EN
A graph-aware model forecasts traffic without retraining
GAS-PFN: Graph-Aware Spatiotemporal Prior-Data Fitted Networks for Zero-Shot Traffic Forecasting · arXiv:2610.37598
In plain words
Traffic changes over time, and connected roads can affect one another. The paper introduces GAS-PFN, a forecasting method that learns from made-up traffic patterns shaped by road maps, then uses recent and older readings from a target network without changing its model parameters.
On six traffic datasets, its average squared forecast error 90 minutes ahead was lower than TimePFN’s under the paper’s comparison rule. The tests do not show how it performs on a live road network.
The research
Can a traffic forecaster use the layout of a target road network without being trained specifically for that network? Existing graph-based forecasters typically need dataset-specific training, while time-series PFNs do not explicitly represent road topology.
This paper tests whether graph-shaped examples and target-network readings can help bridge that gap.
GAS-PFN combines road-shaped synthetic training examples with target-network readings to forecast traffic without updating its parameters on that network.
The researchers test GAS-PFN on six traffic datasets, comparing forecasts with measured speed or flow on chronological test splits.
At 90 minutes, its mean squared error averaged 34.0% below TimePFN’s. For each dataset, TimePFN uses the context option with lower test mean absolute error; GAS-PFN uses three independently pretrained checkpoints, while TimePFN uses one released checkpoint.
This comparison supports lower average squared error under the paper’s evaluation and context-selection procedures. It does not mean GAS-PFN wins on every dataset or under every context choice.
These results come from six chronological dataset test splits, not a live road network. The TimePFN comparison also uses a context-selection rule that can choose different context lengths for each dataset, so the average does not establish that GAS-PFN is better in every setting.
What the evidence shows
- Forecast across road networks
GAS-PFN, or Graph-Aware Spatiotemporal Prior-Data Fitted Network, is a zero-shot method for forecasting traffic on sensor networks represented as graphs. It uses a target network’s structure and readings to make predictions without updating its parameters on that dataset.
The research question is whether this kind of forecasting can account for road connections, which existing time-series PFNs do not explicitly represent, while avoiding dataset-specific training typical of graph-based forecasters.
- Represent connected roads
The method gives the model graph-linked examples during pretraining: it combines generated time patterns with sampled road subgraphs, varies signals by sensor, and lets some signal and noise travel across graph connections.
At prediction time, one pass uses readings from a single sensor, while another can use neighboring sensors. A learned gate combines their representations. This links the synthetic examples and the inference process to the same spatial structure.
- Compare forecast errors
The researchers assess forecasts on six traffic tests, which compare predicted speed or flow with measured values. At the cumulative 90-minute horizon, GAS-PFN’s mean squared error averaged 34.0% below TimePFN’s across the six datasets.
For each dataset, TimePFN is scored using whichever of two context options has lower test mean absolute error, then that option is used for all reported metrics and horizons. This supports an advantage under the study’s comparison procedure, not for every individual case.
- Check context and road coverage
A frozen-checkpoint ablation checks whether historical examples help at inference: removing them worsened mean absolute error or mean squared error in all 36 dataset, horizon, and metric combinations. A separate test compares GAS-PFN with SCPT, trained on records from a fraction of roads and evaluated on held-out roads.
At 10% road coverage, GAS-PFN had lower errors on all six datasets; at 20% and 30%, it remained better on five. These results support usefulness under the tested conditions, while the procedures differ.
Deep
- Original paper authors
- Joseph Natter, Long Vu, Yifan Zhang, Bobin Deng, Kun Suo
- Submitted to arXiv
- License
- arXiv non-exclusive