Understanding cycling distance according to the prediction of the XGBoost and the interpretation of SHAP: A non-linear and interaction effect analysis
Shujuan Ji,Xin Wang,Tao Lyu,Xiaojie Liu,Yuanqing Wang,Eva Heinen,Zhenwei Sun
DOI: https://doi.org/10.1016/j.jtrangeo.2022.103414
IF: 5.899
2022-07-01
Journal of Transport Geography
Abstract:Cycling benefits both the individual and society in terms of public health promotion, traffic congestion relief and vehicle emissions reduction. To better understand cycling behaviors, we analyze non-linear relationships and interaction effects between the built environment and cycling distance. Few studies explore the interaction effects on cycling distance in which road network patterns interact with the demographic, trip, and other built environment characteristics to produce complex effects. Previous research has not examined the effect size or relative contribution of various variables have on cycling distance. Thus, this study adopts eXtreme Gradient Boosting (XGBoost) to examine the non-linear relationships among road network patterns, demographic, trip, and bike lane infrastructure, and other built environment characteristics and cycling distance, and employs SHapley Additive exPlanations (SHAP) to discover complex interaction effects on cycling distance, based on a bike travel survey in Xi'an, China. The results show that road network patterns have the greatest contribution; bike lane infrastructure is also quite important and has larger collective contributions than land use patterns and socioeconomics. Average geodesic distance and network betweenness centrality, two topological indices to describe road network structure, interact with bike lane infrastructure, land use and demographic characteristics to produce interaction effects on explaining cycling behaviors. When the average geodesic distance is smaller than 2.8 and the network betweenness centrality is smaller than 50%; as average geodesic distance increases, the network betweenness centrality has a positive effect on cycling distance. A network with a lower average geodesic distance, and with a higher intersection density makes cyclists ride a longer distance. A network with a higher value of average geodesic distance discourages cyclists to detour.
transportation,economics,geography