Predicting traffic jams across an entire city is a spatial math problem, not just a weather forecast.
A new arXiv preprint proposes LoReST, a model built to solve a specific tension in large-scale traffic forecasting: nearby roads behave differently depending on identity and direction, but capturing citywide patterns by comparing every road segment to every other one is computationally brutal. The researchers split the problem into two layers. Local aggregation handles direction and road-specific relationships between neighboring segments, while a second layer groups the network into regions, pools their data, and runs attention between those regions instead of every individual node. Tested on four datasets from the LargeST benchmark, LoReST cut average error by 4.78% in MAE, 3.60% in RMSE, and 5.75% in MAPE compared to prior approaches.
The real story is the efficiency trick, not the accuracy bump. Full graph attention across a large road network scales quadratically and chokes on real metro-area sizes; region pooling sidesteps that cost while still passing long-range context back down to individual roads. That is the same computational bottleneck that has limited graph neural networks in other domains, from social networks to power grids.
Those error reductions are respectable, not stunning, and they come from benchmark datasets rather than a live city deployment. Whether region pooling holds up when road networks change through construction, closures, or new interchanges is a question this paper does not yet answer.