Accurate ETA Prediction Using Dynamic Route-Aware Graph Neural Networks for Improved Urban Mobility and Well-Being
Guy Tordjman and Nadav Voloch
DSTRA-GNN predicts ETA with a vehicle-centered dynamic graph and explicit route awareness, cutting error on SUMO and Porto taxi data versus static-graph baselines.
Estimated time of arrival usually comes from traffic history without knowing the vehicle's planned route. This paper asks whether a vehicle-centered dynamic graph with route awareness does better.
We introduce DSTRA-GNN (Dynamic Spatio-Temporal Route-Aware GNN). Vehicles are graph nodes, interaction edges capture local traffic, and a GRU aggregates graph snapshots over time. A mixture-of-experts layer combines vehicle dynamics with route intent. We evaluate on SUMO simulations and Porto-G, a graph form of the Porto taxi trajectories.
Ablations show compounding gains from route awareness and temporal context. On duration-matched Porto-G, MAE falls from 93.3 s (static graph only) to 68.1 s. On SUMO, MAE falls from 80.1 s to 54.8 s under the shared protocol. Vehicle-centered dynamic graphs with explicit route state are a competitive alternative to trajectory-native sequence models.