scale IP backbone networks. In order to improve the prediction performance, a novel spatiotemporal graph convolutional recurrent network (SGCRN)—a deep learning framework that incorporates both spatial and temporal dependencies of traffic flows, is proposed to implement TM prediction with high accuracy and efficiency. By learning network‐wide traffic as graph‐structured TM time series, SGCRN jointly utilizes graph convolutional networks (GCN) and gated recurrent units (GRU) networks to extract comprehensive spatiotemporal correlations among traffic flows. Specifically, SGCRN employs GCNs to identify structural spatial features of traffic flows by considering topological properties, and utilizes GRUs to implement temporal features learning by considering short and long‐term dynamics of traffic flows. Extensive experimental results on the inter‐Points of Presence network traffic data from four real IP backbone networks show that SGCRN can effectively predict short‐term network‐wide TM with superior accuracy compared with other four widely used traffic prediction methods.