Nonlinear Effects of Community Built Environment on Car Usage Behavior: A Machine Learning Approach

Abstract

This study aims to guide the community life circle to create a green, travel-supportive built environment. It quantitatively analyzes the variations in car usage behavior based on the level of the built environment of the community and objectively reflects the car usage behavior based on the parking space utilization rate (PSUR). Ordinary least squares (OLS) and gradient boosting decision tree (GBDT) models were developed to describe the impact of the built environment on this utilization rate. An empirical analysis of the model was also conducted using the multisource, heterogeneous parking data of commercial parking facilities in the main urban area of Chongqing, China; the data include records of parking survey, points of interest, and road networks. The results showed that the GBDT model had a better fitting degree than the OLS model considering nonlinear effects. In terms of the contribution of community-built environment variables, distance to business center (14.30%), population density (14.20%), and land use mix (12.60%) considerably affect the PSUR, indicating that these variables have an important influence on the use of private cars. All built environment variables have nonlinear relationships, and the threshold effects reflect a complex relationship between the built environment and car usage behavior. This study provides refined suggestions for the spatial design and transformation of the community life circle.

1. Introduction

With the acceleration of China’s urbanization and the continuous increase in car ownership, traffic congestion in big cities has become the norm [1]. The key to easing traffic congestion is reducing car use. China is also actively promoting community life circle planning [2] to create a people-oriented living environment by configuring various service elements in the community, such that more trips in the community can be controlled to reduce the use of cars and control greenhouse gas emissions. In this study, analyzing the relationship between the built environment of the community and vehicle use behavior is extremely important to alleviate road traffic congestion and reduce greenhouse gas emissions through urban planning.

Many scholars have systematically investigated the relationship between built environments and travel behavior in which car usage behavior is an important variable. Car usage behavior includes owning a car, using a car for travel, and the frequency of using a car. However, in China, a car is not simply a vehicle for transportation; it also reflects the expansion of the spiritual attributes of ordinary people, symbolizing improvement in social status. Moreover, restricting car ownership seems unrealistic. As of 2021, the number of cars in China has reached 384 million. Therefore, the regulation of car use is particularly important. Existing studies typically express car usage behavior through the choice of travel mode or number of vehicle miles traveled, and questionnaire survey is the main method for obtaining data. With the maturity of information technology, characterizing travel behavior based on the objective data generated during travel (such as mobile phone signal and bus card swiping data) is a current research focus. These data can more precisely reflect individual behaviors; however, the analysis of car usage behavior based on these is generally difficult. Accordingly, finding a new data source to objectively determine the car usage behavior of travelers is urgently required.

Car usage behavior is associated with parking and can therefore be expressed through the parking space utilization rate (PSUR). To the best knowledge of the authors, no research has ever been conducted on how to determine travel behavior based on the characteristics of parking behavior. Previous investigations have shown that many residential parking allocation indicators in the city proper of Chongqing in the most recent 10 years have satisfied the standard of one parking space for each household. Moreover, the parking space sales rate exceeds 95%, thus ensuring that the PSUR of parking facilities in a residential area can accurately reflect car usage behavior. Hence, the foregoing provides an excellent basis for our research.

By utilizing the data from the main urban area of Chongqing, this study reflects the aggregate behavior of car owners based on the PSUR. Further, it explores the nonlinear effect of community-built environments on car usage behavior via the machine learning method. The study is anticipated to be used as a scientific reference for the spatial design and transformation of community life circles as well as optimizing the allocation of urban spatial resources.

The organization of this paper is as follows: Section 2 reviews the literature on the connection between built environments and travel behavior. Section 3 introduces the variables selected for the study. Section 4 presents the data. Section 5 introduces the modeling approach. Finally, Section 6 and Section 7 review the key results and implications for planning practice.

2. Literature Review

Since the 1980s, the idea of optimizing the built environment to guide individual travel demand has been considerably promoted and applied to the historical course of urban renewal and suburban development in developed countries. Moreover, some theories, such as “new urbanism,” “compact city,” and “smart growth,” were proposed to reduce car use by optimizing spatial organization. Cervero [3] summarized the built environment as three-dimensional (3D), i.e., composed of density, diversity, and design. In the foregoing, residential, commercial facility, road network, and social (such as population and employment) densities are included. Diversity refers to the variety of alternative land and transportation modes. Design pertains to the design of public spaces, such as street green coverage and traffic facilities (e.g., road network grading structure, street width, and sidewalk width). Based on the 3D elements, Ewing [4] added two indicators, i.e., destination accessibility and distance to transit, to form a five-dimensional (5D) description of the built environment; currently, this is the mainstream measurement dimension of built environments. Subsequently, Ewing [5] added demand management (including parking supply and cost) to the 5D elements. Some scholars have also proposed the inclusion of the distance to the city center based on the 5D composition. Handy [6] viewed built environment as consisting of three parts: land use, urban design, and transportation system.

The results of research on the built environment and travel behavior, especially the impact of the built environment on travel behavior at the beginning and end of each trip, have been empirically tested in different regions [7,8]. However, empirical studies in different countries and regions have come to different conclusions on how the built environment affects individual travel behavior and the extent to which it influences the travel decision-making behavior of residents. The study of Ma considered Beijing as an example and found that the high-density and land use mix around residences can reduce the distance traveled by residents and the dependence of travelers on cars [9]. The empirical results of the study by Limanond in the United States show that the built environment of residences has an inconsiderable impact on the distance traveled by and travel mode of residents [10]. Yang believed that these differences may be caused by different travel purposes [11]. However, for the same travel purpose, Yu found that the relationship between the built environment and travel behavior also differs among countries [9]. In general, the impact of the built environment on travel behavior has a “universality of effect and difference of conclusion,” which may be caused by factors such as geography, culture, social environment, group difference, and travel purpose difference [12]. Therefore, a study on built environment and travel behavior requires empirical research in different countries and regions in order to enrich the conclusions in this field. In addition, Yang proposed that the difference in this relationship may be caused by the simple assumption of the linear relationship between the built environment and travel behavior [13].

The 5D measurement system (composed of density, diversity, design, destination accessibility, and distance to transit) proposed by Ewing [5] is the most widely used measurement index of built environments.

Density is an important index to regulate the scale of regional construction, in which population density is the core control index. It is also the key factor to determine the coverage index of service elements in China’s community life circle planning. Handy’s empirical study in Florida shows that high population density is usually related to low-income families, limited parking supply, and more intensive bus services. Hence, high population density inhibits the ownership and use of cars [5]. The empirical research of Sun in Shanghai shows that population density has a negative impact on car use [14]. However, Ding [15] points out that Chinese cities with extremely dense populations may not be able to reduce the use of cars as expected; thus, more empirical studies on these cities are necessary.

The representative indicator of diversity is land use mix [4]. A low value indicates a single-use environment and a high value denotes a more diverse land use. Empirical studies in many Western countries show that, when the degree of land use mix is high, activities that have been previously performed in scattered locations are held in a small space. Consequently, the use of cars is limited by the reduction in travel distance [6,16]. However, low-density urban sprawl is generally behind urban motorization in the West. In contrast, China already had a high-density and mixed urban form before motorization. Therefore, the empirical analysis of land use mix in China may present a different law.

Design mainly refers to the road network characteristics in an area. China actively promotes the design concept of “dense road network and small blocks” to create a people-oriented travel environment and reduce the use of cars. Therefore, the exploration of the impact mechanism of road density and intersection density on car usage behavior in the design dimension is particularly important. The empirical research of Badland in New Zealand shows that road and intersection densities are negatively correlated with the probability of car use. Higher road density means better connectivity and lower mobility, which can reduce car use [17].

Destination accessibility is closely related to the distance from the city center, which is the key index of urban structure. In planning practices, the distance between the residences and city center is typically reduced by implementing a multicenter urban layout. Buliung used the data from Portland to show that the closer the residence to the central business district, the shorter the daily travel distance. Consequently, residents have a greater tendency to give up using cars for travel [18]. Næss also reached a similar conclusion in the empirical study of Hangzhou, China [19].

The representative indicators of distance to transit are bus stop and rail transit station densities. China is promoting the large-scale optimization of public transport systems. Empirical studies in China and Western countries show that increasing the density of bus stops and rail transit stations around residential areas can effectively curb the demand for and use of cars [20,21].

However, most of the above studies assume a linear relationship between the built environment and travel behavior. With further research, scholars realized the defects of a linear hypothesis. Accordingly, some studies began to explore the nonlinear effects using machine learning in order to improve space design. For example, Tao [21] built a gradient-boosting decision tree (GBDT) model to reveal the nonlinear relationship between the built environment and walking distance. The GBDT model compared with a linear model has a better fitting degree for data. Based on the GBDT model, Zhang [22] explored the nonlinear effect of the built environment on family car ownership, and Tu [23] examined the nonlinear impact of the built environment on carpooling rates.

Previous studies have shown that the built environment has a significant impact on travel behavior, and the nonlinear effect must be further explored. The current research is necessary in view of the following:

(1) The traditional research on travel behavior requires the collection of data on the socioeconomic and travel characteristics of individuals through questionnaire surveys; consequently, certain deficiencies exist in data validity. However, the PSUR can be objectively generated from the car use of travelers. By analyzing the use of parking facilities in the community to determine the collective behavior of car owners at the community level, car usage behavior can be explored from a new perspective and the validity of the conclusion can be improved.

(2) Empirical studies on the nonlinear relationship between the built environment and travel behavior in China are limited. Due to differences in society, economy, and culture, the relationship between the built environment and travel behavior varies between regions [10,11,24,25]. Hence, contributions to the empirical research on Chinese cities are necessary to enrich the relevant conclusions for guiding community life circle planning.

3. Variables and Research Basis

This study explores the impact of the built environment on car usage behavior. Moreover, the measurement of selected variables of the built environment focuses on those that affect the travel mode choice and decision to travel. Based on a literature review, studies on built environment and traffic behavior indicate that the former is typically characterized by 5D elements: density, diversity, design, destination accessibility, and distance to transit [3,26]. Density and diversity can be described by population density and land use mix, respectively. Design can be characterized by the road network and intersection densities. Distance to transit can be described by bus stop and rail stop densities. Destination accessibility can be reflected by the distance from the business district. Because the socioeconomic characteristics of car owners affect their travel behavior, housing price is included in the model as a control variable. At the same time, considering the availability of data and the guidance to urban planning and design, we chose the above built environment variables.

Based on the urban parking facility planning guide issued by the Ministry of Housing and Urban–Rural Development of the People’s Republic of China, the PSUR is defined as the proportion of parking facilities occupied to the total number of parking facilities during a particular time of day. The PSUR reflects the surplus of parking space resources and indirectly indicates the extent of car use.

By considering the PSUR as the dependent variable and built environment variables as independent variables, the nonlinear influence of built environment variables on the utilization rate of parking facilities is described by the machine learning method. The influence path of built environment variables on the PSUR is shown in Figure 1. The built environment has an influence on whether travel demand is generated and whether cars are used, thus indirectly affecting the PSUR. That is, when individuals generate travel demand and choose to use cars, the PSUR decreases. The lower the PSUR, the higher the car use intensity.

4. Study Area and Data

4.1. Sample Selection

To comprehensively understand the parking situation in the city proper of Chongqing, our research team surveyed the parking characteristics (including static characteristics, dynamic characteristics, parking willingness, and parking behavior) of 236 parking facilities in residential areas in 2019 by stratified, random sampling. Static and dynamic characteristics are the key data that were used in this study to calculate the PSUR. Some parking facilities had undergone information transformation; hence, we could easily obtain these data from the platform. For parking facilities without information transformation, our research team obtained the dynamic characteristics through 24-h continuous observations. To ensure the reliability of the research conclusion, residential areas with a high parking allocation index were selected. The sales rate of parking spaces exceeds 95% with no external sharing of spaces. From the districts of Yuzhong, Nanan, Jiangbei, Yubei, Beibei, Shapingba, Jiulongpo, Dadukou, and Banan, 178 samples were collected. The spatial distribution is shown in Figure 2.

By combining electronic data and those obtained by continuous manual investigation of parking facilities, the parking time of vehicles and the characteristics of parking facilities were obtained to calculate the PSUR. The location data of parking facilities were obtained by the inverse geocoding of coordinate positions and crawling of the point of interest (POI) coordinate data required by the built environment from the application programming interface of AMAP through the Python program. Because different data had varied coordinate systems, the data coordinates were unified through the QGIS software; road network data were intercepted by OpenStreetMap; and housing price data were crawled from the website via the Python software.

4.2. Analysis Unit

To avoid the problem of ecological fallacy, scale effect must be considered in spatial data analysis. To analyze the built environment around the relevant parking facilities, selecting the appropriate size of the analysis unit is necessary to determine the impact of the spatial factors of the built environment on the PSUR. The entire modeling process requires that the spatial analysis unit is not extremely large. It must ensure the effectiveness of the measurement of the built environment and highlight the internal differences among the factors of the built environment. Based on the study of the built environment on the spatial unit division of travel behavior [12], the present study selected the 500 m buffer zone of relevant parking facilities as the research scale of the built environment around these parking facilities. Because the rail station has a large service radius, the research scale of the rail station density in the built environment is 800 m.

4.3. Variables to Solve

Through the buffer analysis of ArcGIS, the built environment impact area of relevant parking facilities was generated. Moreover, the number of various types of POIs, length of road network, and population density in the area were calculated using an extraction tool. Population density was derived from street scale census data. The land use mix was mainly utilized to investigate the extent of mixing various types of POIs in each relevant parking facility buffer area. Nine types of POIs were considered: government, bank, hospital, school, gymnasium, shopping mall, hotel, square, and park. Land use mix was characterized by the land use entropy index. In the location analysis, the origin-destination cost matrix was solved by constructing a road network dataset. The starting point was each relevant parking facility. The endpoints were the five major business centers in Chongqing (Jiefangbei in Yuzhong District, Guanyinqiao in Jiangbei District, Wanda Plaza in Nan’an District, Three Gorges Plaza in Shapingba District, and Yangjiaping pedestrian street in Jiulongpo District). Consequently, each relevant parking facility produced five distance values in location calculation. The minimum value was selected as the location value of relevant parking facilities. The specific description and dimensions of built environment variables are summarized in

Table 1. Indicators of built environment.

5D	Variable	Description
Density	Population density	Population per square kilometer
Diversity	Land use mix	Land use mix entropy in 500 m circular buffer zone
Design	Road density	Sum of road lengths in 500 m circular buffer zone
	Intersection density	Sum of intersections in 500 m circular buffer zone
Destination accessibility	Bus stop density	Sum of number of bus stops in 500 m circular buffer zone
	Rail transit station density	Sum of number of rail transit stations in 800 m circular buffer zone
Distance to transit	Distance to business center	Minimum distance between parking facility and five business districts in Chongqing
control variable	Housing price	RMB/m2

. In the table, the calculation formula of the land use entropy index is:

$$L=\frac{-{\displaystyle {\sum }_{}{P}_{ij}·\ln }{P}_{ij}}{\ln N_j}$$

(1)

where $L$ is the land use mix entropy, $P_{ij}$ is the proportion of the number of $j$ POIs in the $i$th buffer to the total number of POIs, and $N_j$ is the number of types of POIs contained in the jth buffer.

4.4. Descriptive Analysis

Table 2. Descriptive statistics of explanatory variables.

Variable	Min	Max	Mean	Standard Deviation	K-S Test
Distance to business center	0.77	35.70	16.62	0.69	0.13
Land use mix	0.61	2.25	1.82	0.33	0.19
Road density	1.47	13.02	0.47	2.42	0.12
Intersection density	2	5	3.28	0.65	0.30
Bus stop density	0	9	2.9	2.34	0.09
Rail transit station density	0	2	0.6	0.81	0.25
Housing price	11,158	22,160	15,385	4830	0.22

lists the statistical analysis of the variables. The Kolmogorov-Smirnov (K-S) test showed that the significance of each variable exceeded 0.05, indicating that the sample data were normally distributed and the sample of this study was suitable. Before conducting regression analysis, analyzing the correlation among independent variables and eliminating those with strong correlations were necessary. Through a multicollinearity test using the SPSS software, the variance expansion coefficient of variables was found to be less than 10, indicating that the variables have no strong collinearity; accordingly, all explanatory variables were included in the model.

5. Model

The relationship between the built environment and the PSUR has an influence on a driver’s psychological behavior, which may be nonlinear. The GBDT model is a tree-based integration method that has excellent robustness and high fitting accuracy. The response of decision trees to an independent variable depends on the values of other independent variables at higher tree levels. Thus, GBDT automatically models the interaction effects among independent variables. Moreover, it can capture subtle and sudden changes in the PSUR and improve the prediction accuracy through boosting. The literature reports that GBDT offers better prediction precision than traditional modeling techniques. Many studies show that GBDT outperforms regression [25], autoregressive integrated moving average type, and random forest (RF) [27] models as well as neural networks (NN) and support vector machine (SVM) [28] methods. However, GBDT has weaknesses; similar to single-equation models, it cannot account for the causal order among independent variables.

5.1. GBDT Model

The base learner of the GBDT model is the regression tree, which is a tree-based ensemble algorithm [29,30]. The trees of the GBDT model are constructed in succession; that is, the first tree trains all samples. The next tree reduces the residual of the previous tree as a goal and iterates continuously until it reaches the number of iterations or the number of preset trees. The final model summarizes the results of each tree by adding their weights, as follows:

$$\widehat{F}(\mathrm{x})={\displaystyle \sum _{j=1}^J{\gamma }_j}{h}_{}(x,c_j)$$

(2)

where $\widehat{F}(x)$ denotes the approximate function to be estimated, $x$ represents the built environment variables, ${\gamma }_j$ represents the weights of weak learners, $c_j$ denotes the parameters of weak learners, $h_j(x)$ is the estimation results of weak learners, and $J$ is the number of weak learners. The weight of each learner and the parameters to be estimated are calculated by the gradient boosting method. Its algorithm can be summarized as follows.

Step 1: Initialize $F_0(x)$ by calculating the constant value that minimizes the loss function, $F_0(x)=\underset{\gamma }{\arg \min }{\displaystyle \sum _{i=1}^{N}L[y_i,\gamma ]}$, where $L[y_i,\gamma ]$ is the square loss function, $\gamma$ is the parameter to be estimated, and M is set as the maximum number of iterations.

Step 2: Calculate the residual of the Mth iteration,

$$r_{im}=-{\left\{\frac{\partial L[y_i,F(x_i)]}{\partial F(x_i)}\right\}}_{F(x)=F_{m-1}(x)}$$

(3)

Step 3: Use the weak learner, $h_m(x,c)$, to fit the residuals, $r_{im}$, in Step 2 by

$$c_m=\underset{c}{\arg \min }{\displaystyle \sum _{i=1}^n(r_{im}-h(x,c))}$$

(4)

Step 4: Estimate multiplier

$${\gamma }_m=\underset{\gamma }{\arg \min }{\displaystyle \sum _{i=1}^{n}L[y_i,F_{m-1}(x_i)+\gamma h_m(x_i,c_m)]}$$

(5)

Step 5: Update model

$$F_{\mathrm{m}}(x)=F_{m-1}(x)+{\gamma }_m{h}_m(x)$$

(6)

Step 6: Determine whether the preset iteration times and accuracy requirements are satisfied. If these are satisfied, the final estimation result is obtained; otherwise, return to Step 2.

5.2. Relative Importance of Factor

By calculating the mean value of all additive trees, the relative importance of a single independent variable to the dependent variable is measured:

$$I_i{}^2=\frac{1}{J}{\displaystyle \sum _{j=1}^J{I}_i{}^2(T_j)}$$

(7)

$$I_i{}^2(T_j)={\displaystyle \sum _{t=1}^{H-1}{\tau }_t{}^2}I[v(t)=i]$$

(8)

where $I_i{}^2$ is the degree of impact on the PSUR, $T_j$ is the additive tree, $I_i{}^2(T_j)$ is the influence of the additive tree on the PSUR, $t$ is the terminal node of the tree, $H$ is the number of terminal nodes, ${\tau }_t{}^2$ is the adjustment factor, and $v(t)$ is a function used to determine whether $i$ belongs to the segmented region (when $i$ belongs to the segmented region, $I$ = 1; otherwise, $I$ = 0).

5.3. Partial Dependence Plots

After the most relevant variables have been identified, the next step is to understand the nature of the dependence of the approximation, $F(x)$, on their joint values. The graphical rendering of $F(x)$ as a function of its arguments provides a comprehensive summary of its dependence on the joint value of input variables. Partial dependence plots can be mathematically defined as follows:

$$F_{x_S}\left(x_S\right)=E_{x_C}\left[F\left(x_S,x_C\right)\right]$$

(9)

where $x_S$ is the explanatory variable of concern, $x_C$ is the other explanatory variable, and $F_{x_S}\left(x_S\right)$ is the partial dependence of explanatory variables on the approximation function, which can be estimated by

$${\overline{F}}_{x_S}\left(x_S\right)=\frac{1}{J}\sum _{i=1}^{n}F\left(x_S,x_C\right)$$

(10)

6. Results

To conduct a quantitative analysis, a fivefold cross-validation procedure was applied to determine the model parameters. To prevent overfitting, a learning rate of 0.001 was set, generating a final model with a low prediction bias and reasonable tree size [31]. The grid search method was used to cross-verify and evaluate all the combinations of hyperparameters. When the number of iterations was 1000, the learning rate was 0.01, tree depth was 4, and the mean square error (MSE) of the model was minimal (Figure 3); accordingly, these were used as the model parameters.

To examine the effectiveness of the GBDT model for explaining the impact of the built environment on PSUR, a comparison with several machine learning models, including SVM, RF, and AdaBoost, was conducted.

Table 3. Accuracy of different machine learning models.

Model	Accuracy
SVM	0.785
RF	0.723
AdaBoost	0.816
GBDT	0.851
OLS	0.322

summarizes the comparison results. According to Table 3, the GBDT model yields the best performance, indicating that this model has advantages in modeling the impact of the built environment on the PSUR using the data obtained in this study.

6.1. Analysis of Overall Effect

Based on the parking survey and the built environment data of Chongqing, this study used the “GBM” package in the R programming language to solve the GBDT model.

Table 4. Model results.

Variable	Rank	Relative Importance (%)
Population density	3	14.20
Land use mix	4	12.60
Road density	8	6.50
Intersection density	7	9.80
Bus stop density	5	11.67
Rail transit station density	6	10.03
Distance to business center	2	14.30
Housing price	1	20.90

lists the effects of the explanatory variables on the PSUR. Because the effects were measured with relative importance, the collective importance of all explanatory variables is 100%. The higher the relative contribution value of an explanatory variable, the stronger the effect of the variable on the corresponding dependent variable. All the explanatory variables are ranked according to their relative contributions.

In terms of the relative importance of variables, the housing price, which has a contribution of 20.90%, was the most significant independent variable for the PSUR. This indicates that the individual economic level of car owners has the most significant impact on car usage behavior. In the built environment variables, the distance to business center (14.30%), population density (14.20%), and land use mix (12.60%) had a more significant impact on the PSUR than the bus stop density (14.67%) and rail transit station density (10.23%). The foregoing shows that increasing the extent of land use mix, mixed commercial and residential development, and control of population density can be an effective strategy to reduce the use of private cars. However, road network variables, such as intersection density (9.80%) and road density (6.50%), may not produce desirable outcomes.

6.2. Analysis of Nonlinear Effect

To further investigate how built environment variables influence the PSUR, partial dependence plots that show the relationship between PSUR and a set of built environment features are provided. A partial dependence plot provides a graphical depiction of the marginal effect of a variable on the response variable after accounting for the average influence of all other variables in the model [32,33]. In the GBDT approach, this relationship is not constrained by the assumption of linearity, and the partial dependence plots show the data-based estimate of the independent variable on driving distance. This aids in understanding the impact of changes in a single built environment variable when all other independent variables are integrated.

Figure 4 demonstrates the effects of seven built environment variables on PSUR. From the overall trend, an inverted V-shaped relationship is observed between population density and PSUR. The distance to business center is negatively correlated with the PSUR. The remaining community-built environment variables are positively correlated with the PSUR.

From the perspective of community population density, the PSUR gradually increases with the population density between 0 and 1.6; after 1.6, the PSUR decreases (Figure 4a). Note that in 2018, the average population density value of China’s built-up areas was approximately 0.8; hence, 1.6 represents a high population density. This may be because a low population density typically represents the lack of community infrastructure resources, which promote long-distance travel and lead to the use of cars. If the population density is extremely high, problems, such as smaller street space per capita, increased public transport congestion, and poor community environment, are expected to occur. These can inhibit people from choosing green travel modes (such as walking and public transportation) and promote the use of cars. Therefore, community space planners must focus on regulating population density. From the perspective of reducing the use of cars, the community population density value must be maintained at approximately 1.6; however, this does not mean that 1.6 is the appropriate density value, because other aspects, such as community safety and comfort, must also be considered.

From the perspective of land use mix in the community, when the variable is within 0.55 or exceeds 0.7, the effect of land use mix is trivial; this has a positive impact on the PSUR (Figure 4b). With the increase in extent of land use mix, diverse travel requisites are easier to satisfy in small-scale space, thus reducing the use of private cars. However, less functional mixing also has a less significant effect on reducing the PSUR. This shows the diversity of the activity space of community residents, and planners must focus on satisfying the diverse requirements in the layout of public facilities in the community.

From the perspective of community road network design, when the road density exceeds a 6 km/buffer (approximately equal to 7.64 km/km²), its effect on improving the PSUR begins to be evident (Figure 4c). In China’s comprehensive urban transportation system planning standard, the recommended road system density in the central urban area must not be less than 8 km/km². Moreover, the intersection density variable starts to have a stimulating effect on the PSUR when its value exceeds 10 (Figure 4d). The layout of multiple intersections with a dense road network is conducive to reducing the mobility of cars. In contrast, high road network density and intersection density mean more connected streets, providing more convenient basic walking conditions. This also confirms the effectiveness of the humanized street space layout form of “small-scale and dense road network” in shaping the current community space in China.

From the perspective of the distance to transit variable, overall, the higher the bus stop and rail transit station densities, the higher the PSUR. Due to the high values of these two variables, which reflect the high accessibility of community public transport, the probability of travelers deciding to stop using private cars for travel is high, thus improving the PSUR. Furthermore, the bus stop and rail transit station densities must exceed 4 and 1, respectively, to increase the PSUR (Figure 4e,f, respectively). The former indicates that buses require “scale”; hence, the density of bus stops must be sufficiently large. This suggests that the planning of community public transport must focus on improving the service density; more than four bus stops in the community scale are necessary to restrain the use of cars. A rail transit density value of 1 indicates that if the community is to be served by rail transit (possibly due to the high connectivity of rail networks) a station is required.

From the perspective of the distance to business center, when this variable is within 2 km, its impact on the PSUR appears to be slight. When it is in the range 2–5 km, the PSUR substantially decreases; beyond 5 km, its effect does not vary (Figure 4g). The closer the community to the business district, the greater the accessibility of public transportation and the better the walking environment; however, the more severe the road congestion problem. Consequently, the use of cars is reduced, and the PSUR improves. This conclusion indicates that the service radius of the comprehensive business district to the community is 2 km, and the influence radius is 5 km, thus providing a reference for the spatial layout of a large business district.

6.3. Synergy between Built Environment and Housing Price

In terms of variable contribution, housing price is the main factor affecting the PSUR. It represents the individual economic attribute characteristics of travelers to a certain extent; hence, it can be compared with the objective characteristics of the built environment. Accordingly, this section discusses the impact of the interaction between the housing price and built environment on the PSUR; the interaction effect diagram is shown in Figure 5. The relationship between some built environment variables and PSUR has not been modified by the change in housing price; thus, these variables have no interaction with the housing price. The absence of interaction between the housing price and population density, road density, intersection density, and rail station density can be easily observed. In contrast, the interaction effect between the housing price and land use mix, bus stop density, and distance from the city center is evident.

Overall, the increase in housing price weakens the impact of the built environment on car usage behavior. In areas with low housing prices, improving the built environment can more effectively inhibit the use of cars. Regarding the land use mix (Figure 5b), in areas where the housing price exceeds 16,000 yuan/m², the inhibitory effect of improving the land use mix on the use of cars begins to weaken. As shown in Figure 5b, when the housing price exceeds the fixed house price of 16,000 yuan/m², the PSUR remains unchanged with the increase in land use mix. For bus stop density, in an area where the housing price exceeds 14,000 yuan/m², the effect of improving the public transport service level on restraining the use of cars begins to weaken. Moreover, the improvement in the public transport service level is more sensitive to the change in individual social and economic levels. That is, as shown in Figure 5e, when the housing price exceeds the fixed house price of 14,000 yuan/m², the PSUR remains unchanged with the increase in bus stop density. Regarding the distance of the community from the business center, in an area where the housing price exceeds 17,000 yuan/m², the effect of reducing the distance from the business center on restraining the use of cars begins to weaken. That is, as shown in Figure 5g, when the housing price exceeds 17,000 yuan/m² (fixed house price), the PSUR remains unchanged despite the reduction in the distance of the community from the business center. This is possibly due to the high level of income of travelers living in areas with high housing prices. The analysis of the travel survey of residents in Chongqing in 2016 indicates that the travel of high-income groups is highly dependent on cars. From the perspective of land planning and the provision of public transport services, adjusting the travel habits of these groups is difficult, and more rigid policies and measures are necessary to limit the use of cars. The foregoing alerts planners to focus on the effect of differentiation in the process of designing or transforming the built environment; that is, the improvement of the built environment has a better effect on areas with a low economic development level.

7. Conclusions and Discussions

This study employed the GBDT model to explore the nonlinear relationship between the built environment and car usage behavior using data from Chongqing. To the best of our knowledge, this is the first study that employs the GBDT method in the field of land use and car usage behavior from the perspective of parking facility use behavior. It offers insightful results to the literature.

First, the GBDT model overcomes the defect of the linear hypothesis of the traditional statistical model. The analysis of the use behavior of the built environment on parking facilities shows that the built environment has a nonlinear effect on car usage behavior, and the importance of various built environment variables differs. A similar result has been observed in previous studies on the built environment and car usage behavior [15,21].

Second, among the variables observed in this study, in addition to the housing price (23.5%), the distance to business center (14.30%), population density (14.20%), and land use mix (12.60%) have a substantial effect on car usage behavior. This indicates that socioeconomic attributes exert a greater influence on car usage behavior than built environment variables. However, a unified conclusion has not been reached on which of the socio-economic attributes and built environments variables have a greater impact on car usage behavior. In contrast, the foregoing reflects the importance of community land planning and population control in influencing the car usage behavior. Ding reached a similar conclusion by exploring the impact of the built environment on driving distance [15]. Notably, this study provides a direction for quantitatively improving low-carbon community life circle planning in the context of China’s urban development. It emphasizes that the transformation of community-built environments must start with land use planning. With the continuous increase in the level of income of the Chinese people, traveler shopping also increases. Thus, reducing the travel distance and regulating the use of cars by building multiple business centers and promoting mixed land development are extremely significant.

Third, the community-built environment variables show salient nonlinear effects (including scale and threshold effects) on car usage behavior. The foregoing alerts planning managers to focus on the following two points when considering the transformation of travel behavior through urban design. The first is to grasp the scale effect: Only a certain number or degree of built environment variables can significantly reduce the use of private cars. The second is to be aware of the thresholds of built environment variables. The transformation of the community-built environment must not be blindly implemented; instead, planners must accurately grasp the mechanism behind the transformation to avoid a waste of resources.

The empirical study of Chongqing, China shows that the optimal community population density is 16,000 people/km². In contrast, in a study on Austin, USA, Ding found that car use was considerably reduced after the population density of the city center exceeded 3000 people/km² [15]. The difference is possibly due to the high land development intensity in Chinese cities; consequently, the population carrying capacity of cities is also high. A previous study by Yang only emphasized the improvement of land use mix and public transport service levels to curb car use [34]. It did not specify the level of land use mix that must be achieved. In the community life circle planning outline, the paper only mentioned the necessity of public transport within the community; it did not emphasize the service level of public transport. This current study provides a quantitative optimization direction through nonlinear relationship analysis. It emphasizes that, due to the dependence of travelers on car use, conventional public transport must have a scale effect to attract car travelers effectively. The study also shows that the service radius of large commercial centers is 5 km; this finding is extremely relevant to the location of new commercial centers that urban planners must note in their designs. In addition, the effective value range of road density and intersection density is also proposed, which is of great significance to road network planning. Interestingly, in China, the higher the road density and intersection density, the more it inhibits the use of cars. Cervero’s research, based on the data of 370 cities in the United States, showed that road density has a positive effect on the use of cars [35]. There may also be differences in the level of public transport services. In contrast, because Chongqing has built a large number of rail transit lines and opened a multilevel bus line network over the last 20 years, this provides travelers with more travel mode choices. The road density and intersection density directly affect the accessibility of public transport stations.

Previous studies have not explored the interaction between economic attributes and the built environment on car usage behavior. This study considered housing price as the representation of economic attributes. Because the housing price has spatial characteristics, understanding the interaction between the housing price and the built environment can further guide the spatial differentiation of policies. The study finds three variables that have interactive effects with the housing price: land use mix, bus stop density, and distance from business center. Therefore, the design of these three variables must consider spatial factors, which have a better marginal effect on low-income areas. In contrast, variables without an interaction effect exert a strong influence on inhibiting the use of cars by all income groups. These variables are population, road, intersection, and rail transit station densities.

Although Chongqing was used as the geographical background for this empirical research, the foregoing conclusions can be used as reference for relevant planning and management in other, similar cities. Moreover, due to variations in cultural and geographical backgrounds, the authors presume that different cities must implement research based on their own urban observation data. The use of parking data is also recommended to better reflect car usage behavior, thus compensating for the defects of traditional survey data. Due to the limited data, the variables we selected were limited, and the conclusion of this study needs more empirical research to be verified. In future research, conducting multiscale built environment analysis will be necessary to draw more reliable conclusions. Moreover, with the maturity of image recognition technology, the measurement of built environments in the future must consider more of the microfeatures of streets affecting travel behavior. Because these variables are easy to optimize in practice, they provide a facile optimization direction that is beneficial to the improvement of travel environment for the purposes of inhibiting the use of cars.