Exploring the Influence of the Built Environment on the Demand for Online Car-Hailing Services Using a Multi-Scale Geographically and Temporally Weighted Regression Model

By inergency On Feb 22, 2024

1. Introduction

The development of digital technology has led to huge changes in the field of transport [1]. The development of internet technology has made it possible for users to conveniently use the web, enabling passengers and drivers to connect through online platforms. As a result, online car-hailing has emerged. The emergence of online car-hailing has rapidly spread to all major cities in China. As of the first half of 2022, the number of active users and drivers in the online car-hailing sector had reached 360 million and 4.53 million, respectively. Online car-hailing is progressively transitioning to electric vehicles, contributing to reduced resource consumption, lower carbon emissions, and achieving sustainable development. Online car-hailing is also essential in achieving China’s dual-carbon goals. Online car-hailing will be further developed to promote ease of travel and lower resource use. However, the rapid expansion of online car-hailing has resulted in various challenges, including long waiting times and high vehicle idling rates. It leads to wasted resources and decreased user satisfaction. To address these issues, gaining a deeper understanding of the root causes of changes in passenger flows is crucial.

Previous studies on online car-hailing have primarily focused on predicting trip demand and analyzing factors influencing passenger flow [2,3]. Such studies have highlighted the significance of online car-hailing as a mode of transportation. Nevertheless, there needs to be more investigation of the link between the built environment and passenger flow.

In mainstream academic journals, studies often use terms that start with a “D” to reflect features of the built environment. The urban built environment comprises “seven Ds”: density, diversity, design, destination accessibility, distance to transit, demographics, and demand management [4]. Existing studies have demonstrated a strong correlation between travel behaviors and the built environment [5]. Features of the built environment have varying impacts on different travel patterns. For bike-sharing, Ma et al. [6] used the GTWR model to compare differences in the built environment’s impact between docked and dockless bike-sharing systems. Chakour et al. [7] used a composite marginal likelihood (CML)-based ordered response probit (ORP) model to conclude that an effective way to increase bus ridership is by improving public transport service and accessibility. Gan et al. [8] applied a gradient boosting regression tree model to investigate the non-linear relationship between built environment characteristics and station-to-station ridership. For online car-hailing, Li et al. [2] employed the GWR model to examine the influence of the built environment on online car-hailing usage. They concluded that online car-hailing services play a significant role as a complementary mode of transportation to public transport services. For walking, Tao et al. [9] adopted the gradient boosting decision tree method to examine the relationships between walking distance and spatial attributes. The studies listed above demonstrate a strong link between different modes of travel and the urban built environment. While studies have been conducted to examine the relationship between the urban built environment and the demand for online car-hailing, further research is warranted on this aspect of the relationship.

Statistical modeling is commonly used in research to examine the relationship between travel behavior and the urban built environment. The ordinary least squares (OLS) model is the most representative and widely used among these models. However, the OLS model assumes that spatial variable relationships are fixed and do not change with spatial location, which is unrealistic [10]. However, the demand for online car-hailing exhibits spatial heterogeneity. Spatial heterogeneity refers to variations in the demand for online car-hailing trips at different spatial locations. Several models have been proposed to address spatial heterogeneity, including distance decay weighted regression, two-stage least squares regression, the passion model, and the GWR model [11]. The GWR model, introduced by Brunsdon in 1998, examines the impact of different factors on a localized scale by constructing regression equations for each sub-region within the study area [12]. The GWR model is used not only in the analysis of factors influencing passenger flow [13,14,15] but also in other fields [16,17]. It is also used in the context of the demand for online car-hailing travel. Bi et al. [3] clustered passenger flows into three patterns based on the time-varying characteristics of passenger flows and established a spatial passenger flow model based on the GWR model. The results verified that the built environment has different degrees of influence on the passenger flow of online car-hailing travels in the spatial and temporal dimensions. Grid size also affects model results. Zhao et al. [18] recognized this issue, partitioning the area into grids of different sizes and establishing a GWR model to investigate the variations in the influence of various factors on the demand for online car-hailing trips.

In addition to spatial heterogeneity, the demand for online car-hailing travel also displays temporal heterogeneity, which is evident in the demand fluctuation over time. Temporal heterogeneity and spatial heterogeneity collectively constitute spatiotemporal heterogeneity. The GWR can only analyze space instability, which makes its spatiotemporal analysis incomplete. Therefore, the GTWR model was developed by improving the GWR model [19]. This model was first applied in the field of house price research and has since been used extensively in other fields as well. Shen et al. [20] analyzed the spatial and temporal distribution pattern of GTWR coefficients and compared the demand for different travel modes. Their results showed substantial spatial and temporal heterogeneity in the demand for different travel modes.

The GTWR and GWR models assume that all independent variables have the same bandwidth, meaning that the influence scope of the variables is the same. However, reality often deviates from this assumption. For instance, transportation facility POIs may attract online car-hailing services from a greater distance, indicating that the influence scale of such POIs is larger. This element was not addressed in any of the studies mentioned above, leading to significant model errors. The range of influence varies between variables, as reflected in the bandwidth of each variable. Therefore, the multi-scale geographically weighted regression (MGWR) model was proposed. Cao et al. [21] constructed MGWR models for different periods to derive significant spatial and temporal heterogeneity in the impact of various factors on the demand for online car-hailing trips.

In addition to GWR-derived models, other regression models are suitable for examining the relationship between travel demand and the built environment. For example, Liu et al. [22] used a generalized additive mixed model to study the relationship between the built environment and taxi demand. Sun et al. [23] clustered periods and then employed the spatial autoregressive moving average (SARMA) to investigate similar relationships. However, such regression models are far less effective in addressing spatial issues than GWR and its derivative models and are less commonly used.

Machine learning has gained popularity in research due to its higher fitting accuracy. Gan et al. [8] used gradient decision boosting trees to explore the relationship between the built environment and rail patronage in an urban setting. Hagenauer et al. [24] conducted a comparative analysis of modal choice modeling using multiple machine learning classifiers and found that random forests were the most accurate. Cheng et al. [25] also chose random forests to explore the effect of the built environment on travel behavior. Although they achieved a high level of model accuracy, they lacked consideration for spatiotemporal heterogeneity. Cheng et al. [26] addressed this limitation by employing an improved version of gradient decision boosting trees, taking spatial heterogeneity into complete account.

Compared to regression models, machine learning exhibits higher accuracy. However, machine learning models can only explore the impact of different built environments on the demand for online car-hailing within the study area, with less consideration of spatiotemporal heterogeneity. Considering spatiotemporal heterogeneity helps better explain the impact of a specific factor on online car-hailing travel in a particular spatiotemporal context.

Spatiotemporal positions influence the impact scales of different built environments. However, as shown in Table 1, limited research has considered both spatiotemporal heterogeneity and impact scales. The GWR, GTWR, and MGWR models cannot account for variable range and time scale. Machine learning can only explore the effect of a variable on travel behavior and cannot consider spatiotemporal heterogeneity. MGTWR, an extension of the GTWR model incorporating impact scales, is a beneficial tool for investigating spatiotemporal heterogeneity. This study will establish a MGTWR model to comprehensively explore the relationship between the built environment and the demand for online car-hailing trips.

Based on the issues mentioned above, our research study has the following contributions:

To fully consider the range of influence of different factors and time scales, MGTWR and GTWR will be used to study the influencing factors of the demand for online car-hailing trips;
There may be differences in travel demand between weekdays and holidays, which will be investigated separately;
The impact of each built environment variable on the demand for online car-hailing trips will be analyzed from a spatiotemporal perspective.

The rest of this paper is organized as follows. Section 2 presents the data sources and data processing. Section 3 describes the methodology used for this study. Section 4 analyzes the results of the models. Section 5 concludes this paper with an outlook for the future.

4. Experiment

In this section, we will compare the model results and analyze the impacts of different POIs on the demand for online car-hailing trips in both the temporal and spatial dimensions.

4.1. Evaluation metrics

The corrected Akaike information criterion (AICc) and the coefficient of determination (R²) were used to evaluate the fit of our model. AICc and R² are calculated as follows [32,33]:

$AICc = n \cdot \ln (RSS) + 2 (p + 1) - n \ln (n) + [2 p (p + 1) / n - p - 1]$

(14)

$R^{2} = 1 - \frac{\sum_{i = 1}^{n} {({\hat{y}}_{i} - y_{i})}^{2}}{\sum_{i = 1}^{n} {(\frac{1}{n} \sum_{i = 1}^{n} y_{i} - {\hat{y}}_{i})}^{2}}$

(15)

where $n$ represents the number of samples, $p$ represents number of variables, and ${\hat{y}}_{i}$ and $y_{i}$ denote the fitted values and actual values, respectively.

The AICc takes values from 0 to positive infinity. The larger the value of the AICc, the worse the fit.

The value R² ranges from 0 to 1. An R² value closer to 1 means that the prediction is better. If the R² value is 0, each predicted sample value equals the mean, the same as the mean model. If the R² value is less than 0, the constructed model is not as good as the mean model.

4.2. Performance of the MGTWR Model

Table 8 presents the fitting accuracy of the GTWR and MGTWR models for weekdays and holidays. For both weekdays and holidays, the AICc of the MGTWR model is smaller than that of the GTWR model, with a reduction of 1540 and 2547, respectively. Additionally, the R2 of the MGTWR model is higher than that of the GTWR model for both workdays and holidays, with an increase of 0.062 and 0.057, respectively. These indicators indicate that the MGTWR model outperforms the GTWR model’s fitting accuracy for workdays and holidays. Therefore, the subsequent analysis will use the MGTWR results to explore the impact of different types of POIs on the demand for online car-hailing trips.

The impact of different types of POIs on the demand for online car-hailing is different. Table 9 and Table 10 present the regression coefficients’ mean, standard deviation, and minimum and maximum values for different types of POIs in the MGTWR model. A positive regression coefficient indicates that the POIs facilitate demand for online car-hailing trips. A more significant value corresponds to a greater facilitating role. Conversely, a negative coefficient suggests a disincentive effect on demand. The mean value provides an overall estimate of the impact of POIs on online car-hailing trips in the study area. At the same time, standard deviation indicates the stability of the effect. For both weekdays and holidays, the mean regression coefficients for shopping service POIs, address information POIs, and car service POIs are close to zero, indicating a weak impact on the demand for online car-hailing trips. On the other hand, corporate enterprise POIs and government agency POIs have a more significant impact on weekdays.

4.3. Temporal Features of Variable Coefficients

Figure 6 indicates that shopping service POIs have a more pronounced effect on promoting online car-hailing during holidays. This can be attributed to increased demand for online car-hailing due to more people shopping during holidays. Furthermore, holidays have an even more noticeable effect on demand for online car-hailing, with the boosting effect starting to increase sharply from around 10:00. This change is because passengers tend to leave later during holidays. During part of the weekday period, the regression coefficient is less than 0, which means that shopping service POIs create less demand for online car-hailing trips during this time. The regression coefficient is close to 0 for most of the weekday period. This suggests that shopping service POIs have little to no effect on the demand for online car-hailing trips.

Compared to shopping service POIs, catering service POIs have a more significant impact on demand for online car-hailing trips during holidays, and the boost on weekdays is lesser than during holidays. Catering service POIs contribute to demand for consistent online car-hailing trips. The highest demand occurs between 20:00 and 22:00. Demand experiences a rapid growth between 12:00 and 14:00 due to the need for transportation after meals.

There is minimal variation in the promotion of online car-hailing by POIs of corporate enterprises during weekdays and holidays. This illustrates the existence of holiday work in some businesses. As people do not work during the evening hours, these POIs tend to create much less demand for online car-hailing services during this time. The peak of the off-duty period is at 16:00–18:00, which is when corporate enterprise POIs have the greatest impact on boosting demand for online car-hailing trips.

Due to their high traffic volume, transport facility POIs, such as ports, airports, and train stations, are the main contributors to the demand for online car-hailing trips. Their effect is more pronounced during working days than during holidays. However, between 2:00 and 6:00, the number of passengers is low, and the positive effect is the lowest. As the number of passengers increases, the boost in demand becomes more evident.

Figure 10 shows that government agency POIs consistently contribute to the demand for online car-hailing trips. However, on holidays, some government institutions are closed or operate with limited staff, resulting in less promotion of online car-hailing travel than on weekdays. The demand for online car-hailing trips is greatest between 16:00 and 20:00, when employees are likely to take a taxi to reach their destination. Increased demand is also more pronounced between 10:00 and 14:00 due to the need to travel or go home for a lunch break. Additionally, some government agencies are in regular business on holidays and thus generate some demand.

Figure 11 demonstrates that the impact of healthcare service POIs on the demand for online car-hailing trips is highly variable. However, the changes are relatively consistent. The most significant increase is observed between 8:00 and 12:00, but this effect decreases rapidly in the afternoon and becomes negative. It indicates that healthcare service POIs switch from facilitating to suppressing demand for online car-hailing trips. There is a partial rebound during the evening peak on weekdays, after which the effect declines again.

The impact of accommodation service POIs on the demand for online car trips peaks mainly between 8:00 and 14:00, as shown in Figure 11. During holidays, hotel occupants increase, resulting in an even greater boost. Accommodation service POIs mainly refer to hotels, and check-out times are between 8:00 and 14:00, leading to a higher demand for online car-hailing trips. However, after this period, the demand for travel decreases, and the promotion effect rapidly drops. The demand is lowest between 20:00 and 24:00, and during this time, accommodation service POIs even suppress the demand for online car-hailing trips.

Car service POIs mainly include places for car repair, car sales, car cleaning, and other car-related services. The impact of car service POIs on the demand for online car-hailing trips is complex. For instance, car cleaning services are likely to have a negative effect on the demand for online car-hailing trips, whereas the impact of car repair and car sales on online car-hailing trips is uncertain. It is puzzling that most car service POIs are closed at night. However, they contribute the most to the demand for online car-hailing trips during this time.

4.4. Spatial Feature of Variable Coefficients

Figure 14 indicates that as the research area shifts from west to east, the impact of shopping service POIs on the demand for online car-hailing trips changes from negative to positive. The figure also highlights that the positive effect of shopping service POIs is more pronounced during holidays, especially in the upper right-hand corner of the study area. Furthermore, the distribution chart of POIs illustrates a positive correlation between the number of shopping service POIs and the demand for online car-hailing trips, with the number of shopping service POIs increasing from left to right. In other words, a higher quantity of shopping service POIs corresponds to a greater demand for online car-hailing travel.

Catering service POIs consistently positively impact the demand for online car-hailing trips within the study area, which is further pronounced during holidays. The impact of catering services POI on online car-hailing trips gradually increases from the lower left corner to the upper right corner of the study area. Similar to the effects of shopping service POIs, there is a positive correlation between the density of catering service POIs and the demand for online car-hailing trips, where a higher density leads to a more prominent positive impact.

The demand for online car-hailing trips from corporate enterprise POIs remains constant across different types of dates. The distribution map of corporate enterprise POIs reveals a gradual increase in their number from the lower left corner to the upper right corner of the study area. However, the positive impact of these POIs on the demand for online car-hailing progressively decreases in the same direction. A negative correlation exists between the number of corporate enterprise POIs and their resulting impact, which may be due to differences in travel convenience. As transportation facilities in the lower left corner are fewer and more inconvenient to use, employees in this area may have a higher demand for online car-hailing services.

Overall, transport facility POIs contribute to the demand for online car-hailing trips. Transport facility POIs have the greatest impact on the demand for online car-hailing trips in the lower left corner of the study area. This may be due to the low number of transport facility POIs in the area, making travel more inconvenient and increasing the demand for online car-hailing.

The impact of government agency POIs on the demand for online car-hailing is highly dependent on the type of day. The positive effect of government agency POIs on the demand for online car-hailing trips during workdays is greater than during holidays, indicating that some POI staff may not work during holidays or some agencies may close. Like corporate enterprise POIs, the bottom left region, where the demand for online car-hailing trips is most influenced by government agency POIs, has the most significant positive impact. In the other areas, government agency POIs are evenly distributed, which leads to a slight variation in the spatial distribution of the coefficients.

The influence of healthcare service POIs on the demand for online car-hailing trips is consistently negative during holidays, while only the top left corner has a boosting influence on weekdays. The regression coefficient in the study area presents a bottom–up increase, with sharper changes on the left and gradual increases on the right. In general, there is a positive correlation between the density of healthcare service POIs and the demand for online car-hailing trips.

Accommodation service POIs are primarily utilized by tourists, who often require online car-hailing services. Consequently, during the holiday season, accommodation service POIs positively impact online car-hailing, and the promotion is more noticeable. The facilitating effect of accommodation service POIs on the demand for online car-hailing travel increases from the lower right to the upper left corner of the study area. This effect is more prominent in areas with low POI density and less noticeable in areas with high POI density, as indicated by the POI distribution map.

The regression coefficient for car service POIs is positive in most regions, indicating that car service POIs contribute to the demand for online car-hailing travel. However, there is a negative impact of car service POIs on the demand for online car-hailing in the bottom right part of the region. The effect of car service POIs on demand for online car-hailing trips changes from negative to positive as one moves from the lower left to the upper right corner, eventually decreasing to approximately zero. This change is consistent with the change in the density of automotive service POIs, indicating a positive correlation between the resulting impact and density.

5. Conclusions

This paper explored the correlation between travel demand and different categories of POIs by constructing a GTWR model and an MGTWR model employing POI data and online car-hailing data. The results showed that the MGTWR model has a better regression effect than the classical GTWR model. Then, the result of the MGTWR model was used to analyze the spatiotemporal heterogeneity of the influence of different POIs on travel demand. Finally, we also obtained some findings with the regression coefficients for each type of POI. Transport facility POIs had the greatest boosting effect on the demand for online car-hailing trips, while healthcare services had the opposite and the most significant inhibiting effect. Recreational POIs were had the most significant promoting effects on holidays, while corporate enterprise POIs, government agency POIs, and healthcare service POIs were more prominent on weekdays.

In the future, the following research questions will be explored:

First, the generalizability of the model has yet to be substantiated. Due to the current limitations of the available data, it has not been possible to validate its generalization capability in other regions. In subsequent research, data from different areas will be utilized to demonstrate its generalization ability.

Second, the current study performed zoning with relatively simple meshing, which is too idealistic and may increase errors.

Finally, the data limitations resulted in unmatched online orders not being captured, leading to a potential discrepancy between the actual demand and orders. Subsequent efforts will be made to obtain these missing data.