Data Mining in Transportation Networks with Graph Neural Networks: A Review and Outlook

Jiawei Xue, Ruichen Tan, Jianzhu Ma, Satish V. UkkusuriarXiv 2025 · DOI · arXiv

Abstract

Graph Neural Networks (GNNs) are revolutionizing transportation science by enabling sophisticated analysis of complex, non-Euclidean spatio-temporal data. This review provides a comprehensive roadmap of GNN applications, extending from traditional traffic forecasting to system operations, safety management, and large-scale industrial deployments. We explore the integration of GNNs with Large Language Models (LLMs) and physics-informed architectures, bridging the gap between academic research and real-world infrastructure.

Introduction

Transportation data is inherently complex, characterized by intricate spatio-temporal dependencies and a graph-like structure that traditional deep learning models often fail to capture. While sequence-based models handle temporal data well, the non-Euclidean nature of road networks requires specialized architectures. This paper addresses the limitations of existing literature, which often focuses narrowly on traffic prediction, by expanding the scope to include traffic operations and industrial practices.

Overview of the review framework, illustrating the connection between data mining problems, GNN models, and industry practices.

Methodology: The GNN Landscape

The paper categorizes the evolution of GNNs into three primary pillars:

  1. Generic and Variant GNNs: The field has progressed from spectral methods like Graph Convolutional Networks (GCNs) to spatial attention mechanisms (GATs) and scalable aggregation techniques like GraphSAGE. Modern variants now focus on robustness against adversarial attacks and handling out-of-distribution data.
  2. Integration with Large Language Models (LLMs): A cutting-edge frontier involves the synergy between GNNs and LLMs. GNNs can provide topology-aware embeddings as tokens for LLMs, while pre-trained LLMs can initialize node features in GNNs, leveraging vast "world knowledge" to compensate for sparse traffic data.
  3. Physics-Informed GNNs (PIGNN): To increase reliability for engineering applications, PIGNNs incorporate fundamental traffic flow laws—such as the conservation of flow—directly into the model's loss functions or architecture.
The fundamental mechanism of GNN layers, involving node embedding, neighborhood aggregation, and non-linear transformations.

Data and Benchmarking

The review emphasizes a shift from aging benchmarks like METR-LA to more expansive, modern datasets. Notable examples include LargeST, featuring data from 8,600 sensors, and the U.S. 20 Dataset, which covers 20 major cities. Multi-modal data, including vehicle trajectories and global vessel movements, are also becoming central to the field.

Industrial Insights and Results

A key finding is the "Node Disconnect": while academia often treats sensors as nodes, industry leaders like Google and Baidu treat road segments as nodes, making models more user-centric for tasks like Estimated Time of Arrival (ETA) prediction.

  • Google Maps Impact: Implementation of GNN-based ETA models led to a 16% to 51% reduction in highly inaccurate predictions across 19 cities.
  • LinkedIn Scalability: GNNs scaled to graphs with hundreds of billions of nodes, improving ad clicks by 2% and job applications by 1%.
  • Traffic Assignment: GNNs solved the Traffic Assignment Problem (TAP) with less than 2% error on standard test networks.

Future Directions and Challenges

Despite these successes, several challenges remain:

  • Scalability vs. Complexity: Bridging the gap between academic models (hundreds of nodes) and city-scale networks (millions of nodes).
  • Interpretability: Tools like Traffexplainer are essential for understanding the "why" behind model predictions.
  • Efficiency: The use of "Graph-less Neural Networks" (GLNNs) to distill complex GNN knowledge into faster MLP models for real-time inference.
  • Data Privacy: The barrier created by proprietary industry data highlights the need for more aggregated, non-sensitive data sharing.

Conclusion

GNNs represent a paradigm shift in transportation data mining. By moving beyond simple prediction to complex system operations and integrating with emerging technologies like LLMs, GNNs are poised to become the backbone of modern, intelligent transportation systems.

Summary updated

BibTeX
@misc{xue2025data,
  author = {Jiawei Xue and Ruichen Tan and Jianzhu Ma and Satish V. Ukkusuri},
  title = {{Data Mining in Transportation Networks with Graph Neural Networks: A Review and Outlook}},
  year = {2025},
  eprint = {2501.16656},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.16656},
}
Figure