[Paper Review] Graph Neural Networks for Road Safety Modeling: Datasets and Evaluations for Accident Analysis

Nippani, Abhinav, Dongyue Li+3 more|arXiv (Cornell University)|Oct 31, 2023

Traffic Prediction and Management Techniques4 citations

TL;DR

This paper introduces a large-scale, unified dataset of 9 million traffic accident records from U.S. states, integrated with road network and traffic volume data, to evaluate graph neural networks (GNNs) for road safety modeling. Using GraphSAGE with multitask and transfer learning, the method achieves <22% mean absolute error in accident count prediction and >87% AUROC for binary accident occurrence classification across states.

ABSTRACT

We consider the problem of traffic accident analysis on a road network based on road network connections and traffic volume. Previous works have designed various deep-learning methods using historical records to predict traffic accident occurrences. However, there is a lack of consensus on how accurate existing methods are, and a fundamental issue is the lack of public accident datasets for comprehensive evaluations. This paper constructs a large-scale, unified dataset of traffic accident records from official reports of various states in the US, totaling 9 million records, accompanied by road networks and traffic volume reports. Using this new dataset, we evaluate existing deep-learning methods for predicting the occurrence of accidents on road networks. Our main finding is that graph neural networks such as GraphSAGE can accurately predict the number of accidents on roads with less than 22% mean absolute error (relative to the actual count) and whether an accident will occur or not with over 87% AUROC, averaged over states. We achieve these results by using multitask learning to account for cross-state variabilities (e.g., availability of accident labels) and transfer learning to combine traffic volume with accident prediction. Ablation studies highlight the importance of road graph-structural features, amongst other features. Lastly, we discuss the implications of the analysis and develop a package for easily using our new dataset.

Motivation & Objective

• To address the lack of public, comprehensive accident datasets for evaluating traffic accident prediction models.
• To develop a unified, large-scale dataset integrating traffic accident records, road network topology, and traffic volume data across multiple U.S. states.
• To evaluate existing deep learning methods—particularly graph neural networks—on accident prediction using the new dataset.
• To investigate the impact of structural road features and cross-state variability on model performance.
• To enable reproducible research by releasing a software package for accessing and using the new dataset.

Proposed method

• Constructed a large-scale, unified dataset from official U.S. state accident reports, enriched with road network structures and traffic volume data.
• Applied GraphSAGE, a GNN architecture, to model spatial dependencies and structural relationships in road networks for accident prediction.
• Employed multitask learning to jointly predict accident counts and binary accident occurrence, improving generalization across states with varying label availability.
• Used transfer learning to integrate traffic volume features into the GNN framework, enhancing predictive performance.
• Conducted ablation studies to isolate the contribution of graph-structural features versus other input features.
• Developed and released a software package to facilitate access and reuse of the dataset and models.

Experimental results

Research questions

• RQ1 How accurately can graph neural networks predict traffic accident occurrences on road networks using real-world data?
• RQ2 What is the impact of incorporating road network topology and traffic volume on accident prediction performance?
• RQ3 How do cross-state variations in data availability and labeling affect model generalization?
• RQ4 Which features—especially graph-structural features—are most critical for accurate accident prediction?
• RQ5 Can multitask and transfer learning improve the robustness and accuracy of accident prediction models?

Key findings

• GraphSAGE achieved a mean absolute error of less than 22% in predicting the number of accidents on road segments, relative to actual counts.
• The model achieved an AUROC score of over 87% for classifying whether an accident occurs on a given road segment.
• Ablation studies demonstrated that graph-structural features significantly improve prediction performance, highlighting their importance.
• Multitask learning effectively handled cross-state variability in accident label availability, improving model robustness.
• Transfer learning enabled effective fusion of traffic volume data with GNNs, further enhancing prediction accuracy.
• The proposed dataset and software package provide a foundation for reproducible and scalable research in road safety modeling.