Capturing Traffic Propagation and Topological Interactions in Cloud Networks via Graph Neural Networks

Abstract

Network traffic in cloud computing environments is jointly influenced by multi-tenant concurrency, service dependency relationships, and dynamic scheduling policies. It therefore exhibits strong volatility, temporal non-stationarity, and structural coupling, which pose significant challenges for accurate prediction. To address the limitation of traditional time series methods in capturing inter-node correlations and traffic propagation mechanisms, this study proposes a graph neural network-based approach for cloud traffic fluctuation prediction from a structural modeling perspective. The cloud network is abstracted as a time-evolving graph, where node-level traffic observations are treated as graph signals. Spatial dependencies among communication entities are modeled through graph structures, while temporal dynamics of traffic evolution are learned via sequence encoding mechanisms. This enables a unified representation of spatiotemporal features within a single framework. Under consistent data settings and evaluation protocols, the proposed method is systematically compared with several representative forecasting models. The results demonstrate clear advantages in error control and prediction stability. The method is more effective in handling traffic fluctuations caused by topological interactions and workload migration in complex cloud scenarios. These findings indicate that structure-aware spatiotemporal modeling can substantially improve the reliability of cloud traffic prediction and provide a more robust foundation for network resource management and operational state analysis.