[Paper Review] Traffic Flow Forecasting with Spatial-Temporal Graph Diffusion Network

Xiyue Zhang, Chao Huang+6 more|arXiv (Cornell University)|Oct 8, 2021

Traffic Prediction and Management Techniques21 citations

TL;DR

This paper proposes ST-GDN, a spatial-temporal graph diffusion network that captures multi-resolution temporal dynamics via a multi-scale self-attention mechanism and models both local and global spatial dependencies through a hierarchically structured graph neural network. The framework achieves state-of-the-art performance on multiple real-world traffic datasets by explicitly integrating local geographical correlations and long-range cross-region dependencies.

ABSTRACT

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.

Motivation & Objective

• To address the limitations of existing traffic forecasting models that focus only on local spatial correlations and fail to capture complex, time-dependent, multi-resolution traffic patterns.
• To explicitly model both local geographical dependencies and global cross-region inter-dependencies in traffic flow for improved prediction accuracy.
• To develop a unified framework that jointly learns hierarchical temporal dynamics and comprehensive spatial relationships across urban regions.
• To enhance representation learning in spatial-temporal forecasting by integrating multi-grained temporal signals and global semantic relationships in a graph-based architecture.

Proposed method

• A multi-scale self-attention network is designed to encode temporal dynamics across different time resolutions (e.g., hourly, daily, weekly), capturing hierarchical temporal patterns.
• A temporal hierarchy aggregation layer is introduced to model dependencies between multi-level temporal representations, enabling collaborative learning of granular temporal signals.
• A hierarchically structured graph neural network combines graph attention networks and convolution-based graph diffusion to model both local spatial adjacency and global region-wise dependencies.
• The graph diffusion mechanism enables the propagation of traffic state information across distant regions, capturing long-range spatial correlations beyond immediate neighbors.
• The framework integrates spatial and temporal representations through a joint learning paradigm, enhancing the expressiveness of the final traffic flow predictions.
• The model is trained end-to-end using a reconstruction loss on historical traffic flow tensors, with attention weights and diffusion processes jointly optimized.

Experimental results

Research questions

• RQ1 Can modeling multi-resolution temporal dynamics improve the accuracy of traffic flow forecasting in urban environments?
• RQ2 How can global inter-region dependencies—beyond immediate spatial neighbors—be effectively captured in traffic prediction models?
• RQ3 To what extent does combining local spatial correlations with global spatial semantics enhance forecasting performance?
• RQ4 Can a graph diffusion-based architecture effectively model high-order spatial dependencies in traffic networks?
• RQ5 How does the integration of multi-scale attention and hierarchical graph learning compare to conventional GNN or RNN-based approaches in traffic forecasting?

Key findings

• ST-GDN outperforms multiple state-of-the-art baselines on three real-world datasets, including BJ-Taxi, NYC-Taxi, and NYC-Bike, in terms of MAE, RMSE, and MAPE metrics.
• On the NYC-Bike dataset, ST-GDN achieves a MAE of 569.26, RMSE of 891.63, and MAPE of 0.20, surpassing all compared methods including DCRNN, ST-GCN, and GMAN.
• The model demonstrates superior performance in capturing complex, multi-periodic traffic patterns due to its ability to model multi-grained temporal dynamics through the multi-scale attention mechanism.
• The hierarchical graph neural architecture enables ST-GDN to effectively learn both local and global spatial dependencies, significantly improving prediction accuracy over models that only consider local neighborhoods.
• Despite higher computational cost than ST-GCN due to the attention-based graph propagation, ST-GDN achieves a favorable trade-off between accuracy and efficiency, with competitive inference times.
• Ablation studies confirm that both the multi-scale temporal modeling and global spatial dependency learning components contribute significantly to the model's performance gains.