Graph Transformer Boosts Edge Computing Traffic Forecasting Accuracy

Accurate traffic forecasting is critical for managing resources proactively in edge computing environments, where service demand fluctuates spatially and temporally. Cellular edge systems exhibit complex traffic patterns, including strong spatial correlations among neighboring regions and long-range temporal dependencies influenced by user mobility and application behavior. Traditional recurrent forecasting methods often struggle with these long-horizon, non-stationary conditions. To overcome these limitations, a new spatiotemporal graph Transformer framework has been developed. This framework uniquely integrates graph neural networks to model the spatial interactions between service regions and leverages Transformer-based self-attention mechanisms to capture long-range temporal patterns from historical traffic data. By decoupling spatial representation learning from temporal reasoning, the approach offers an effective solution for large-scale spatiotemporal traffic modeling. Extensive experiments conducted on a real-world cellular network dataset demonstrate that this graph Transformer consistently surpasses recurrent graph-based baselines, such as GCN-RNN and GCN-LSTM, across various forecasting horizons. The improved forecasts enable more effective proactive resource provisioning, significantly reducing the risk of overload compared to reactive management strategies, highlighting the potential of graph-enhanced attention for adaptive edge computing.