NetSight: Graph Attention Based Traffic Forecasting in Computer Networks

2025-05-11

AI Summary

Addressing the challenge of modeling nonlinear and dynamic spatiotemporal dependencies in network traffic forecasting, existing graph neural network (GNN) approaches typically decouple spatial and temporal modeling and struggle to jointly capture local interactions and global topological structure. This paper proposes the first end-to-end graph attention framework for joint spatiotemporal dependency modeling. It introduces a dynamic spatiotemporal adjacency matrix to explicitly encode time-varying spatial relationships among nodes, and a multi-scale node normalization mechanism to jointly learn local neighborhood interactions and global graph-structural awareness. Evaluated on two large-scale real-world network traffic datasets, the method consistently outperforms all state-of-the-art models, achieving an average 21.3% reduction in prediction error. The source code and benchmark datasets are publicly released.

Technology Category

• Planning, Routing, and Scheduling: Optimization of Spatio-temporal Systems
• Machine Learning: Graph-based Machine Learning
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Social Networks and Social Media: Influence propagation, information diffusion, and the prediction on networks
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

The traffic in today's networks is increasingly influenced by the interactions among network nodes as well as by the temporal fluctuations in the demands of the nodes. Traditional statistical prediction methods are becoming obsolete due to their inability to address the non-linear and dynamic spatio-temporal dependencies present in today's network traffic. The most promising direction of research today is graph neural networks (GNNs) based prediction approaches that are naturally suited to handle graph-structured data. Unfortunately, the state-of-the-art GNN approaches separate the modeling of spatial and temporal information, resulting in the loss of important information about joint dependencies. These GNN based approaches further do not model information at both local and global scales simultaneously, leaving significant room for improvement. To address these challenges, we propose NetSight. NetSight learns joint spatio-temporal dependencies simultaneously at both global and local scales from the time-series of measurements of any given network metric collected at various nodes in a network. Using the learned information, NetSight can then accurately predict the future values of the given network metric at those nodes in the network. We propose several new concepts and techniques in the design of NetSight, such as spatio-temporal adjacency matrix and node normalization. Through extensive evaluations and comparison with prior approaches using data from two large real-world networks, we show that NetSight significantly outperforms all prior state-of-the-art approaches. We will release the source code and data used in the evaluation of NetSight on the acceptance of this paper.

Problem

• Modeling joint spatio-temporal dependencies in network traffic
• Handling graph-structured data with local and global scales
• Improving GNN-based traffic forecasting accuracy

Innovation

• Graph attention based joint spatio-temporal dependency learning
• Simultaneous modeling of local and global network scales
• Novel spatio-temporal adjacency matrix technique