Source-linked AI summary
Traffic Flow Forecasting with Spatial-Temporal Graph Diffusion Network
Xiyue Zhang, Chao Huang, Yong Xu, Lianghao Xia, Peng Dai, Liefeng Bo, Junbo Zhang, Yu Zheng
TL;DR
Traffic forecasting methods often miss global inter-region dependencies and the complex, multi-resolution temporal regularities needed for citywide prediction. ST-GDN combines multi-scale temporal attention with hierarchical graph diffusion to model local and global spatial structure. Across several datasets, the paper reports that ST-GDN consistently outperforms diverse baselines.
Problem
Existing traffic forecasting methods commonly focus on adjacent-region correlations and do not adequately encode global dependencies or multi-resolution temporal dynamics.
Method
ST-GDN combines resolution-aware self-attention and temporal hierarchy aggregation with hierarchical graph neural modeling of local and global traffic dependencies.
Results
ST-GDN consistently yields the best performance across the reported evaluation cases and outperforms baselines of different types over several datasets.
Takeaways & Limitations
The paper supports jointly integrating multi-resolution temporal patterns with global region-wise dependencies and geographical relations for spatial-temporal traffic modeling.
Abstract
from arXiv · showhide
Accurate forecasting of citywide traffic flow has been playing critical role in a variety of spatial-temporal mining applications, such as intelligent traffic control and public risk assessment. While previous work has made significant efforts to learn traffic temporal dynamics and spatial dependencies, two key limitations exist in current models. First, only the neighboring spatial correlations among adjacent regions are considered in most existing methods, and the global inter-region dependency is ignored. Additionally, these methods fail to encode the complex traffic transition regularities exhibited with time-dependent and multi-resolution in nature. To tackle these challenges, we develop a new traffic prediction framework-Spatial-Temporal Graph Diffusion Network (ST-GDN). In particular, ST-GDN is a hierarchically structured graph neural architecture which learns not only the local region-wise geographical dependencies, but also the spatial semantics from a global perspective. Furthermore, a multi-scale attention network is developed to empower ST-GDN with the capability of capturing multi-level temporal dynamics. Experiments on several real-life traffic datasets demonstrate that ST-GDN outperforms different types of state-of-the-art baselines. Source codes of implementations are available at https://github.com/jill001/ST-GDN.
Introduction
Traffic forecasting must capture both nonlocal spatial dependencies and complex multi-resolution temporal dynamics. ST-GDN addresses these gaps with a hierarchical graph architecture and multi-scale temporal attention, and experiments show stronger forecasting performance than diverse baselines.
- Motivation: Traffic forecasting supports intelligent transportation, public risk assessment, and mobility management based on predicted regional traffic volumes.The forecasting objective is to infer regional inflow and outflow from past traffic observations.
- Research gaps: Traffic patterns are multi-periodic, requiring collaboration across hourly, daily, and weekly resolutions.Recurrent approaches are described as effective mainly for short-term, smooth dynamics and limited on high-order multi-dimensional horizons.
- Research gaps: Existing methods often model temporal effects and correlations between adjacent regions but overlook broader cross-region dependencies.Regions with similar urban functions may be correlated despite being geographically distant.
- Proposed framework: ST-GDN uses multi-scale self-attention and temporal hierarchy aggregation to encode and combine multi-level temporal dynamics.The framework is designed to represent traffic transition regularities across different time resolutions.
- Proposed framework: ST-GDN jointly preserves local geographical relations and global traffic dependencies through a hierarchically structured graph neural architecture.Its graph diffusion design incorporates spatial semantics from adjacent regions to citywide traffic-pattern representations.
- Evaluation: Experiments on three real-world datasets report better forecasting performance for ST-GDN than baselines of different types.The study also evaluates model efficiency in the traffic prediction process.
Problem Definition
The problem is defined over grid-partitioned cities and uses historical incoming and outgoing traffic tensors as input. The model must infer future traffic volumes for the regions.
- Spatial regions: A city is partitioned into I × J disjoint grids, with each grid treated as a spatial region r_i,j for prediction.The regions are defined from geographical coordinates.
- Traffic flow tensor: Historical citywide traffic distributions are represented as a three-way tensor X ∈ R^I×J×T over regions and T time slots.Each tensor entry records a traffic-volume measurement for one region at one time slot.
- Traffic flow tensor: Two tensors, Xα for incoming traffic and Xβ for outgoing traffic, represent the two traffic-flow directions.These tensors support prediction of both incoming and outgoing traffic.
- Task formulation: Given Xα and Xβ from the past T time slots, the task is to infer unknown future traffic volumes for the regions.The output is a predictive function operating on the observed citywide traffic information.
Methodology
ST-GDN combines multi-resolution temporal modeling with attentive global and local spatial dependency learning, then fuses external factors for regional traffic prediction.
- Temporal Hierarchy Modeling: ST-GDN uses multi-scale self-attention to encode traffic dynamics across different temporal resolutions.Resolution-aware traffic series are transformed into query, key, and value matrices for attentive temporal aggregation.
- Temporal Hierarchy Modeling: Temporal resolution p represents the sampling interval, such as hourly, daily, or weekly measurements.Each resolution produces a corresponding traffic series with length T_p.
- Traffic Dependency Learning with Global Context: The global-context graph module learns dynamic traffic dependencies across regions through attentive message aggregation.The region graph contains spatial regions and pairwise relationships, while attention captures local and global dependencies.
- Traffic Prediction Phase: The framework fuses learned regional representations with external factors before predicting each region’s traffic volume.Weather, temperature, and wind speed are embedded and concatenated with regional representations in the prediction layer.
- Optimized Loss Function: ST-GDN jointly optimizes estimated incoming and outgoing traffic volumes using a weighted loss function.The parameter λ determines the relative influence of the two traffic-flow components.
- Model Complexity Analysis: The attentional graph module dominates ST-GDN’s computational cost, taking O(3 × I^2 × J^2 × d′).The multi-scale self-attentive network additionally takes O(3 × T × I × J × d) for projections and O(3 × T^2 × d) for weighted summation.
Evaluation
ST-GDN is evaluated on three real-world traffic datasets using RMSE and MAPE against diverse baselines. Results examine overall performance, component contributions, temporal resolutions, parameter sensitivity, and efficiency.
- Experimental Settings: Experiments use BJ-Taxi, NYC-Taxi, and NYC-Bike datasets, with traffic sampled every half hour for the taxi datasets and hourly for NYC-Bike.BJ-Taxi contains 34,000+ trajectories on a 32 × 32 grid; NYC-Taxi contains 22,000,000+ trajectories on a 10 × 20 grid; NYC-Bike uses a 16 × 8 grid.
- Experimental Settings: RMSE and MAPE are used to evaluate ST-GDN against traditional, neural, recurrent-convolutional, attention-based, and graph-based forecasting methods.The comparison includes ARIMA, SVR, Fuzzy+NN, RNN, and other baselines with varied model structures.
- Overall Performance: ST-GDN consistently achieves the best performance across all evaluated cases and outperforms the compared baselines in prediction-error visualization.The reported advantage is attributed to jointly modeling multi-level temporal dynamics and global-level region-wise dependencies.
- Ablation Study: The full ST-GDN outperforms variants that remove multi-scale attention, global graph attention, graph diffusion, neighborhood context, or external-factor fusion.This ablation result supports the positive contribution of each designed sub-module to prediction performance.
- Multi-Resolution Temporal Effects: The ST-GDNh,d,w configuration achieves the best temporal-resolution accuracy, while adding hourly, daily, and weekly patterns improves over singular-dimensional granularity.The study reports that decomposing temporal effects into multiple resolution-specific representations helps model temporal regularity and resolution-aware region relations.
- Parameter Sensitivity and Efficiency: Third-order graph-attention relations are sufficient for global dependencies, K = 3 gives the best graph-diffusion performance, and longer sequences or larger channels can introduce noise or overfitting.ST-GDN also achieves competitive efficiency against ST-GCN, although its attention-based propagation layer has higher computational cost than adjacent-matrix graph convolution.
Related Work
Prior traffic forecasting work uses neural architectures and graph methods to model spatial-temporal patterns, but ST-GDN emphasizes hierarchical temporal dynamics and globally enhanced region dependencies.
- Recurrent networks encode temporal traffic features, while later methods combine recurrent layers with convolutional networks for spatial-temporal modeling.
- ST-GDN preserves hierarchical temporal dynamics and globally enhanced region-wise dependencies in spatial-temporal representations.
- Graph-based approaches use graph convolutions and attention mechanisms to model correlations or aggregate information across regions and adjacent roads.
Conclusion
ST-GDN is a graph neural architecture that combines resolution-aware temporal encoding with local and global spatial dependency modeling. Experiments across several datasets show consistent superiority over many baselines.
- ST-GDN first encodes multi-level temporal signals with a resolution-aware self-attention network.
- The architecture then integrates local spatial context and global traffic dependencies across regions to enhance spatial-temporal representations.
- ST-GDN consistently significantly outperforms many baselines across several datasets.
- Future work targets deploying the developed prototype in a cloud-based system for real-time traffic prediction.