transportformer

Transportformer is a family of neural network architectures designed for transportation data. It combines graph representations of transportation networks with transformer-based temporal modeling to capture both the spatial structure of roads and the dynamic evolution of traffic. The goal is to support tasks such as traffic forecasting, route optimization, and mobility planning in urban environments.

The architecture typically integrates a graph neural network component to encode network topology and interactions among nodes and edges, with a temporal transformer that applies self-attention over time to capture long-range dependencies. Some variants use hierarchical graphs, multi-scale representations, adaptive adjacency, and combined spatial-temporal encoders. Time-of-day and day-of-week embeddings are commonly used.

Inputs and outputs. Inputs include a static network graph (nodes, edges), dynamic features such as speeds, volumes, occupancy, external signals like weather or events, timetable data, and historical travel times or origin-destination matrices. Outputs cover predicted speeds and travel times, arrival time estimates, demand forecasts for mobility services, and routing costs for decision-support systems.

Applications include urban traffic forecasting, transit scheduling, logistics and last-mile delivery, autonomous vehicle routing, and incident management.

Development and impact. The approach emerged in the mid-2020s as researchers adapted transformer models to irregular graphs and real-time streams. Transportformer methods are evaluated on standard traffic datasets and city-scale networks, often showing improvements over baseline GNNs and recurrent models in short-term forecasting and in robustness to missing data.

Limitations. Limitations include computational cost, data requirements, and sensitivity to data quality; generalization across cities can be limited; interpretability remains a challenge for decision-makers.