ST-GAT: a spatial–temporal graph-augmented transformer for cellular network traffic forecasting

Karimi Kashani, Alireza, Soleimani, Hossein
Telecommunication Systems · Springer Nature · 2026-10-07 · DOI: 10.1007/s11235-026-01516-x

Аннотация

Accurate cellular traffic forecasting requires modeling both local topology-driven interactions and long-range spatio-temporal dependencies. This paper proposes ST-GAT, a dual-branch framework that combines Chebyshev graph convolution with multi-head spatial self-attention and a Temporal Transformer. An Adaptive Spatio-Topological Gating Mechanism performs input-dependent fusion of the local graph representation and the global attention representation. On the Milan Internet-traffic test set, ST-GAT with adaptive gating achieves an MAE of 77.7318, an RMSE of 136.8921, and an $$R^{2}$$ R 2 of 0.9683, outperforming the considered baselines under the evaluated experimental setting. Additional experiments on SMS and call traffic show that the same architecture remains competitive across different cellular service types, although the magnitude and consistency of the improvement are service-dependent. A stratified analysis of public holidays, weekends, and non-holiday weekdays further shows that the reported prediction metrics vary across calendar conditions. These results support the applicability of the proposed architecture within the evaluated offline forecasting setting, while continual adaptation, cross-city transfer, and evaluation under broader temporal conditions remain directions for future work.

Читать у издателя →

Похожие статьи