[Paper] Adaptive Graph Pruning with Sudden-Events Evaluation for Traffic Prediction using Online Semi-Decentralized ST-GNNs

Source: arXiv - 2512.17352v1

Overview

The paper tackles a real‑world bottleneck in deploying Spatio‑Temporal Graph Neural Networks (ST‑GNNs) for traffic forecasting across edge‑distributed compute nodes (cloudlets). By introducing an adaptive graph‑pruning technique and a new Sudden‑Event Prediction Accuracy (SEPA) metric, the authors show how to slash inter‑node communication while still catching rapid traffic changes—something that traditional error metrics often overlook.

Key Contributions

• Adaptive Graph Pruning: A dynamic algorithm that trims redundant neighbor features per cloudlet based on recent prediction performance, preserving only the most informative spatial context.
• SEPA Metric: A novel evaluation measure that explicitly rewards correct detection of traffic slow‑downs and recoveries, addressing the blind spots of RMSE/MAE on sudden events.
• Online Semi‑Decentralized Framework: Integration of the pruning method into three distributed learning paradigms—classic Federated Learning (FL), server‑free FL, and Gossip Learning—demonstrating broad applicability.
• Extensive Empirical Validation: Experiments on two large‑scale traffic datasets (PeMS‑BAY, PeMS‑D7‑M) across short (5 min), mid (15 min), and long (30 min) horizons, confirming that communication overhead can be cut by up to 45 % without degrading SEPA or standard accuracy.
• Insightful Analysis of Spatial Connectivity: Evidence that preserving selective edges, rather than dense connectivity, is sufficient for capturing dynamic traffic patterns.

Methodology

1. Problem Setting: Sensors across a city form a graph where nodes are traffic stations and edges encode spatial proximity. Each cloudlet processes a sub‑graph but must exchange overlapping node features with neighboring cloudlets to maintain global context.
2. Adaptive Pruning Loop:
   • After each prediction round, the cloudlet computes a short‑term performance score (e.g., recent MAE).
   • If performance degrades, the pruning rate is reduced (i.e., more neighbor features are kept); if performance improves, the rate is increased (i.e., more edges are dropped).
   • Pruning decisions are made per‑edge using a lightweight importance estimator (e.g., gradient‑based saliency or historical variance).
3. SEPA Calculation: For each prediction horizon, the metric checks whether the model correctly flags a traffic slowdown/recovery within a predefined time window. It combines true‑positive rate on events with a penalty for missed or delayed detections.
4. Training Regimes: The authors plug the pruning logic into three semi‑decentralized learning loops:
   • Traditional FL: Periodic aggregation on a central server.
   • Server‑Free FL: Peer‑to‑peer model averaging without a coordinator.
   • Gossip Learning: Randomized neighbor exchanges, naturally aligning with the pruning decisions.
5. Evaluation Pipeline: Both standard metrics (MAE, RMSE) and SEPA are logged while measuring the volume of transmitted feature tensors per epoch.

Results & Findings

• SEPA shines: While MAE/RMSE differences were within statistical noise, SEPA revealed that the pruned models caught 15‑20 % more sudden congestion events than the unpruned baseline.
• Latency impact: The adaptive scheme kept inference latency stable (≈ 10 ms per cloudlet) because the smaller feature set also speeds up the GNN forward pass.
• Robustness across horizons: Gains persisted for 5‑min, 15‑min, and 30‑min forecasts, indicating that the method adapts to both immediate and longer‑term traffic dynamics.

Practical Implications

• Edge‑Centric Traffic Services: City‑scale traffic prediction platforms can deploy lighter‑weight ST‑GNN agents on edge nodes (e.g., 5G MEC servers) without flooding the backhaul network.
• Cost‑Effective Scaling: Reducing inter‑node traffic translates directly into lower bandwidth bills and less congestion on the control plane, especially valuable for municipalities with limited fiber connectivity.
• Improved Incident Response: Higher SEPA scores mean that traffic management systems (e.g., dynamic routing, variable speed limits) receive timely alerts about emerging jams, enabling faster mitigation.
• Plug‑and‑Play with Existing FL Toolkits: Since the pruning logic is a pre‑processing step on feature tensors, it can be integrated into popular FL frameworks (TensorFlow Federated, PySyft) with minimal code changes.
• Generalizable to Other Sensor Networks: Any domain where spatially distributed time‑series data are processed (smart grids, environmental monitoring) can benefit from adaptive graph pruning to curb communication overhead.

Limitations & Future Work

• Pruning Overhead: The importance‑estimation step adds a small compute cost; on ultra‑low‑power devices this could offset communication savings.
• Event Definition Sensitivity: SEPA relies on domain‑specific thresholds for “sudden” events; tuning these for different cities or sensor densities may require expert input.
• Static Topology Assumption: The current method assumes a relatively stable graph structure; handling dynamic addition/removal of sensors (e.g., temporary road closures) remains an open challenge.
• Broader Benchmarks: Experiments were limited to two California traffic datasets; validating on heterogeneous networks (e.g., European cities with different road layouts) would strengthen generalizability.

Future research directions include joint optimization of pruning and model architecture, adaptive thresholding for SEPA, and extending the approach to multimodal data (e.g., combining traffic flow with weather or incident reports).

Authors

• Ivan Kralj
• Lodovico Giaretta
• Gordan Ježić
• Ivana Podnar Žarko
• Šarūnas Girdzijauskas

Paper Information

• arXiv ID: 2512.17352v1
• Categories: cs.LG, cs.AI, cs.DC
• Published: December 19, 2025
• PDF: Download PDF