Spectral Adaptive Graph Transformer for Efficient and Scalable Traffic Forecasting

Key points

• It costs cubic time, is fixed by the road network, and provides a symmetric operator that cannot represent the directed propagation of traffic congestion.1
• Attention-based spatio-temporal transformers remove these limitations but generate a dense sensor-by-sensor attention map at each layer and time step, so memory, rather than accuracy, becomes the binding constraint on large networks.1
• We introduce the spectral adaptive graph transformer, whose spatial operator is learned rather than derived from the Laplacian, directed rather than symmetric, and conditioned on the current traffic state rather than a static state, at a cost linear in the number of sensors.1
• A sparse road graph diffusion branch provides the strictly local propagation that a low-rank operator cannot express, and a zero-initialized gate learns the balance between the two.1

Sources (1)

• [1]Spectral Adaptive Graph Transformer for Efficient and Scalable Traffic Forecasting
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 09:48 AM

  Traffic forecasting models based on spectral graph convolution offer an interpretable account of how information is mixed between sensors, but they rely on an eigendecomposition of the graph Laplacian.

  It costs cubic time, is fixed by the road network, and provides a symmetric operator that cannot represent the directed propagation of traffic congestion.

Extractive summary: sentences quoted from the sources.