Gao, Xiaowei; (2025) Advanced Graph Deep Learning for Urban Traffic Crash Research. Doctoral thesis (Ph.D), UCL( University College London).

Preview

Text Gao_10210801_thesis.pdf Download (43MB) | Preview

Abstract

Traffic crashes remain a critical public health and economic challenge globally, causing approxi- mately 1.19 million deaths annually and imposing substantial economic costs of up to 3% of GDP based on the 2023 report of WHO. Despite significant advances in urban transportation systems, predicting and preventing crashes effectively remains challenging due to their complex, nonrecurring nature and the diverse factors influencing their occurrence. These challenges are compounded by several persistent limitations in current crash modelling approaches, including the prevalence of missing data in crash records that compromises prediction reliability, the com- plex spatiotemporal dependencies in crash patterns that traditional models struggle to capture, and the significant uncertainties in fine-grained predictions that limit practical applications. This thesis develops a unified advanced approach to crash analysis through Graph Deep Learning (GDL), handling both structured tabular records (containing crash details and casual- ties) and unstructured geographic crash patterns (reflecting spatial distributions and temporal evolution). By leveraging GDL’s advantages in graph representations, the research captures both explicit spatial-temporal relationships in geographic data and implicit correlations in tab- ular records, enabling comprehensive crash pattern analysis benefitting both crash risk and crash severity tasks. This unification provides a robust foundation for addressing increasingly complicated prediction challenges while maintaining practical applicability. The thesis develops a GDL framework that addresses fundamental data quality issues and crash prediction challenges. Firstly, the research addresses the pervasive problem of missing data in crash records, introducing a unified inexact matching bipartite graph representation learning approach that preserves complex relationships between observed and missing values at various missing rates. This foundation ensures reliable data for subsequent analysis while maintaining the integrity of critical feature relationships for injury severity prediction, which enhances the chance of the models being used in real-world applications. Building on this reliable data foundation, the research develops a multi-view adaptive hypergraph learning method for regional-level traffic risk prediction that captures both local and global dependencies. The innovation lies in the model’s ability to automatically learn meaningful relationships from diverse urban data sources, without relying on predetermined spatial relationships. Finally, the research culminates in addressing the challenges of road-level prediction, where data sparsity and uncertainty quantification are critical concerns. The final methodological innovation introduces a probabilistic graph framework integrated with a statistical model that not only handles extreme data sparsity but also provides reliable uncertainty estimates, enabling more informed decision-making. Using the UK’s STATS19 dataset as a case study, the research demonstrates how these advanced GDL-based methodological advances translate into practical improvements in crash prediction. The models consistently outperform existing state-of-the-art approaches while providing interpretable insights for safety interventions. These practical validations confirm the framework’s ability to bridge the gap between theoretical advances and real-world applications.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item