Evaluating the Nonlinear Impacts of Built Environment on Online Car-Hailing: Focus on the Whole Day

Abstract

Understanding the relationship between the built environment and online car-hailing travel behavior can greatly influence sustainable urban development. These complex impacts are not fully understood from a whole-day perspective. This paper took the central urban area of Nanning, China as a case study, and analyzed the data of 1,154,759 online car-hailing service orders collected by the Nanning Information Center over the course of a week. Thiessen polygons based on road networks were used to divide the research area. A partial dependence plot and random forest regression were used to examine the nonlinear impacts of the built environment (home-work attributes, traffic facility attributes, land use, and diversity) on online car-hailing trip demand throughout the day. The results show that the built environment elements that affect the demand for online car-hailing differ at different times of the day and that they have different ranges of influence. For example, residential areas play an important role in the commuter period for demand with a contribution rate of 33.2%, while the contribution rates for the remaining periods are less than 5%. In the whole day, cultural services, health services, leisure sports services, and catering services are important factors. And if there are enough training institutions to form a separate trading area within a city, this will have a profound impact on travel demand. In comparison to subway stations, bus stations contribute more to the demand. The study results will assist the government in improving urban transportation coordination.

Keywords

online car-hailing built environment nonlinear impacts machine learning sustainability

Over the past decade, on-demand platform mobile services have been fully developed as a result of rapid advances in information and communication technology. Among the emerging modes of shared transportation, shared bikes, electric scooters, and online car-hailing (OCH) have gained a lot of popularity ( 1 , 2 ). OCH can reduce urban transport emissions if the scale of the service is sufficient to minimize the cruising time of drivers ( 3 ), which demonstrates that shared transportation can reduce urban traffic congestion and pollutant emissions to some extent. Over the past six years, the OCH business has grown rapidly in China, and it plays an important role in people’s travel activities. According to the China OCH regulatory information exchange platform, 214 OCH companies had been granted business licenses in China as of December 31, 2020. As of December 2020, the platform had received 810 million orders ( 4 ). With its flexibility and door-to-door service, OCH can almost match the convenience of private cars, particularly in cases where users go to places where parking is difficult and parking fees are high as OCH not only reduces the need for parking, but also reduces waiting times ( 5 ).

While OCH offers many benefits, its scale in cities must be limited within a reasonable range. It is very important to ensure the load factor. By reducing the vehicle’s empty-loading cruise, we can reduce energy consumption, which further reduces traffic congestion and environmental pollution. Before picking up customers, having a clear understanding of the OCH demand distribution can greatly improve the load factor ( 3 ). In light of this, exploring the urban OCH ridership is an important task, and the research shows that the built environment has a significant impact on transit ridership ( 6 – 8 ) and online car-hailing ( 9 – 14 ). These empirical studies point out that factors related to home-work attributes, land-use density, and diversity and transport facility characteristics all significantly affect the passenger flow.

It is worth noting that our study is different from the previous ones. To distinguish from previous studies, we summarized the main features of relevant past studies, as listed in Table 1. Table 1 indicates that most of these studies explored the linear relationship between environmental factors and travel demand (e.g., least squares, geographically weighted regression, multiscale geographically weighted regression, logistic regression), and the scale of influence on individual environmental factors is limited. In fact, when the built environment features are at different ranges, the impact on passenger flow may also be different ( 15 ). This nonlinear relationship has been demonstrated for other modes of travel (e.g., subway, bus, and walking) ( 6 , 16 , 17 ). In addition, most of these studies neglected to examine the relationship between the two on a time scale (e.g., peak hours were used). Few studies explored it from a full-day perspective and still used the assumption of linearity. It is important to note that such typical analyses may mask important partial information about the impact of a policy change and have limitations in examining its temporal heterogeneity ( 18 ). Last but not least, most of the existing studies have used a spatial grid division method when carrying out the division of study units. This simple and straightforward division ensures consistency of area but ignores the impact of road network distribution on spatial travel.

Table 1.

Characteristics of the Study of the Relationship between OCH and the Built Environment

Author	Focus	Method	Division method of study units	Temporal perspectives
Wang et al. (2022) ( 19 )	The influence mechanism of urban built environment on OCH travel	The spatial Durbin model and the mixed geographically weighted regression model	0.5 km spatial grids division	Peak hours: 13:00–14:00
Du et al. (2022) ( 20 )	Spatiotemporal variation of high-efficiency ride-hailing orders	Ordinary least squares and geographically weighted regression	0.5 km spatial grids division	12:00–06:00am, 06:00–noon, noon–17:00, 17:00–midnight–2:00am
Zheng et al. (2022) ( 21 )	The impact of the built environment on ride-hailing	Ordinary least squares and geographically weighted models	Divided into 5,415 research units according to the natural blocks surrounded by the road network	Weekdays versus weekends
Gehrke and Huff (2022) ( 22 )	Spatial equity implications and neighborhood indicators of ride-hailing trip frequency	Geographically weighted regression	One-mile hexagon zonal system	Every year
Zhao et al. (2022) ( 9 )	The effect of the urban built environment on OCH ridership	Ordinary least squares/ geographically weighted regression/ multiscale geographically weighted regression	0.5 km~5 km spatial grids division	Indiscriminating
Wang et al. (2021) ( 11 )	Influencing factor analysis and demand forecasting of intercity OCH travel	Geographically weighted regression	0.5 km × 0.66 km spatial grids division	Indiscriminating
Dean and Kockelman (2021) ( 23 )	Spatial variation in shared ride-hail trip demand	Ordinary least squares/spatial autoregressive model/spatial error model	Census tract	07:00–10:00, 10:00–14:00, 14:00–18:00, 18:00–22:00, 122:00m–07:00
Wang and Noland (2021) ( 24 )	Variation in ride-hailing trips	Ordinary least squares and geographically weighted regression	0.2 km × 0.2 km spatial grids division	06:00–10:00,16:00– 19:00, 23:00–03:00
Zhang et al. (2020) ( 13 )	The influence of different types of POIs on the OCH travel intensity	Ordered logistic regression	0.19 km × 0.22 km spatial grids division	Early morning, morning, afternoon, evening
Sabouri et al. (2020) ( 15 )	The influence of built environment on Uber demand	Hierarchical linear modeling	Census block group	Indiscriminating
Zhang et al. (2020) ( 25 )	Spatiotemporal varying effects of built environment on ride-hailing ridership	Improved parallel-based GTWR	Transportation analysis zone	Every month
Bi et al. (2020) ( 14 )	The effect of built environment on OCH ridership	Geographically weighted regression	Voronoi diagram cell based on bus stops	06:00–10:00, 10:00–14:00, 14:00–18:00, 18:00–23:00, 23:00–06:00
Li et al. (2019) ( 12 )	The impacts of built environment variables on DiDi’s demand	Geographically weighted regression	0.5 km spatial grids division	Peak hours: 13:00–14:00, Slack hours: 03:00–04:00

Note: OCH = online car-hailing; POI = Point of interest; GTWR = Geographically Temporal Weighted Regression.

In this context, this study aims to enrich the existing literature by relaxing the predefined linearity assumptions and revealing the complex nonlinear/linear relationship between the built environment and OCH trips at different travel times throughout the day. This study, therefore, seeks to address the following questions: 1) Is OCH demand linearly influenced by environmental factors at different times of the day? 2) Is there a difference in the relationship between the built environment and OCH travel demand at different times of the day? 3) How to better manage OCH? There are, therefore, relative to previous studies, two area in which this study makes an innovative contribution. Firstly, time-scale factors were incorporated into the exploration of the relationship between demand for OCH travel and the built environment. The time periods were divided according to residents’ travel habits and modeled separately to investigate how the impact of built environment attributes varies at different times of the day. The machine learning method random forest partial dependence plot, which does not assume a linear relationship, performed this task, highlighting the need for differentiated urban resource allocation to promote sustainable transport. Secondly, a study unit division method based on the road network was proposed. The division method with a focus on intersections fully takes into consideration the subtle relationship between the road network structure and the production of OCH travel demand.

This paper is structured as follows. We briefly review the research on the relationship between OCH travel demand and the built environment. We then introduce the data and processing methods, followed by a summary of the model methods used in the study. We discuss the results and data analysis before concluding with the main findings and proposals for some relevant policy implications.

Literature Review

Influence Factors and Online Car-Hailing Ridership

The intensification of urbanization will result in several traffic problems as blind urban expansion takes place. It is possible to guide the development of green and sustainable urban transportation by studying the relationship between built environment characteristics and travel behavior. The impact of sociodemographic factors, the built environment, weather conditions, and so forth ( 14 , 26 , 27 ) on travel behavior has been examined in recent research. Specifically, taxi travel demand is related to population density, employment density, and educational background. Among these, income will further affect the taxi travel demand—the higher the income, the higher the demand ( 28 , 29 ). There is a tendency for families with more vehicles to use OCH less frequently ( 30 ). Secondly, rainfall has a greater impact on taxi travel demand than snowfall ( 31 ). Moreover, there are several characteristics of the built environment that influence taxi travel demand, including the density of development, the diversity of land use, and the design of the street network ( 32 ). Based on an analysis of Uber data from 24 different regions in the United States, Sadegh found that Uber demand is positively correlated with land-use entropy and bus station density, and negatively correlated with intersection density and public transportation accessibility ( 15 ). According to Yu and Peng ( 33 ), the impact of the built environment and socio-economic variables on OCH travel varies from city to suburb based on a geographically weighted Poisson regression method. Based on Tian’s geographically weighted regression model, he found that entertainment interest points and residential areas are the main factors influencing OCH travel at night, and that land use has a positive impact as well ( 12 ).

Considering the current state of research, we can conclude that most conclusions are based on global analyses of the influencing factors. The threshold effects of influencing factors are missing in many models that consider capturing local spatial heterogeneity. An important detail of the relationship between the built environment and travel behavior is the threshold effects of influencing factors. Moreover, these more specific relationships can be used to develop sustainable urban planning strategies and effective transportation policies.

Methodologies and Study Unit Division

In previous studies, ordinary least squares regression models, discrete choice models, multiple regression models, as well as other models were used to uncover the relationship between travel behavior and the built environment ( 26 , 34 , 35 ). However, there are several methods for estimating the mean of variables in a study, but they are only able to estimate a global mean, which makes it difficult to capture some partial changes, and they presume that all study variables are static throughout the entire study area. In the absence of partial changes, the model’s reliability will be reduced and our understanding of the spatial variation of OCH demand will be distorted. There have been several recent studies that have used spatial and spatiotemporal models, including geographically weighted Poisson regression, geographically weighted regression models, and semiparametric geographically weighted Poisson regression models ( 12 , 33 , 36 ). This model, however, is based on the assumption that variables are linearly related. In fact, the effect between the built environment and travel behavior may not fully conform to the assumption of a linear relationship between independent and dependent variables ( 37 ). Therefore, in this paper, for more specific recommendations on urban planning management, machine learning methods will be used to capture the nonlinear relationship between the two.

Despite some studies exploring the impact of the built environment on OCH travel demand, most of these studies are limited to using some traditional regression models, which are based on linear assumptions between variables, and lack of data available to complete the exploration of the relationship from different perspectives during the day. Secondly, there have been few studies which have used machine learning to examine the nonlinear relationship between OCH travel demand and the built environment elements at different time periods. As a result, we need to further study this contents to confirm these subtle and detailed relationships, so that we can adopt strategies for sustainable urban planning and effective transportation policies.

In addition to the methodologies, most of the relevant studies require the completion of a division of the study area as the basis of the study (analysis of the relationship between the built environment and influencing factors based on small study units). As can be seen from Table 1, the common division method currently used is the spatial grid division method, which ensures that the area of each study unit is of the same size, but this approach may neglect the impact of the road network environment. For example, the spatial grid may ignore the original distribution of the road network in a small area (some areas have many road network nodes and more intersections, such as downtown areas; some areas have fewer road network nodes, such as suburban areas). To weaken the impact of the study unit division, this study will propose a study area division method based on road network intersections as the focus, which makes our study more complete.

Description of Data

Study Area

A permanent population of 8.741 million makes Nanning the core city of the Beibu Gulf urban agglomeration. In Southwest China, it is a comprehensive transportation hub connecting the sea passages. This paper’s research area is set within the urban area of Nanning, which is framed by Xiangzhu Street, Zhuxi Street, Baisha Street, Shajing Street, Qingchuan Street, and Xiuxiang Street, taking into account the subregions of the city and the distribution of OCH travel demand. And the research areas involve four main urban areas, namely Xingning District, Xixiangtang District, Jiangnan District, and Qingxiu District, with a total of 103 square kilometers, as shown in Figure 1.

Figure 1.

The schematic diagram of research area.

Data Source and Data Processing

The primary objective of this paper is to examine the nonlinear relationship between OCH travel demand and land use in different periods of time. There are three types of data used: OCH travel records, built environment, and road network information. Among them, the OCH travel data is from December 2, 2019 to December 8, 2019 in Nanning, with a total of 1,154,759 trips. A complete OCH travel order includes eight pieces of information: the order number (ID), the starting time, the starting longitude, the starting latitude, the ending time, the ending longitude, the ending latitude, and the distance, as presented in Table 2.

Table 2.

A Sequence of OCH Records

Number	Starting Time	Starting Longitude	Starting Altitude	Ending Time	Ending Longitude	Ending Altitude	Distance
17652721657532	20191202140101	108.38396	22.814472	20191202140634	108.367893	22.816085	2.6
17652723649085	20191202140102	108.29046	22.817994	20191202142106	108.258896	22.776413	7.1
…	…	…	…	…	…	…	…
17654277449087	20191208200718	108.30637	22.83610	20191208202753	108.33776	22.81877	5.8
17654278030888	20191208200922	108.28260	22.832635	20191208203152	108.29843	22.88934	8.4

Note: OCH = online car-hailing.

Since some data cannot be used in this paper, data preprocessing needs to delete records that meet the following criteria: 1) the necessary information is missing; 2) the order number is repeated (the first occurrence of the record is retained); 3) the distance is less than 300 m; 4) the travel time is less than 1 min or more than 2 h; and 5) the latitude and longitude range exceeds the scope of Nanning. After filtering according to these rules, there were 1,136,203 pieces of valid data remaining. The road network used in this study can be obtained from the OpenStreetMap official website. A total of 210,000 POI data were obtained from Gaode’s map data open platform through Python, including the built environment of Nanning. In the field of transportation, POI data is a common source of data that can reflect the functional characteristics of an area to some degree. To avoid errors caused by subsequent data fusion, OCH travel data and POI data must simultaneously be transformed into the same coordinate system. The purpose of this paper is to analyze travel data in different time periods, so we count travel data per hour in this study, as shown in Figure 2.

Figure 2.

24-h travel data.

By taking into account the working hours of companies and the balance of travel during each time period, we divided the 24-h data into five time periods, namely 06:00 to 10:00, 10:00 to 14:00, 14:00 to 18:00, 18:00 to 22:00, and 22:00 to 02:00, to examine the nonlinear relationship between built environments and OCH travel demand during these five time periods. To obtain an overview of the travel situation within the research area, travel data were visualized on ArcGIS software, as shown in Figure 3.

Figure 3.

Spatial distribution of travel data in five periods and travel volume in five periods: (a) 14:00 to 18:00; (b) 18:00 to 22:00; (c) 22:00 to 02:00; (d) 06:00 to 10:00; (e) 10:00 to 14:00; and (f) travel volume in five periods.

As shown in Figure 3, a to e , the majority of OCH travel takes place in Nanning’s main urban area, with a noticeable spatial and temporal heterogeneity. From the perspective of travel volume (Figure 3f), however, there is no obvious peak during five periods, suggesting that OCH is not the main mode of transportation for residents of Nanning, contrary to our previous findings ( 38 , 39 ). Additionally, we must examine the nonlinear relationship between OCH trips and built environment in different periods more carefully by dividing the research area.

Methodology

Research Areas Division

To ensure the rationality of the refinement of the research areas, the research areas were further divided based on the structure of the road networks. The specific steps are as follows:

A complete record containing the intersection name, longitude, and latitude can be obtained by crawling the intersection data in Nanning through Python. Delete the intersection outside the research scope, and coordinates with longitude and latitude to form a finite set of points ${p_{1}, p_{2}, p_{3}, p_{4}, \dots \dots, p_{m}}$ , each point represents an intersection. The details are presented in Table 3.

The intersections (points) automatically construct Delaunay triangulation denoted as ${a_{1}, a_{2}, a_{3}, a_{4}, \dots \dots, a_{n}}$ . Each triangle formed by the intersections (points) will be assigned a unique number, and the intersections (points) constituting each triangle will be recorded.

Find and record the numbers of all triangles adjacent to each point, and find all triangles with one same vertex in the constructed triangulation.

The triangles adjacent to each discrete point are sorted clockwise or counterclockwise. The details are as follows: Take any point $p_{1}$ , find a triangle with $p_{1}$ as vertex, set $A_{1}$ ; the other vertex of triangle $A_{1}$ except $p_{1}$ is set as $p_{2}$ , and the other vertex is $p_{3}$ ; set the next triangle with $p_{1} p_{3}$ edge as triangle $A_{2}$ ; the other vertex of triangle $A_{2}$ is $p_{4}$ , so the next triangle is edged by $p_{1} p_{4}$ . This repeats until you return to the $p_{1} p_{2}$ side.

Calculate the circumcircle center of each triangle in turn and record the center point set as ${c_{1}, c_{2}, c_{3}, c_{4}, \dots \dots, c_{l}}$ .

A Thiessen polygon is constructed based on the intersection of the circumcircle centers of the adjacent triangles surrounding each point. Each Thiessen polygon contains only one intersection, as shown in Figure 4.

Table 3.

The Intersection Data in Nanning

Number	Name	Longitude	Altitude
1	Anhu/Guangyuan intersection	108.354628	22.828318
2	Anhu/Pinghu intersection	108.353035	22.828329
…	…	…	…
1947	Beidaluxiyili/Beidabei intersection	108.298173	22.823974
1948	Beidaluxiyili/Longteng intersection	108.29325	22.828956

Figure 4.

Thiessen polygons in research areas.

Random Forest

Based on the intersection Thiessen polygons, a random forest regression model was used to explore the potential impact of built environment on OCH travel demand in different periods.

The random forest model ( 40 ) is one of the most powerful ensemble learning algorithms for regression problems. It provides good performance in stability and correctness as a bagging ensemble model based on the classification and regression trees (CART) tree and has been widely used. To carry out effective ensemble learning, several weak classifiers are combined to create a strong classifier. To summarize, multiple decision trees are used to form a “forest” and the trees are classified by out-of-bag errors obtained from the voting results of multiple weak classifiers. A training set for out-of-bag estimation is derived by randomly placing back samples when building decision trees to establish a training set. This leads to a result where approximately 30% of the samples, referred to as out-of-bag data, are not used in the random forest generation process. The random forest model provides a way to compensate for overfitting when the data set contains many predictive variables, which can lead to complicated machine learning predictions. Furthermore, this model is capable of handling highly nonlinear variables and interactive variables, without the need for additional dummy variables, as well as ranking the contributions of each variable.

Given the discontinuous and non-monotonic nature of this technical approach, the random forest regression model can capture the detailed relationship between the built environment and the demand for OCH travel. And it can process continuous data and discrete data simultaneously. By using random forests, we can investigate the impact of various types of built environment elements on the OCH travel demand over time. Random forest adopts the bootstrap resampling method, which is composed of a series of decision tree models ${h (X, θ_{k}), k = 1, 2, 3, \dots}$ , where $X$ is an independent variable, $θ$ is a random variable, and $k$ represents the number of decision trees. Multivariate nonlinear regression can be constructed when the predictive variables are numerical. Assuming that the training set is an independent distribution set selected from random vectors $Y and X$ , the mean square generalization error of any numerical prediction $h (X)$ is equal to $E_{X, Y} [Y - h (X)]^{2}$ .

The predicted value of random forest regression is the average value of ${h (X, θ_{k})}$ of $k$ trees, which has the following properties:

E_{X, Y} {[Y - \bar{h} (X, θ_{k})]}^{2} \to E_{X, Y} {[Y - E_{θ} h (X, θ_{k})]}^{2}, k \to \infty

(1)

For all $θ$ , $E (Y) = E_{X} h (X, θ_{k})$ , there is:

P E^{*} (forest) \leq \bar{ρ} P E^{*} (tree)

(2)

where

$\bar{h} (X, θ_{k})$ is the average value of $h (X, θ_{k})$ ,

$P E^{*} (forest)$ is the generalization error of random forest model,

$P E^{*} (tree)$ is the average generalization error of a tree,

$\bar{p}$ is the correlation coefficient of residual $Y - h (X, θ_{k})$ and $Y - h (X, θ_{k}^{'})$ , and

$θ_{k}$ and $θ_{k}^{'}$ are independent of each other.

Various types and specifics of the built environment are the independent variables in the model. An OCH travel demand in the Thiessen polygon is the dependent variable for a complete record, which includes all POI types in that polygon as well as their corresponding data. To prevent overfitting and over complexity, it is necessary to select some super parameters for the random forest model. In five periods, the random forest model parameters were selected using the grid search method, as listed in Table 4.

Table 4.

The Selected Parameters of the Random Forest Model in Five Periods

Time frame	N_estimators	Max_depth	Min_samples	Min_simples_split
06:00–10:00	182	19	2	5
10:00–14:00	71	14	3	9
14:00–18:00	31	17	1	2
18:00–22:00	11	13	1	7
22:00–02:00	194	14	3	2

The marginal effect of the prediction model can be estimated using the partial dependence plot of the machine learning model after the corresponding random forest model has been established. An analysis of the partial dependence plot is useful in determining the linear, monotonic, or complex relationship between objects and features. The partial dependence function of the regression is

{\hat{f}}_{xS} (x_{S}) = E_{xC} [\hat{f} (x_{S}, x_{C})] = \int \hat{f} (x_{S}, x_{C}) d F (x_{C})

(3)

where $x_{S}$ is the sum of the features of drawing partial dependence function and $x_{C}$ is the other features used in the machine learning model $\hat{f}$ . The partial dependence function works by marginalizing the machine learning model on the distribution of features in set $C$ , so that the function shows the relationship between the features of set $S$ we are interested in and the predicted result. By marginalizing other features, we get a function that depends only on features in $S$ , including interactions with other features. The partial function ${\hat{f}}_{x_{S}}$ estimates marginal effects by computing the mean in the training data, also known as the Monte Carlo method, as follows:

\frac{1}{n} \sum_{i = 1}^{n} \hat{f} (x_{S}, x_{C}^{(i)})

(4)

where $x_{C}^{(i)}$ is the actual value of the feature in the data set that we are not interested in and $n$ is the number of instances in the data set.

Result and Discussion

Descriptive Analysis of Online Car-Hailing

(1) Duration. We conduct statistical analysis on the travel time of OCH in the research areas, as shown in Figure 5.

Figure 5.

Distribution map of OCH travel time.

From Figure 5, it is evident that the service time for OCH trips in Nanning is mainly concentrated in 6 to 15 min. From the cumulative curve, it is also evident that 85% of travel users spend less than 24.5 min traveling. Furthermore, very few OCH services last longer than 50 min. Therefore, we can conclude from the above that most travel distances are less than 30 min for OCH, and thus, OCH may not be the best mode of transportation for long distances. When faced with a long-duration trip, people prefer to select other transportation modes instead of OCH given the higher cost ( 41 ). There are still some long-duration trips, however, which may result from the lack of other modes of transport or to avoid the traffic peak hours. Those passengers are willing to choose OCH instead of the metro as their first priority, and their second concern is avoiding the metro’s inconvenience resulting from the interchanges ( 41 ).

(2) Distance. We perform statistical analysis on the travel distance of OCH in the research areas, as shown in Figure 6.

Figure 6.

Distribution map of OCH travel distance.

From the distribution of OCH travel distance in Figure 6 we can see that the single travel distance of OCH in Nanning is mainly within the range of 2 to 7 km. According to the cumulative curve, 85% of users travel within 11.5 kilometers. Additionally, few users have a single service distance over 25 km, reflecting that users seldom use OCH for long-distance travel to some extent ( 23 ), which is perhaps a result of the high cost of long-distance travel.

(3) Space distribution. ArcGIS is used to visualize the spatial distribution of OCH in the research areas, as shown in Figure 7.

Figure 7.

Spatial distribution of OCH travel demand: (a) 06:00 to 10:00, (b) 10:00 to 14:00, (c) 14:00 to 18:00, (d) 18:00 to 22:00, and (e) 22:00 to 02:00.

Hotspots in Figure 7a mainly involve Xingping Community, Dermatology Hospital and Textile Industry School, Jiangnan Xinxingyuan Community, Haotian Garden Community, Qingshan Community, Huayu Luyuan Community, and Guangxi University. We can find hotspots that are mainly composed of residential areas. Comparing Figure 7, a and b, there has been an increase in the points of presence of health care services, such as the People’s Hospital and First Affiliated Hospital of Medical University. Similarly, comparing Figure7, c and a , the study primarily increases leisure sport POIs, such as the first and second affiliated hospitals of Medical University, people’s hospitals, and zoos, while reducing some hotspots, such as residential areas. Compared with Figure 7c, Figure 7d mainly increases some leisure sport POIs, and reduces zoos and some health care POIs. The travel hotspot area of Figure 7e is significantly reduced.

Relationship Analysis

Built Environment Elements

To investigate the relevance of the demand for OCH, three main categories of built environment elements were selected for this study with reference to previous research findings ( 14 ). Specifically, home-work attributes ( 42 , 43 ), land use and diversity attributes ( 24 , 44 ), and traffic facility attributes ( 36 , 45 ). The importance of these three main categories of influencing factors has been confirmed in previous studies. In addition to the residential area, the home-work attribute includes the company. We have singled it out because commuting between the workplace and home is one of the key features of daily urban travel, and this commuting behavior has been identified as one of the key factors influencing OCH travel. There are two variables included in the traffic facility attribute: the subway station and the bus station. Among the parameters of the land use and diversity attribute, schools, government groups, hotels, health care, leisure sports, car services, motorcycle services, cultural services, financial insurance, shopping services, public facilities, scenic spots, catering services, and entropy are included. In total, 14 variables are included in this attribute. The descriptive statistics of the above 18 variables are listed in Table 5.

Table 5.

Descriptive Statistics of Independent Variables

Variables	Variables description	Mean	SD
Home-work attributes
Residential area	Number of dwelling, dormitory, and other facilities in one Thiessen polygon	6.668	9.435
Company	Number of company in one Thiessen polygon	4.948	10.976
Traffic facility attributes
Subway station	Number of subway station in one Thiessen polygon	0.293	1.280
Bus station	Number of bus station in one Thiessen polygon	0.834	1.933
Land use and diversity
School	Number of School in one Thiessen polygon	3.121	6.062
Government group	Number of administrative, traffic, judicial, and other departments in one Thiessen polygon	6.911	10.366
Hotel	Number of hotel in one Thiessen polygon	1.904	3.372
Health care	Number of hospitals, clinics, emergency centers, and other facilities in one Thiessen polygon	7.095	12.050
leisure sport	Number of swimming pools, game room, sports venues, entertainment venues, and other facilities in one Thiessen polygon	3.864	7.649
Car service	Number of car maintenance, car rental, and other facilities in one Thiessen polygon	3.199	15.665
Motorcycle service	Number of motorcycle sales, motorcycle maintenance, and other facilities in one Thiessen polygon	0.158	0.576
Cultural service	Number of science museums, and other facilities in one Thiessen polygon	2.357	5.057
Financial insurance	Number of banks, insurance agents, and other facilities in one Thiessen polygon	2.034	4.321
Shopping service	Number of shopping malls, shops, and other facilities in one Thiessen polygon	32.385	66.979
Public facility	Number of facilities such as public toilets and emergency shelters in one Thiessen polygon	0.856	2.218
Scenic spot	Number of park plaza, city square, scenic spot, and other facilities in one Thiessen polygon	0.374	1.049
Catering service	Number of restaurants, bars, cafes, and other facilities in one Thiessen polygon	10.777	18.155
Entropy	POIs spatial entropy in one Thiessen polygon	0.820	0.110

Note: SD = standard deviation; POI = Point of interest.

Importance of Independent Variables

The R² values of the random forest model in the five periods of 06:00 to 10:00, 10:00 to 14:00, 14:00 to 18:00, 18:00 to 22:00 and 22:00 to 02:00 are 0.725, 0.776, 0.736, 0.780, and 0.703, respectively. These values indicate that the random forest model in this paper exhibits good prediction performance and can effectively capture the complex relationship between the built environment and OCH travel demand. The contribution and ranking of each variable in the prediction are presented in Table 6.

Table 6.

Importance Analysis of Independent Variables in Different Periods

Variable	06:00-10:00(P1)		10:00-14:00(P2)		14:00-18:00(P3)		18:00-22:00(P4)		22:00-02:00(P5)
Variable	Rank	Proportion	Rank	Proportion	Rank	Proportion	Rank	Proportion	Rank	Proportion
Home-work attributes
Residential area	1	33.21%	8	2.31%	7	4.38%	10	2.24%	6	3.50%
Company	8	1.42%	13	1.14%	12	1.58%	16	0.23%	15	0.86%
Traffic facility attributes
Subway station	17	0.23%	15	0.60%	14	1.33%	5	4.18%	5	4.34%
Bus station	4	12.79%	3	13.84%	4	11.01%	3	12.21%	11	2.00%
Land use and diversity
School	10	0.85%	14	0.77%	15	1.14%	6	3.90%	16	0.54%
Government group	13	0.74%	9	2.26%	10	1.75%	14	1.16%	13	1.03%
Hotel	9	1.07%	5	3.48%	13	1.47%	11	2.07%	12	1.34%
Health care	3	15.41%	1	35.34%	3	12.00%	7	2.82%	8	2.55%
Leisure sport	5	4.89%	4	11.77%	2	18.16%	1	39.58%	2	27.31%
Car service	14	0.73%	16	0.51%	16	1.07%	15	0.39%	14	0.91%
Motorcycle service	18	0.17%	18	0.05%	18	0.05%	17	0.06%	18	0.06%
Cultural service	2	22.13%	7	2.45%	1	25.50%	4	8.80%	3	12.29%
Financial insurance	7	1.83%	10	2.00%	6	5.06%	9	2.37%	7	3.18%
Shopping service	12	0.80%	12	1.71%	9	3.52%	8	2.52%	4	5.88%
Public facility	15	0.36%	6	3.29%	8	3.61%	12	1.85%	10	2.15%
Scenic spot	16	0.32%	17	0.29%	17	0.56%	18	0.04%	17	0.16%
Catering service	6	2.21%	2	16.40%	5	6.20%	2	14.04%	1	29.68%
Entropy	11	0.84%	11	1.79%	11	1.61%	13	1.54%	9	2.22%

Table 6 indicates the impact of different built environment elements of OCH travel demand in different periods. The sum of relative importance of each independent variable is 100%. As a result of our research, different POIs in different periods contribute differently to the use of OCH, which is in contrast to our previous research findings. Previous studies suggest that the company and residential area play an important role in the use of OCH ( 21 ). According to the results, however, OCH use in residential areas is only significant for the period P1, with a contribution rate of 33.2%, while it is less than 5% for the other four periods. The company, which was valued in the past, contributed less than 5% in each period. This may be a result of Nanning residents often using electric bikes as travel tools. According to the results of the P1 analysis, 13 land use characteristics are responsible for about 50% of the prediction ability. Over 80% of the contributions were made in the remaining four periods. It is possible that this difference is a result of residents traveling for relatively single purposes during P1. There is a greater variety of travel purposes during the other four periods. For example, when traveling during other periods, it is more likely that travelers engage in varied activities such as catering, sports, and entertainment. There are four relatively important factors involved in land use: cultural services, health care, leisure sports, and catering services. In contrast to previous research conclusions, shopping services do not figure among them ( 14 ).

Furthermore, traffic facility attributes have a significant impact on OCH demand. However, five different periods occupy different positions in the model. According to the bus station, the contribution rates during P1, P2, P3, P4, P5 are 12.79%, 13.84%, 11.01%, 12.21%, and 2.00%, respectively. There is a contribution rate of less than 5% for the subway station over the five periods. Compared with subway stations, the impact of the bus station on OCH travel requirements is more obvious, indicating that people are more sensitive to existing bus services than subway services. Generally speaking, there is a greater contribution to OCH travel demand from the bus station than from the subway station. Over the course of the four periods, P1, P2, P3, and P4, the bus station has contributed significantly to OCH travel. There may be a reason for this, such as the inaccessibility of public transport which leads to the use of OCH, which is considered to be an important connection to public transportation ( 12 , 21 ). It is also possible that residents who originally planned to travel by public transportation gave up after waiting for an extended period of time and preferred OCH.

Nonlinear Impacts of Independent Variables

Based on the partial dependence plot generated by the random forest model, the nonlinear impacts of independent variables are identified. For analysis of the impact of built environment elements on OCH travel over time, all other independent variables have been set to constants. The partial dependence plot of home-work attribute is shown in Figures 8 and 9. The partial dependence plot of land use and diversity is shown in Figure 10, and traffic facility attribute is shown in Figure 11.

Figure 8.

Nonlinear impacts of residential area in different periods on OCH travel demand.

Figure 9.

Nonlinear impacts of company in different periods on OCH travel demand.

Figure 10.

Nonlinear impacts of land use and diversity in different periods on OCH travel demand.

Figure 11.

Nonlinear impacts of traffic facility attribute in different periods on OCH travel demand.

It can be seen from Figure 8 that there is a positive correlation between residential areas and OCH trips, which is consistent with previous research ( 12 ). However, it is readily apparent that for the period P1, when the number of residential areas in the fixed area increases from 0 to 12, there will be a significant increase (about 50 times) in OCH trip needs. Travel demand change will slow when there are more than 12 residential areas in a fixed area. In addition, during the other four periods, we can observe that when the residential areas are between 0 and 5, an increase in residential areas does not result in a significant increase in OCH trips. There is a significant impact elasticity of residential areas on OCH during the period of P1, about twice that of other periods.

As can be seen from Figure 9, there is a significant impact of increasing the number of company POIs on the use of OCH travel demand in P2 and P3, but not in P1, P4, or P5. There is a noticeable elastic relationship between the number of company POIs and OCH travel demand during the period of P2 and P3. There is a possibility that this is because most workers choose OCH to go out during their work hours. OCH can provide door-to-door service to meet workers’ timeliness requirements and enable them to complete their tasks before the end of the working day.

The 15 subgraphs in Figure 10 are the relationship between the three most important land-use variables and the corresponding OCH travel demand at different periods. According to Figure 10, the relationship between land-use variables and the demand for OCH travel is nonlinear during different periods. Additionally, all land-use variables have a positive impact on OCH travel in general, but the effects of each period may differ in magnitude. During the P1 period, the growth of cultural services from 0 to 15 had little effect on the demand for OCH. It has been shown that as the number of cultural services increases from 15 to about 20, the number of OCH travel services increases as well, reaching a steady value of about 450 times when there are about 20 cultural services. Cultural services mainly involve some training institutions. The results show that only a few training institutions fail to attract enough OCH trips. It should be noted, however, that when a region has enough training institutions and establishes its trading area in a city, it will have a profound impact on the need for OCH in that area. It is noteworthy that health care service is an important factor influencing the period of P2. There is little impact on the OCH demand when the number of health care services is less than five. In contrast, the number of OCH trips has increased by about 250 times when the number increased from 5 to 25. This may reflect most patients preferring to leave home early in the morning for treatment at hospital, so the end of treatment falls mostly during this time frame. As a consequence of physical discomfort, they prefer a more comfortable and quieter means of transportation, which significantly impacts the travel needs of OCHs in the health care area.

When the number of leisure sports services is between 0 and 6, it has little impact on demand during the period of P3. When the number increased from 6 to 16, the demand for OCH trips shows a sharp growth trend. Health care is also a major influencing factor during this period. The trend presented is consistent with period P2. The impact of health care service on OCH demand is a relatively new finding. There are three major land-use variables that affect OCH travel during the P4 and P5 periods: leisure sports, catering services, and cultural services. These two periods show a certain similarity in land-use variables related to OCH travel demand. It is possible that these two periods are both rest times for residents following a long day of work, so travel activities during these two periods are consistent. To effectively affect OCH needs during P4, the number of leisure sports services should reach at least 13. Meanwhile, the demand for OCH travel is not always affected by the catering service. There is a greater impact of catering during the P4 and P5 periods, whereas it is not a major factor during the daytime.

The impacts of traffic facilities in different periods on the use of OCH are shown in Figure 11. It is almost impossible to discern any impact from the number of bus stations on the amount of OCH travel demand during P5. In contrast, there is a significant positive elastic relationship between the number of bus stations and demand during the period P1∼P4. From the perspective of the five periods, the impact of bus stations on OCH demand reaches saturation around 15, indicating that bus stations have a limited impact on OCH demand. Generally speaking, the subway station has no negative impact on the demand for OCH, but it does not have a significantly positive impact at all times. For example, in the periods P1 and P2, the subway stations increased from 1 to 4, and the increase in the demand for OCH is about 20. In the three periods P3, P4, and P5, the subway system has a significant positive impact on OCH demand. The subway stations increased from 1 to 4, and the increase in OCH demand reaches about 50∼80. This indicates that the subway system has little effect on the demand during P1 and P2, while it has a significant impact on the demand for OCH during P4 and P5. This may be because most buses’ last bus time is about 22:00. Given the shutdown of bus lines during P4 and P5 at night, there is limited travel mode available in the last kilometer, and more individuals use the OCH to connect to the subway system ( 24 , 46 ). Furthermore, it indicates that there are few users who use OCH to connect to the subway in Nanning in the morning, and the subway system has a limited impact on OCH use. Based on this result, it appears that OCH and the subway system are capable of cooperating on service timing.

Conclusions and Policy Implications

From the micro perspective, this study examined the relationship between OCH travel and the built environment in different periods of time. Specifically, a comprehensive analysis of the balance between working hours in urban areas and travel during each period was conducted by dividing the research period into five categories (6:00–10:00, 10:00–14:00, 14:00–18:00, 18:00–22:00, 22:00–02:00). Furthermore, to obtain the correlation analysis unit, the research area division method was proposed based on intersection, and it was then combined with a random forest partial dependence plot to explore the impact of built environmental factors on OCH use over five typical periods. As a result of this study, we will further examine the mechanism by which built environment elements influence OCH demand in different periods, as well as discover the discontinuous and non-monotonic relationship between the two in different periods. It provides results useful for follow-up research, as follows:

In different periods, OCH travel demand may differ as a result of heterogeneity in the built environment. According to some theories, this phenomenon is potentially caused by there being an inverse correlation between the purpose of urban residents’ travel and the time of their travel, which results in an uneven distribution of their travel demand over time and space. First of all, results show that residential areas, which have always been highly valued, have a significant influence on OCH travel demand during the period P1, with a contribution rate of 33.2%, while the predicted contribution rates for the remaining four periods are less than 5%. To prevent excessive invalid no-load patrolling, a decline in order receiving rates, and meaningless exhaust emissions, the OCH and taxi management department must control the vehicles patrolling near the residential area in the other four periods. In contrast, the contribution rate of the company to the OCH requirements in Nanning has been below 5% in five periods, which is significantly different from previous studies. Secondly, generally speaking, there is a greater contribution to OCH travel demand from the bus station than from the subway station. Specifically, there was a significant contribution rate of more than 10% for the bus station in the four periods P1 through P4, and the contribution rate in the period P5 was less than 5% for the subway station. The results confirm that passengers who planned to travel by bus during P1 to P5 may be disappointed with the bus service given long waiting times or crowded bus passenger flow, and instead turn to OCH. Furthermore, it demonstrates the potential to connect OCH and buses in Nanning.

In addition, the nonlinear relationship between land-use variables and demand for OCH trips varies at different times of the day and has different threshold effects. The contribution of land-use attributes to the demand varies between time periods, and the main influencing variables differ. Specifically, during P1 only about 50% of OCH travel demand was attributed to land use and diversity attributes. During the other four periods, the contribution exceeded 80%. Among these 13 attribute variables, cultural service, health care service, leisure sport service, and catering service are four relatively important factors, which apply to five different time periods. A major contribution comes from cultural services in P1 and P3. It is important to note, however, that in cases where there are enough training institutions to form a separate trading area within a city, this may have a profound impact on the OCH travel demand. In contrast, the contribution of health care services is the most important of the land-use variables in the P2 period, which is a relatively new finding. But it also has a certain threshold effect, when the number of health care services in the study unit is less than 5, the effect on OCH demand is small. When the number increased from 5 to 25, the number of trips made by the OCH increased by a factor of about 250. For the P4 and P5 periods, the main land-use variables are also more similar because of some similarities in the purpose of the trips. Leisure sports services and catering services make up the majority of the contributions in periods P4 and P5. Catering service, which was the focus of previous studies ( 12 , 24 ), did not always greatly affect the demand for OCH, such as P1 and P3 periods.

This paper provides effective guidance for the scheduling of urban occupational health and safety services. To improve cruising efficiency and reduce waiting times for residents, traffic guidance and scheduling are effective methods to recommend routes that are in high demand to OCH drivers. It is important to note that these findings are of great importance to other cities around the world that have similar travel habits and structures to Nanning. Secondly, to reduce the interference between static traffic and dynamic traffic flow for the main contribution POI types in different periods, the government may adopt some necessary traffic organization and control measures, such as optimizing the parking area at the OCHs to reduce the interference caused by static traffic. Thirdly, when there are high traffic levels in some areas, it would be helpful to check the quality of the operation of the bus lines, to avoid residents becoming dissatisfied with public transportation services and turning to OCH travel, resulting in higher levels of traffic. Finally, since the built environment and OCH travel display heterogeneous and nonlinear relationships in different periods, relevant policies can be developed according to the scale characteristics of urban infrastructure. Meanwhile, the balance between supply and demand of transportation facilities in high-demand areas of the built environment is adjusted to promote sustainable development of public transportation.

Although our work is a comprehensive study that explores the nonlinear effects of the built environment on the demand for OCH travel on a time scale, there are still some limitations that require further research. First, the independent variables studied in this paper do not have access to data related to sociodemographic and service factors (ease of use) given data limitations. These factors may be co-linear with some land-use types and thus have an impact on the demand. Secondly, this study is limited to data on pick-up points for OCH trips, while no relevant research has been conducted on the distribution of drop-off points, and a subsequent comparison of pick-up and drop-off points could be interesting. Thirdly, to capture the nonlinear relationship between the built environment and the demand for OCH, this study focuses more on interpreting the machine learning results to provide some policy guidance to relevant departments, which is why it uses only the random forest model. To understand why different machine-learning models differ, it might be helpful to pay attention to the differences between some of them. Then, considering that in many countries car-hailing is not integrated with other transport modes, our approach can be extended to ride-hailing more generally. So exploring the relationship between the built environment and on-demand buses and minibuses could be a good direction to take in the future. Last but not least, given space constraints, we have not compared weekdays and weekends, and it will be interesting to compare and analyze the model performance on weekdays and weekends in future studies to find the differences.

Footnotes

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: X. Zhou, Y. Ji; data collection: X. Zhou, M. Lv, Y. Ji; analysis and interpretation of results: X. Zhou, M. Lv; draft manuscript preparation: X. Zhou, M. Lv, Y. Ji. All authors reviewed the results and approved the final version of the manuscript.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was funded by the National Key R&D Program of China (2018YFB1600900).

Data Availability Statements

Given the nature of this research, the data provider requested that the data be kept confidential, so supporting data is not available.

References

Orozco-Fontalvo

Llerena

Cantillo

Dockless Electric Scooters: A Review of a Growing Micromobility Mode. International Journal of Sustainable Transportation, Vol. 17, No. 4, 2023, 406–422.

Dey

B. K.

Tirtha

S. D.

Eluru

Konduri

K. C.

Transformation of Ridehailing in New York City: A Quantitative Assessment. Transportation Research Part C: Emerging Technologies, Vol. 129, 2021, p. 103235.

Chen

Zhang

Cai

Sui

Long

Song

Shibasaki

GPS Data in Urban Online Ride-Hailing: A Simulation Method to Evaluate Impact of User Scale on Emission Performance of System. Journal of Cleaner Production, Vol. 287, 2021, p. 125567.

Yangshixinwen. Data on ride-hailing industry in December 2020. https://baijiahao.baidu.com/s?id=1688504881935515118&wfr=spider&for=pc. Accessed May 11, 2022.

Acheampong

R. A.

Siiba

Okyere

D. K.

Tuffour

J. P.

Mobility-on-Demand: An Empirical Study of Internet-Based Ride-Hailing Adoption Factors, Travel Characteristics and Mode Substitution Effects. Transportation Research Part C: Emerging Technologies, Vol. 115, 2020, p. 102638.

Gan

Yang

Feng

Timmermans

H. J. P.

Examining the Relationship Between Built Environment and Metro Ridership at Station-to-Station Level. Transportation Research Part D: Transport and Environment, Vol. 82, 2020, p. 102332.

Liu

Wang

Xie

Evaluating the Effects of Public Transport Fare Policy Change Together with Built and Non-Built Environment Features on Ridership: The Case in South East Queensland, Australia. Transport Policy, Vol. 76, 2019, pp. 78–89.

Scott

D. M.

Ciuro

What Factors Influence Bike Share Ridership? An Investigation of Hamilton, Ontario’s Bike Share Hubs. Travel Behaviour and Society, Vol. 16, 2019, pp. 50–58.

Zhao

Shang

Yang

How Does the Urban Built Environment Affect Online Car-Hailing Ridership Intensity Among Different Scales?

International Journal of Environmental Research and Public Health, Vol. 19, No. 9, 2022, p. 5325.

10.

Wang

Hua

Wei

Exploring Taxi Demand Distribution of Comprehensive Land-Use Patterns Using Online Car-Hailing Data and Points of Interest in Chengdu, China. Transportation Research Record: Journal of the Transportation Research Board, 2021. 2675: 1268–1286.

11.

Wang

Mao

Ren

Chen

Gao

Influencing Factor Analysis and Demand Forecasting of Intercity Online Car-Hailing Travel. Sustainability, Vol. 13, No. 13, 2021, p. 7419.

12.

Jing

Sun

Yan

Revealing the Varying Impact of Urban Built Environment on Online Car-Hailing Travel in Spatio-Temporal Dimension: An Exploratory Analysis in Chengdu, China. Sustainability, Vol. 11, No. 5, 2019, p. 1336.

13.

Zhang

Chen

Tang

Analysis on Spatiotemporal Urban Mobility Based on Online Car-Hailing Data. Journal of Transport Geography, Vol. 82, 2020, p. 102568.

14.

Wang

Chen

Shao

How Built Environment Impacts Online Car-Hailing Ridership. Transportation Research Record: Journal of the Transportation Research Board, 2020. 2674: 745–760.

15.

Sabouri

Park

Smith

Tian

Ewing

Exploring the Influence of Built Environment on Uber Demand. Transportation Research Part D: Transport and Environment, Vol. 81, 2020, p. 102296.

16.

Wang

Zhao

Liu

Chen

Wang

Zhang

Non-Linear Effects of the Built Environment and Social Environment on Bus Use Among Older Adults in China: An Application of the XGBoost Model. International Journal of Environmental Research and Public Health, Vol. 18, No. 18, 2021, p. 9592.

17.

Yang

Zhang

Helbich

Ettema

Chen

Examining Non-Linear Associations Between Built Environments Around Workplace and Adults’ Walking Behaviour in Shanghai, China. Transportation Research Part A: Policy and Practice, Vol. 155, 2022, pp. 234–246.

18.

Chen

Wang

Zhang

Discovering the Spatio-Temporal Impacts of Built Environment on Metro Ridership Using Smart Card Data. Cities, Vol. 95, 2019, p. 102359.

19.

Wang

Fan

Liu

Revealing the Influence Mechanism of Urban Built Environment on Online Car-Hailing Travel Considering Orientation Entropy of Street Network. Discrete Dynamics in Nature and Society, Vol. 2022, 2022, pp. 1–15.

20.

Kwan

Yang

Liu

Understanding the Spatiotemporal Variation of High-Efficiency Ride-Hailing Orders: A Case Study of Haikou, China. ISPRS International Journal of Geo-Information, Vol. 11, No. 1, 2022, p. 42.

21.

Zheng

Zhang

Rong

Qin

Understanding the Impact of the Built Environment on Ride-Hailing from a Spatio-Temporal Perspective: A Fine-Scale Empirical Study from China. Cities, Vol. 126, 2022, p. 103706.

22.

Gehrke

S. R.

Huff

M. P.

Spatial Equity Implications and Neighborhood Indicators of Ridehailing Trip Frequency and Vehicle Miles Traveled in the Phoenix Metro Region. Transportation, 2022, pp. 1–25. https://doi.org/10.1007/s11116-022-10327-3.

23.

Dean

M. D.

Kockelman

K. M.

Spatial Variation in Shared Ride-Hail Trip Demand and Factors Contributing to Sharing: Lessons from Chicago. Journal of Transport Geography, Vol. 91, 2021, p. 102944.

24.

Wang

Noland

R. B.

Variation in Ride-Hailing Trips in Chengdu, China. Transportation Research Part D: Transport and Environment, Vol. 90, 2021, p. 102596.

25.

Zhang

Huang

Zhu

Spatiotemporal Varying Effects of Built Environment on Taxi and Ride-Hailing Ridership in New York City. ISPRS International Journal of Geo-Information, Vol. 9, No. 8, 2020, p. 475.

26.

Liu

Fang

Jing

How Urban Land Use Influences Commuting Flows in Wuhan, Central China: A Mobile Phone Signaling Data Perspective. Sustainable Cities and Society, Vol. 53, 2020, p. 101914.

27.

Guidon

Becker

Dediu

Axhausen

K. W.

Electric Bicycle-Sharing: A New Competitor in the Urban Transportation Market? An Empirical Analysis of Transaction Data. Transportation Research Record: Journal of the Transportation Research Board, 2019. 2673: 15–26.

28.

Qian

Ukkusuri

S. V.

Spatial Variation of the Urban Taxi Ridership Using GPS Data. Applied Geography, Vol. 59, 2015, pp. 31–42.

29.

Yang

Gonzales

E. J.

Modeling Taxi Trip Demand by Time of Day in New York City. Transportation Research Record: Journal of the Transportation Research Board, 2014. 2429: 110–120.

30.

Correa

Xie

Ozbay

Exploring the Taxi and Uber Demand in New York City: An Empirical Analysis and Spatial Modeling. SSRN, 2017. https://doi.org/10.2139/ssrn.4229042

31.

Kamga

Yazici

M. A.

Singhal

Hailing in the Rain: Temporal and Weather-Related Variations in Taxi Ridership and Taxi Demand-Supply Equilibrium. 2013.

32.

Ewing

Cervero

Travel and the Built Environment: A Meta-Analysis. Journal of the American Planning Association, Vol. 76, No. 3, 2010, pp. 265–294.

33.

Peng

Exploring the Spatial Variation of Ridesourcing Demand and its Relationship to Built Environment and Socioeconomic Factors with the Geographically Weighted Poisson Regression. Journal of Transport Geography, Vol. 75, 2019, pp. 147–163.

34.

Jun

Choi

Jeong

Kwon

Kim

Land Use Characteristics of Subway Catchment Areas and Their Influence on Subway Ridership in Seoul. Journal of Transport Geography, Vol. 48, 2015, pp. 30–40.

35.

Kim

Ahn

Choi

Kim

Sustainable Mobility: Longitudinal Analysis of Built Environment on Transit Ridership. Sustainability, Vol. 8, No. 10, 2016, p. 1016.

36.

Chen

Feng

Ding

Yao

Examining the Spatial-Temporal Relationship Between Urban Built Environment and Taxi Ridership: Results of a Semi-Parametric GWPR Model. Journal of Transport Geography, Vol. 96, 2021, p. 103172.

37.

Van Wee

Handy

Key Research Themes on Urban Space, Scale, and Sustainable Urban Mobility. International Journal of Sustainable Transportation, Vol. 10, No. 1, 2016, pp. 18–24.

38.

Luo

Research on Oversight Mechanisms of Online Car-Hailing Platforms Based on Three-Party Game. Journal of Human Resource and Sustainability Studies, Vol. 7, No. 1, 2019, pp. 24–36.

39.

Tang

Long

Delineating the Regional Economic Geography of China by the Approach of Community Detection. Sustainability, Vol. 11, No. 21, 2019, p. 6053.

40.

Breiman

Random Forests. Machine Learning, Vol. 45, No. 1, 2001, pp. 5–32.

41.

Fan

Exploring Spatiotemporal Patterns of Long-Distance Taxi Rides in Shanghai. ISPRS International Journal of Geo-Information, Vol. 6, No. 11, 2017, p. 339.

42.

Osorio-Arjona

García-Palomares

J. C.

Social Media and Urban Mobility: Using Twitter to Calculate Home-Work Travel Matrices. Cities, Vol. 89, 2019, pp. 268–280.

43.

Liu

Ding

Chen

A Panel Analysis of the Effect of the Urban Environment on the Spatiotemporal Pattern of Taxi Demand. Travel Behaviour and Society, Vol. 18, 2020, pp. 29–36.

44.

Jain

Tiwari

Measuring Density and Diversity to Model Travel Behavior in Indian Context. Land Use Policy, Vol. 88, 2019, p. 104199.

45.

Guan

Wang

Influences of the Built Environment on Travel: A Household-Based Perspective. Transportation Research Part A: Policy and Practice, Vol. 130, 2019, pp. 710–724.

46.

Yang

Franz

M. L.

Zhu

Mahmoudi

Nasri

Zhang

Analysis of Washington, DC Taxi Demand Using GPS and Land-Use Data. Journal of Transport Geography, Vol. 66, 2018, pp. 35–44.