Uncertainty-Aware Probabilistic Graph Neural Networks for Road-Level Traffic Accident Prediction

TL;DR

This work tackles fine-grained road-level crash prediction under spatial sparsity and uncertainty by introducing STZITD-GNN, an end-to-end spatiotemporal graph neural network that combines a GRU-based temporal encoder with a GAT-based spatial encoder and a four-parameter Zero-Inflated Tweedie decoder. By modeling both crash frequency and severity through a four-parameter distribution (pi, mu, phi, rho) and explicitly handling zero inflation, the approach yields calibrated multi-step forecasts (p = 14) with improved point estimates and uncertainty quantification. Empirical results on 2019 London crash data across three boroughs show substantial gains in MAE/MAPE/RMSE, narrower yet reliable prediction intervals (MPIW with competitive PICP), and superior zero-crash identification and high-risk crash detection compared to baselines. The framework offers interpretable uncertainty and actionable insights for targeted road monitoring, contributing to safer urban mobility and scalable deployment across cities with similar data characteristics.

Abstract

Traffic accidents present substantial challenges to human safety and socio-economic development in urban areas. Developing a reliable and responsible traffic accident prediction model is crucial to addressing growing public safety concerns and enhancing the safety of urban mobility systems. Traditional methods face limitations at fine spatiotemporal scales due to the sporadic nature of highrisk accidents and the predominance of non-accident characteristics. Furthermore, while most current models show promising occurrence prediction, they overlook the uncertainties arising from the inherent nature of accidents, and then fail to adequately map the hierarchical ranking of accident risk values for more precise insights. To address these issues, we introduce the Spatiotemporal Zero-Inflated Tweedie Graph Neural Network STZITDGNN -- the first uncertainty-aware probabilistic graph deep learning model in roadlevel traffic accident prediction for multisteps. This model integrates the interpretability of the statistical Tweedie family model and the expressive power of graph neural networks. Its decoder innovatively employs a compound Tweedie model,a Poisson distribution to model the frequency of accident occurrences and a Gamma distribution to assess injury severity, supplemented by a zeroinflated component to effectively identify exessive nonincident instances. Empirical tests using realworld traffic data from London, UK, demonstrate that the STZITDGNN surpasses other baseline models across multiple benchmarks and metrics, including accident risk value prediction, uncertainty minimisation, non-accident road identification and accident occurrence accuracy. Our study demonstrates that STZTIDGNN can effectively inform targeted road monitoring, thereby improving urban road safety strategies.

Figures (10)

• Figure 1: The overall framework of STZITD-GNNs. STZITD-GNNs utilise the ST-GNN Encoder $\mathcal{ST}$ (composed of GRU and GAT encoders) to encode the history time window $1:t$ crash risk $Y_{1:t}$ and road features $X_{1:t}$ with the use of the road connection graph in the spatial-temporal embedding of the road $\mathcal{Z}$. After encoding, the four parameter decoders map $\mathcal{Z}$ into the ZITD parameter space and obtain $\pi_{t+1:t+p}, \mu_{t+1:t+p}, \phi_{t+1:t+p},\rho_{t+1:t+p}$ for the predicted time window $t+1:t+p$, which determine the predicted road traffic crash risk distribution $f(Y_{t+1:t+p})$.
• Figure 2: The map of three boroughs, London
• Figure 3: The distribution of 2019-crash risk scores frequency in the three boroughs
• Figure 5: Real v.s. Predicted Crash Risks on Lambeth, Tower Hamlets, and Westminster.
• Figure 6: Temporal variation of MPIW for baseline and STZITD-GNN models.