An Efficient and Effective Architecture for Large-Scale Traffic Prediction via Geometry-Adaptive Square Partitioning

The paper introduces SqLinear, a novel architecture for large-scale traffic prediction that employs a geometry-adaptive algorithm called Square Partition, which creates balanced, non-overlapping spatial regions for sensor data. This approach addresses the limitations of existing partitioning methods by ensuring high-quality data segmentation, and it incorporates a Hierarchical Linear Interaction (HLI) module that replaces traditional attention mechanisms with a linear interaction scheme, resulting in linear computational complexity. Extensive evaluations demonstrate that SqLinear achieves a 2.30% reduction in mean absolute error (MAE) on average and significantly improves training efficiency, making it a valuable advancement for practitioners working with large-scale spatiotemporal models in traffic systems.