Uncertainty Quantification in the Road-level Traffic Risk Prediction by Spatial-Temporal Zero-Inflated Negative Binomial Graph Neural Network(STZINB-GNN)

Xiaowei Gao, James Haworth, Dingyi Zhuang, Huanfa Chen, Xinke Jiang

Abstract

Urban road-based risk prediction is a crucial yet challenging aspect of research in transportation safety. While most existing studies emphasize accurate prediction, they often overlook the importance of model uncertainty. In this paper, we introduce a novel Spatial-Temporal Zero-Inflated Negative Binomial Graph Neural Network (STZINB-GNN) for road-level traffic risk prediction, with a focus on uncertainty quantification. Our case study, conducted in the Lambeth borough of London, UK, demonstrates the superior performance of our approach in comparison to existing methods. Although the negative binomial distribution may not be the most suitable choice for handling real, non-binary risk levels, our work lays a solid foundation for future research exploring alternative distribution models or techniques. Ultimately, the STZINB-GNN contributes to enhanced transportation safety and data-driven decision-making in urban planning by providing a more accurate and reliable framework for road-level traffic risk prediction and uncertainty quantification.

Introduction

Predicting traffic risk at the road level is notoriously difficult due to the "rare event" nature of accidents. On any given day, most road segments experience zero accidents, a phenomenon known as zero-inflation. Traditional grid-based models often fail to provide the granularity needed for specific street-level interventions. Furthermore, standard models typically offer only "point estimates" without quantifying the associated uncertainty. For city planners, understanding the confidence of a prediction is as vital as the prediction itself. This research addresses these challenges by introducing STZINB-GNN, a framework designed for high-resolution prediction and robust uncertainty quantification in sparse data environments.

Methodology

The STZINB-GNN framework integrates three core components to model the complex dynamics of traffic risk:

  1. Temporal Learning (GRU): A Gated Recurrent Unit identifies time-based patterns in historical traffic data.
  2. Spatial Learning (GAT): A Graph Attention Network models the road network as a graph, capturing "spillover effects" where traffic conditions on one segment influence adjacent roads.
  3. Probabilistic Output (ZINB): The model predicts parameters for a Zero-Inflated Negative Binomial distribution. The zero-inflated component ($\pi$) accounts for the high probability of zero accidents, while the Negative Binomial component handles over-dispersion and variance.
The STZINB-GNN architecture, showing the transition from road network data to deterministic spatial-temporal learning and finally to probabilistic parameter estimation.

By outputting a distribution rather than a single value, the model provides inherent confidence intervals, effectively quantifying its own uncertainty.

Experimental Setup

The model was evaluated using 2019 data from the Lambeth Borough in London, UK. The dataset included 5,659 road segments and 1,335 accidents. Notably, the data exhibited a 98.7% zero-inflation rate, presenting a significant "needle in a haystack" challenge for predictive modeling.

Results and Discussion

The STZINB-GNN was compared against several baselines, including standard Gaussian models (STG) and Historical Averages (HA).

Performance comparison of STZINB-GNN against baseline models across various metrics for long-term and short-term predictions.

Key Findings:

  • Predictive Accuracy: In long-term scenarios, STZINB-GNN achieved a Mean Absolute Error (MAE) of 0.054, significantly outperforming the STG baseline (0.118).
  • Handling Sparsity: The model's True-Zero Rate (ZR) of 0.641 demonstrates its superior ability to correctly identify safe roads compared to the STG model (0.199).
  • Hit Rate: The framework achieved a hit rate of approximately 61.8% when identifying accident locations within the top 20% of predicted high-risk roads.
  • Uncertainty as a Tool: By calculating Kullback-Leibler Divergence (KLD) and information entropy, the model identifies high-uncertainty areas, signaling to policymakers where improved data collection or additional sensors may be required.

Conclusion

The STZINB-GNN provides a more accurate and reliable framework for road-level traffic risk prediction by explicitly modeling zero-inflation and quantifying uncertainty. While the Negative Binomial distribution is highly effective for count data, future work will explore alternative distributions for continuous risk scores and integrate diverse data sources like real-time weather to further refine prediction reliability.

Summary updated

Abstract

Urban road-based risk prediction is a crucial yet challenging aspect of research in transportation safety. While most existing studies emphasize accurate prediction, they often overlook the importance of model uncertainty. In this paper, we introduce a novel Spatial-Temporal Zero-Inflated Negative Binomial Graph Neural Network (STZINB-GNN) for road-level traffic risk prediction, with a focus on uncertainty quantification. Our case study, conducted in the Lambeth borough of London, UK, demonstrates the superior performance of our approach in comparison to existing methods. Although the negative binomial distribution may not be the most suitable choice for handling real, non-binary risk levels, our work lays a solid foundation for future research exploring alternative distribution models or techniques. Ultimately, the STZINB-GNN contributes to enhanced transportation safety and data-driven decision-making in urban planning by providing a more accurate and reliable framework for road-level traffic risk prediction and uncertainty quantification.
BibTeX
@inproceedings{gao2023uncertainty,
  author = {Xiaowei Gao and James Haworth and Dingyi Zhuang and Huanfa Chen and Xinke Jiang},
  title = {{Uncertainty Quantification in the Road-level Traffic Risk Prediction by Spatial-Temporal Zero-Inflated Negative Binomial Graph Neural Network(STZINB-GNN)}},
  booktitle = {GIScience},
  year = {2023},
  url = {https://openreview.net/forum?id=zBoLYNOSgp},
}
Figure