Local or Neighborhood? Examining the Relationship between Traffic Accidents and Land Use Using a Gradient Boosting Machine Learning Method: The Case of Suzhou Industrial Park, China

Abstract

In cities, road traffic accidents are critical endangerment to people’s safety. A vast number of studies which are designed to understand these accidents’ leading causes and mechanisms exist. The widely held view is that emerging analysis methods can be a critical tool for understanding the complex interactions between land use and urban transportation. Using a case study of Suzhou Industrial Park (SIP) in Suzhou, China, this paper examines the relationship between different land use types and traffic accidents using a gradient boosting model (GBM) machine learning method. The results show that the GBM can be used as an effective accident model for a variety of research and analysis methods by (1) ranking the influential factors, (2) testing the degree of interpretation of each variable as the complexity of iterations changes, and (3) obtaining partial dependence plots, among other methods. The findings of this study also suggest that land use types—including facility points—demonstrate differing degrees of influence at two geographical scales: local level and neighborhood level. In the ranking of relative importance at both scales, the variables of education institutions, traffic lights, and service institutions are all ranked high—with a more significant influence on the occurrence of accidents. However, residential land and land use mix variables differed significantly in both scales and showed a significant deviation compared to the other results. When adjusting the complexity of the decision tree, the local level is more suitable for measuring variables such as residential areas and green parks where pedestrians and vehicles have fixed mobility periods and moderate flows. On the contrary, the nearest neighborhood level is more suitable to a small number of variables related to public service facilities at fixed locations, such as traffic lights and bus stops. In the partial dependence plots, all variables, except educational institutions and residences, show a positive correlation for accidents in the fitting process. The results of this study can ideally help inform transportation planners to reconsider transport accident occurrence rates in the context of the proximity to various land use types and public service facilities.

KGCN‐LSTM: A graph convolutional network considering knowledge fusion of point of interest for vehicle trajectory prediction

“…The neighbour information of the target node is aggregated and the parameters are updated. ReLu is considered as the activation function, and the forward propagation of the whole feature fusion part is shown in formula (13), where NV j is the j th node representation vector of the ith trajectory.…”

“…POI has been analyzed by Chen et al [12] by creating a spatial and temporal knowledge graph, and it has also been mined with bike-sharing to explore the significance of features for different commuting purposes [11]. POI semantic information is also extracted to identify the characteristic of land use [13], implement urban functional zoning, and realize the fast-charging station planning for urban electric vehicles [14]. Therefore, analyzing vehicle trajectory by considering POI as the spatial feature can effectively improve the efficiency of traffic management.…”

Risk Measurement Model for Vehicle Group Based on Temporal and Spatial Similarities

“…Conventional methods primarily use crash data and statistical techniques for traffic safety analysis. Generally, the crash rate or death rate is combined with possible influencing factors, including speed [10], population [11], traffic volume [12], and land use [13], to analyze traffic safety. For instance, the Smeed model uses regression analysis to combine population, motor vehicle ownership, and crash fatalities to analyze traffic safety [14].…”

Measuring and comparing the spatiotemporal evolution of accident hot spots

“…A set of different statistical analysis methods has been applied in efforts to determine the correlation between categorical variables over accident datasets. For example, Negative Binomial regression (Msengwa and Ngari, 2021;Mustefa and Belayhun, 2019;Wang et al, 2021;Zou et al, 2021) and Poisson regression (Genowska et al, 2021;Khan and Hussain, 2021;Msengwa and Ngari, 2021;Sagamiko and Mbare, 2021;Twenefour et al, 2021;Wang et al, 2021;Yang et al, 2021), have been widely used to model of count accident dataset to verify the impact of independent variables 𝑋 on the given dependent variables 𝑌. The main difference between Poisson and Negative Binomial when the later releases the restricted hypothesis and the mean value is equal to the variance that made by the Poisson model (Maxwell et al, 2018).…”