PatchSTG: Scalable Spatiotemporal Graph Transformers for Traffic Forecasting on Irregular Sensor Networks

Researchers introduce PatchSTG, a new graph Transformer architecture that addresses scalability challenges in traffic forecasting by partitioning unevenly distributed sensors into geographic patches. The model reduces computational complexity from quadratic to near-linear while maintaining competitive forecasting accuracy across multiple prediction horizons.

PatchSTG represents a meaningful advancement in applying deep learning to real-world infrastructure challenges where data collection is inherently irregular. Traditional spatiotemporal models struggle with unevenly distributed sensor networks—a common problem in traffic systems, weather monitoring, and smart city applications—because they attempt to model all spatial relationships uniformly, creating computational bottlenecks at scale.

The core innovation centers on hierarchical spatial decomposition. By clustering sensors into locality-preserving patches based on geographic coordinates, the model transforms an irregular problem into a structured one. The dual attention mechanism then operates efficiently at two scales: intra-patch attention captures localized traffic dynamics (vehicle flow between nearby intersections), while inter-patch attention models broader patterns (traffic movement across city districts). This architectural choice directly addresses a fundamental tension in graph neural networks—balancing expressiveness with computational efficiency.

For intelligent transportation systems and smart city infrastructure, this work has practical implications. Current traffic forecasting models often fail to scale economically to metropolitan regions with hundreds or thousands of sensors. PatchSTG's near-linear complexity enables deployment on resource-constrained edge devices, potentially accelerating real-time traffic management systems. The method's success on Rhode Island data and large-scale datasets suggests generalizability across different urban configurations.

Beyond traffic, the patch-based approach applies to any spatiotemporal forecasting task with irregular sensor distributions—air quality monitoring, power grid management, and epidemic tracking. As cities increasingly deploy IoT sensor networks with uneven coverage, scalable modeling techniques become critical infrastructure components. The ablation studies validating both spatial partitioning and dual attention mechanisms provide evidence that this isn't merely an optimization trick but a conceptually sound modeling choice.

• →PatchSTG reduces computational complexity from quadratic to near-linear by partitioning irregular sensor networks into geographic patches
• →Dual attention mechanism separately models local intra-patch interactions and global inter-patch dependencies for efficient spatiotemporal learning
• →Model achieves competitive forecasting performance across multiple prediction horizons on real-world Rhode Island traffic data
• →Patch-based approach generalizes beyond traffic forecasting to any application with unevenly distributed spatiotemporal sensors
• →Architectural innovation enables deployment of traffic prediction systems on resource-constrained edge devices and metropolitan-scale networks