M$^{2}$STL: Multi-Range Multi-Level Spatial-Temporal Learning Model for Network Traffic Prediction

Network traffic prediction is crucial for network management, resource allocation, congestion control, and network anomaly detection. However, existing studies on network traffic prediction often fail to fully utilize the complex spatial-temporal dependencies present in network traffic data. This paper introduces a deep learning-based network traffic prediction model, named Multi-range Multi-level Spatial-Temporal Learning (M
$^{2}$
STL), specifically designed to learn the complex spatial-temporal characteristics inherent in a dynamic network environment. Within M
$^{2}$
STL, we devised three Spatial-Temporal Aggregation (STA) modules, each designed to learn from three different ranges of spatial-temporal features: recent, daily, and weekly. In particular, we propose three primary techniques within the STA modules to extract these spatial-temporal features: 1) A Dilated Conv Net based on 1D-dilated convolutions is proposed to learn long correlations in the temporal dimension with a small number of parameters. 2) A Mix-hop GCN Net, based on graph convolutions, is proposed to effectively extract spatial features. 3) Instead of providing the Mix-hop GCN Net with a pre-fixed adjacency matrix, a Graph Learning Net is proposed to learn multiple adjacency matrices, progressing from the shallower level to the deeper level. Extensive experiments have been conducted on two real-world network traffic datasets, Abilene and GÉANT. The results demonstrate that our method outperforms the other seven baselines across different prediction intervals. The Normalized Mean Absolute Error (NMAE) is 28.4% lower than the best baseline algorithm when the prediction interval is 36 on the GÉANT dataset.