GAS-PFN

GAS-PFN (Graph-Aware Spatiotemporal Prior-Data Fitted Network) is a method that explicitly incorporates graph structure into the PFN prior to achieve cross-network zero-shot traffic forecasting without parameter updates.

The paper uses random-walk diffusion to incorporate sampled road topologies into the synthetic task prior, enabling the model to learn graph-coupled dynamic patterns during pretraining.

GAS-PFN uses two masked-conditioning forward passes sharing the same Transformer backbone: the first pass models the temporal dynamics specific to each sensor, the second pass focuses on spatiotemporal interactions among neighboring sensors, and the outputs of the two passes are integrated for the final prediction.

On six benchmark datasets with a 90-minute forecast horizon, GAS-PFN reduces MSE by an average of 34.0% relative to its closest PFN predecessor, TimePFN.

In the same evaluation, GAS-PFN reduces MSE by an average of 27.7% relative to four recently industry-developed general zero-shot forecasters.

In GAS-PFN's evaluation comparisons, all models receive 90 minutes of continuous recent history, while additional context is extracted from each dataset's chronologically split training set.

Masked-conditioning forward propagation means controlling the scope of information flow in the Transformer input via attention masks; in GAS-PFN the first pass uses a mask that only allows attention between time steps of the same sensor, and the second pass uses a mask that allows attention across neighboring nodes.