Dynamic correlation analysis of urban crashes using Tucker-net based SIRS model: A case study in New York City

Understanding the spatiotemporal dynamics of traffic crash hotspots remains a critical challenge in urban safety management, hindered by two fundamental limitations: the inherent randomness and sparsity of crash data impede robust hotspot identification, while existing frameworks fail to model dynamic correlations between geographically dispersed zones. This paper introduces a new data-driven and transferable analysis model: Tucker-Net based Susceptible-Infectious-Recovered-Susceptible (SIRS) model (TNBSM) to address this gap. TNBSM designs a novel network named Tucker-Net for investigating the correlation between sparse crash data with a modified epidemic SIRS model, effectively capturing the spatiotemporal correlations of crash zones. Empirically validated using data in the New York City, TNBSM reveals significant discrepancies in crash trends between weekdays and weekends, and long-term semantic correlations between crash zones, highlighting its effectiveness in understanding highly dynamic correlation of urban traffic crash. The robustness and potential of the TNBSM model are further substantiated through three key aspects: comparative predictive capability analysis with three baseline models, assessment of parameter stability, and examination of model unbiasedness.