Traffic Forecasting: Transformers Not Always Needed for Global Spatial Info.

Traffic forecasting models often focus on capturing spatial dependencies, particularly global spatial information that represents interactions between all nodes in a traffic network. Many current models rely on complex Transformer architectures with adaptive attention mechanisms for this purpose. This study investigates whether such high-degree-of-freedom attention is truly necessary, or if simpler global aggregation operators could achieve similar results. Researchers designed an ablation framework to directly compare attention-based global interaction with a uniform full-range mixing approach, which significantly reduces spatial mixing complexity from O(N²) to O(N). Across six traffic benchmarks, both methods achieved comparable Mean Absolute Error (MAE), with only a 0.14% difference on average. The simpler uniform mixing performed better on three datasets, and standard spatial attention on the other three. A detailed analysis showed that spatial attention's contribution could be decomposed into a uniform global background and a non-uniform residual, with the residual offering only marginal, dataset-dependent value. This suggests that the use of complex spatial attention should be justified by stable gains beyond what a simpler global aggregation can provide.