Traffic forecasting models built on spectral graph convolution are attractive because they show explicitly how information is mixed between sensors. But that clarity comes at a price: the underlying eigendecomposition of the graph Laplacian costs cubic time and is fixed once computed, making it hard to scale to large or changing road networks.
The new paper introduces a Spectral Adaptive Graph Transformer designed to keep the interpretability of spectral methods while removing the expensive, rigid eigendecomposition step. By combining spectral graph convolution with a transformer architecture, the approach aims to make forecasting both efficient and scalable.
The abstract is brief, so the precise mechanism is not fully detailed in the source. What is clear is the motivation: existing spectral models trade scalability for interpretability, and this work proposes a way to get both.