GAS-PFN adds road graphs to zero-shot traffic forecasts
Original title: GAS-PFN: Graph-Aware Spatiotemporal Prior-Data Fitted Networks for Zero-Shot Traffic Forecasting
At a 90-minute horizon it cuts MSE 34.0% versus TimePFN and 27.7% versus four general zero-shot forecasters.
GAS-PFN puts sampled road topology into Prior-Data Fitted Network pretraining through random-walk diffusion. Two mask-conditioned passes of a shared Transformer model sensor-specific dynamics and neighbor interactions, then combine them without updating parameters on the target network. Across six benchmarks, the authors report lower mean squared error than TimePFN and four industry zero-shot forecasters given 90 minutes of recent history.
- TimePFN is described as the closest PFN-based predecessor.
- Additional context comes from each dataset's chronological training split.
- Inference needs no parameter updates on the target network.
- The comparison covers six benchmarks at the 90-minute horizon.
From the original article
1 min read at arxiv.orgTraffic forecasting requires modeling complex temporal dynamics and the spatial dependencies imposed by road topology. Supervised graph-based models achieve strong performance, but are often trained for a particular dataset and sensor graph, making deployment to new or evolving networks costly.
By Joseph Natter · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.