Hierarchical Bidirectional State-Space Graph Fusion Network for Long-Horizon Network Traffic Forecastings

HB-SSGFN is a unified deep learning framework for scalable long-horizon network traffic forecasting that combines four components—a Hierarchical Selective State-Space Encoder for multi-resolution temporal patterns, a Bidirectional Graph Propagation Block for both downstream congestion diffusion and upstream TCP feedback, a Bilinear Spatiotemporal Fusion layer using multiplicative cross-modal interactions, and a Stability-Aware Multi-step Decoder that constrains the spectral norm to curb long-horizon error amplification—achieving linear computational complexity while outperforming eight competitive baselines on three real-world backbone datasets, with MAE reductions of 17.3% over Graph WaveNet and 13.6% over Mamba-SSM at a 12-step horizon (p<0.01).