Abstract:

The bike-sharing(BS) flow rate reflect the degree of surplus and shortage of vehicles in urban spatial environment. Understanding its changes and incentives is of great significance for urban BS scheduling. Due to the complexity and variability of travel purposes and external environmental factors, it is difficult to analyze the relationship between the BS flow rate and the characteristics of the built environment through a statistical model with linear assumptions. Therefore, this study explored the contribution of the built environment to the BS flow rate and the nonlinear effects on the flow rate, as well as the changes of the nonlinear model of the BS flow rate on weekdays and weekends based on the data of BS in the downtown of Shanghai, through the extreme gradient boosting tree model (XGBoost) and the interpretive method partial dependence plot (PDP) of machine learning. The results show that the feature importance and nonlinear mechanism are significantly different in the two periods. The density of residential population, educational facilities and residential facilities has a high degree of explanation for the weekday BS flow rate, which is 19.18%, 13.16% and 12.92%, respectively, and has a significant threshold effect. The density of residential population and the density of educational facilities have a positive impact on the net BS outflow rate, reaching the maximum at 11 600 person per km² and 8 educational facilities per km² respectively; the density of residential facilities has a negative impact on the net BS outflow rate, and the corresponding threshold is 40 residential facilities per km².There is little difference in the explanatory degree of each variable to weekend BS flow rate, nevertheless the nonlinear relationship cannot be ignored. Specifically, the distance to the city center and bus line number density have a significant positive impact on the weekend net BS inflow rate, with the effective range of 18~23 km and 28~52 routes per km². The positive influence range of plot ratio on net BS outflow rate at weekends is 0.89~1.41. The above findings show that XGBoost model can effectively compensate for the bias of linear assumption of traditional regression model (MLR), and the disclosure of the contribution degree and influence scope of built environment characteristics also provides decision-making suggestions for the management department for BS dispatching in areas with different built environment levels.

Key words: bike-sharing flow rate, built environment, extreme gradient boosting tree model, non-linear, scheduling management