Answers are grounded in this paper and cite their sources.
Driven by the rapid evolution of digital infrastructure, the exponential growth of network traffic poses severe load-balancing challenges to electronic flow-correlation systems characterized by strong spatio-temporal coupling and flash-crowd dynamics. Traditional “sense-then-react” algorithms based on real-time connection counts or static weight assignments struggle with tidal traffic bursts induced by 5G, IoT, and cloud-native technologies, resulting in traffic hotspots, underutilized resources, and cascading connection avalanches. To address these challenges, this study proposes a prediction-driven, self-adaptive load-balancing framework based on a spatio-temporal graph neural network (ST-GNN). Network traffic is modeled as a dynamic spatio-temporal graph and processed through a hybrid pipeline integrating graph convolution, Bi-LSTM, and spatio-temporal attention. The model processes 168-step historical matrices of bytes, packets, and flows to predict the next six time steps, and converts predicted resource utilization into differentiable weights for an adaptive weighted-least-connections (AWLC) policy. Experiments on the Abilene and GEANT datasets show that ST-FNet achieves an MSE of 1.31 and R 2 of 0.92, improving MSE by 42.3 % and R 2 by 3.4 % over LSTM while reducing inference latency by 12.1 %. Across Fat-tree, Leaf-Spine, and 3-tier topologies under CPU, memory, link-capacity, and RTT perturbations and traffic loads of 0–20 Mbit/s, AWLC maintains a traffic loss rate of ≤ 1.35 % ≤ 1.35% increases throughput by 10–14 %, reduces idle resources by 41 %, and lowers node-overload probability by 52 %. These results demonstrate the effectiveness of prediction-driven scheduling for adaptive load balancing under dynamic network conditions.
No takes yet. Share an insight, caveat, or question.
Ben Yang (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: