Spatiotemporal heterogeneity in built environment impacts on urban mobility: Insights from calendar events and new towns in Beijing

Abstract

The spatiotemporal heterogeneity of the impact of BE on urban transportation remains underexplored, especially in nonlinear dynamics and the comprehensive effects of CEs and spatial configuration in sustainable transportation systems. Using smart card and multi-source urban data, this study analyzes Beijing’s mobility patterns under CEs and NTs, modeling 20 BE factors with XGBoost to uncover nonlinear BE-UM relationships, outperforming conventional OLS. The SHAP value is further introduced to quantify the nonlinear threshold effect of BE on UM, supporting local interpretation to global multi-scale analysis. The results show that (1) Plot ratio and network centrality drive UM, but their effects show significant temporal and spatial differences. CAC relies on high plot ratio development to improve commuting efficiency, while NTs promote um through high network centrality. (2) Non-workday mobility exhibits stronger sensitivity to the mixed degree of functions than workdays, reflecting BE’s adaptive demand for flexible travel patterns. (3) The impact of BE on UM has an apparent threshold effect, thus requiring differentiated planning strategies aligned with CEs’ and NTs’ functional roles. These findings inform multicentric urban optimization and sustainable transport strategies, guiding new towns’ core function decentralization through built environment threshold-based planning.

Graphical abstract

Keywords

Built environment / Urban mobility / New towns / Smart card data / Explainable machine learning

1 Introduction

Rapid urban expansion and growing demand for sustainable transportation systems have positioned efficient urban mobility (UM) as a core governance challenge. UM intersects urban planning, traffic management, economic resilience, public health, and safety (Yabe et al., 2025; Zhao et al., 2024). Public transit systems―subways, buses, taxis, and shared bikes―form the backbone of mobility in megacities, particularly in developing countries, due to their coverage, speed, accessibility, affordability, and environmental benefits (Kim et al., 2024).

The built environment (BE), shaped by socioeconomic elements and spatial configurations (Acevedo and Porta, 2025), defines UM as the macroscopic dynamic process of movement (e.g., people, goods, information) across urban spaces, manifesting as flowlines and basins (Zhang and Cheng, 2024). Unlike traditional static origin-destination (OD) models, UM captures real-time urban dynamics, reflecting structural characteristics and operational patterns.

Established studies have shown that BE and UM are significantly related, including the analytical dimension of the evolutionary expansion from 3Ds to 5Ds (Cervero and Kockelman, 1997; Cervero and Kockelman, 1997). However, existing studies are generally based on linear assumptions, and the analysis of nonlinear roles and elemental interaction mechanisms between BE-UM is significantly underdeveloped. On the time dimension, previous studies are mostly limited to intraday regular time divisions (e.g., peak hours and off-peak periods, morning, midday, and evening, etc.), either focusing on weekly regular events (e. g., workdays and weekends) and ignoring the effects of CEs such as holidays on behavioral patterns, which are important for improving the accuracy of daily commuting forecasts, capturing travel and leisure patterns, and optimizing resource allocation (Chen et al., 2021; Wang et al., 2023); the spatial dimension is overly focused on central urban areas and lacks attention to the mobility characteristics of peripheral NTs. The spatial dimension is overly focused on the central urban area and lacks attention to the mobility characteristics of peripheral NTs. In particular, China’s NTs have adopted a large-scale mixed-use development model that is different from that of the West (Xu, 2022), and there is an urgent need to build an analytical framework suitable for the Chinese context.

This study takes Beijing, China, as an example; based on smart card data and points of interest (POI) and other urban multi-source heterogeneous big data, it analyzes the spatial structure relationship of NTs through urban mobility theory. It provides a basis for promoting a public transport-oriented compact development model and facilitating the functional decentralization of megacities. The research focuses on the following questions: (1) How does UM change with different calendar events (e.g., Labour Day, Have to work, Workday, Weekend)? (2) What are the characteristics of UM between NTs and the core area? (3) Which key indicators in BE factors may promote UM improvement? The study integrates machine learning and spatio-temporal analysis methods, breaks through the limitations of traditional linear models, and reveals the dynamic thresholds and spatial differentiation laws of BE-UM under CEs for different NTs, which provides a scientific basis for the optimization of polycentric urban structure and the threshold-oriented planning of NTs for the deconcentration of CAC functions.

2 Literature review

2.1 The impact of calendar events (CEs) on urban mobility (UM)

Calendar events (CEs) refer to planned or periodic activities that trigger fluctuations in urban mobility at specific time points, covering statutory holidays (such as the Spring Festival and Christmas), cultural festivals (such as music festivals), and large-scale public activities (such as sporting events) (Liu et al., 2023). Including CEs not only refines the shortcomings of the BE-UM study in the time dimension but also provides a new perspective for studying travel behavior patterns on public holidays in China. Early studies relied on static data such as census and statistical yearbooks, making it challenging to capture the impact of different CEs (Chang et al., 2024). With the popularization of mobile sensing technology, multi-source dynamic data such as cardswiping and trajectory data have significantly improved the spatiotemporal resolution of UM measurement (Ariffin et al., 2021; Pei et al., 2020). However, existing studies primarily rely on a single data source and lack a multisource heterogeneous data fusion framework, which limits a comprehensive analysis of the spatiotemporal heterogeneity impact of CEs.

Previous studies have generally verified the significant disturbance effect of calendar events on urban mobility, but the intensity and direction of their impact show temporal heterogeneity. For example, statutory holidays such as the Spring Festival and National Day reset social rhythms, triggering interregional population migration and reshaping consumption-related travel (Liu et al., 2020); large-scale activities such as sporting events and concerts cause a surge in local road network traffic, and the intensity of their impact decays exponentially with the scale of the event (Vainio et al., 2024); weekends reorganize the spatiotemporal distribution of urban mobility by changing travel purposes (leisure substitutes for commuting) (Li et al., 2025). Most existing studies compare UM differences based on the dichotomy of working days and non-working days but ignore the impact of other calendar events (such as public holidays) on urban mobility, resulting in insufficient precision and timeliness of management measures (Kim, 2020; Vilhelmson, 2007).

The impact of CEs on UM does not exist in isolation but shapes the spatiotemporal pattern of urban mobility together with BE (Ding et al., 2019). For example, higher UM on working days is most likely to occur in the morning and evening due to concentrated commuting demand during peak hours, which is also affected by the location of workplaces and residences along the city’s main roads, while travel behaviors on weekends are relatively dispersed. UM is closely related to BE indicators such as leisure and shopping services (Bao et al., 2023). High-density mixed-use development areas can alleviate local congestion during large-scale activities through pedestrian-friendly designs (Zeng et al., 2024), while low-density suburban areas are more vulnerable to CE shocks due to reliance on long-distance motorized travel (Sharath et al., 2024). However, existing policy designs are primarily based on empirical rules and lack a quantitative assessment of the synergistic effects of CEs and BE, resulting in insufficient precision of management measures.

2.2 The influence of new towns (NTs) spatial structure on urban mobility (UM)

New towns (NTs) concept originates from the Garden City theory (Howard, 2003). After the United Kingdom initiated the New Towns Movement to alleviate the overcrowding problem in London, European countries, America, and Japan followed suit (Broughton et al., 2003). The concept of NTs was first proposed in the Chinese government report in 2010, which defines it as an urban development unit with specific functions and relative independence, formed at a particular spatial and temporal distance outside the main urban area according to the overall planning, and an integral part of constructing a multicenter and high-density urban pattern (Li, 2014).

In the multi-center, multi-axis, and multi-group spatial restructuring of mega-cities in China, NTs undertake the dual missions of relieving the pressure of the central city and promoting local urbanization. For example, the Beijing Urban Master Plan (2016—2035) proposes a spatial structure of one core, one central city, one sub-city, two axes, multiple nodes, and one district to promote the transformation from a monocentric to a polycentric model (Li and Zhao, 2022); Shanghai positions five new towns as comprehensive node cities with radiating and driving effects in the Yangtze River Delta urban agglomeration (Zhou et al., 2022).

Existing researches on the spatial structure of NTs and the relationship between BE—UM focus on land use and functional layout (Lin et al., 2021), optimization of transportation systems (Kong et al., 2020), improvement of the ecological environment (Shi et al., 2023; Zhang et al., 2023) and socioeconomic effects (Xu et al., 2025) and other aspects. Scientific planning of the spatial structure of NTs can improve the city’s operational efficiency, enhance residents’ quality of life, and promote coordinated regional development (Sharma and Patil, 2024).

2.3 Revealing built environment (BE) and urban mobility (UM) relationship through machine learning (ML)

In the past, traditional statistical methods and machine learning methods (Abideen et al., 2025) have mainly been used to reveal the relationship between the built environment (BE) and urban mobility (UM). Traditional methods such as land use regression (LUR) (Meng and Xing, 2019), geographically weighted regression (Xiang et al., 2022), autoregressive moving average (Xu et al., 2016), Kalman filter (Achar et al., 2020), and ordinary least squares (OLS) (Ouyang et al., 2022) can analyze the BE‒UM relationship to a certain extent. However, due to their reliance on small samples, assumptions of linear relationships, and specific data distributions, it is difficult for them to capture the complex non-linear interactions between BE and UM.

Machine learning (ML) methods provide a new approach to revealing the non-linear relationships between BE and UM by processing high-dimensional data and automatically recognizing patterns (Tao et al., 2024; Zhao et al., 2023). Methods such as support vector regression (Nidhi and Lobiyal, 2022), random forest (Yang et al., 2023), and gradient boosting decision trees (GBDT) (Gao et al., 2023) have been widely applied. Although GBDT performs reliably in non-linear modeling, its potential for analyzing the spatial structure of NTs under the CEs has not been fully explored.

Feature engineering is the key to optimizing the performance of ML models, and it is used to improve model performance and interpretability (Klomp et al., 2016). Standard features include road network density in geographic information, land use type, and POI distribution (Su et al., 2025); population density, GDP, and housing prices in socioeconomic data (You and Guan, 2024); subway station accessibility, distance from the city center and average passenger flow in traffic data (Peng et al., 2025). The prediction accuracy and stability can be further improved by optimizing model parameters through cross-validation, grid search, and ensemble learning.

3 Materials and methodology

Figure 1 shows all the research materials and key steps in this study. The dependent variable is the spatio-temporal distribution of urban mobility, which is derived from the smart card data of Beijing public transportation and subway. The temporal dimension includes four categories of calendar events, namely Holiday, Have to work, Workday, and Weekend, and the spatial dimension focuses on the seven study areas constituted by CACs, BMCs, and NTs. The independent variables are 20 BE data constructed based on POI and other urban multi-source heterogeneous big data.

First, it was decided to use XGBoost to analyze the nonlinear relationship between UM-BE by comparing the fit and prediction accuracy of XGBoost (nonlinear model) and OLS (linear model). Next, the SHAP value was used to interpret the XGBoost model and highly explanatory BE indicators were selected. Finally, focusing on CAC and NTs to identify the key threshold effects, planning strategies based on threshold effects, and policy recommendations to match the functional positioning of different CEs and NTs are proposed.

3.1 Study area

This study uses Beijing, China’s capital city, to identify urban mobility patterns. It has 14 subway lines involving 326 metro stations (Beijing Municipal Bureau of Statistics, 2020). Beijing consists of 16 administrative districts, among which 12 have operational subway services: Dongcheng District (DC), Xicheng District (XC), Chaoyang District (CY), Haidian District (HD), Fengtai District (FT), Shijingshan District (SJS), Changping District (CP), Shunyi District (SY), Tongzhou District (TZ), Daxing District (DX), Fangshan District (FS), and Mentougou District (MTG). In the current version of Beijing’s master plan for 2016—2035, the city has established one core area, one sub-city, and five new towns. D0 is the core area of the capital (CAC), and D3 is the Beijing Municipal Administrative Center (BMC), which was upgraded from the Tongzhou New Town (TZNT). D1, D2, D4, D5, and D6 are Changping New Town (CPNT), Shunyi New Town (SYNT), Yizhuang New Town (YZNT), Daxing New Town (DXNT), and Fangshan New Town (FSNT) respectively. Beijing is the first city in China to propose implementing reduction development, and the primary task is to relieve the non-capital functions of CAC from BMC and NTs in an orderly manner. Beijing is also the first city in China to have an operational subway system, and the abundant smart card data is crucial for identifying urban mobility. The spatial distribution of Beijing’s subway stations and lines is shown in Fig. 2, with CAC, BMC, and the NTs constituting seven research areas.

3.2 Data and variables

3.2.1 Data description

This study utilizes multi-source heterogeneous big data from the city, including seven types: smart card data, administrative division and planning zone data, transportation infrastructure data, road network data, land use data, Point of Interest (POI) data and urban building data.

Smart card data is collected through the Automatic Fare Collection system (AFC), which includes data from both bus smart cards (BSC) and subway smart cards (SSC). With a utilization rate of over 90%, it can effectively represent urban mobility (Ma et al., 2018). The raw data is provided by the Beijing Smart Transportation Development Center, including 19 fields such as serial number, IC card number, OD point mode, route of OD point, direction of OD point, station number of OD point, station name of OD point, latitude and longitude of OD point, departure and arrival time (see Appendix A.). The raw data suffered from geographic location anomalies, swipe time anomalies, order time anomalies, missing values, redundancy, duplicate values, etc., and finally, 90,219,797 valid data were obtained after data cleansing, of which 57,282,360 were BSC entries, and 32,937,437 were SSC entries. The time interval for the smart card data is from May 1 to 12, 2019, spanning 12 days. Four types of CEs were involved in the study period (see Fig. 3): Holiday (Labour Day), Have to work, Workday, and Weekend. China’s Labour Day is on May 1st every year. According to the General Office of the State Council’s Notice on Adjusting the Labour Day Holiday Arrangement in 2019, the Labour Day holiday in 2019 was adjusted from May 1st to 4th, totaling 4 days. Sunday, May 5th, 2019, was adjusted to be a workday, as shown in Fig. 3.

Beijing’s administrative divisions’ base geographic background map data is sourced from the National Geographic Information Resources Catalog Service System. Subway and bus services have been operationalized in all seven districts of this study, with mature station operations. The research period is more than four months after the opening of new subway and bus lines, allowing for an accurate representation of the independent variables’ indicators. Beijing’s planning and zoning data are obtained from the Beijing Municipal Planning and Natural Resources Commission. Specifically, the boundary of CAC is defined according to the Regulatory detailed planning of the core functional area of the capital (2018—2035), BMC’s boundary follows the Regulatory detailed planning of Beijing Municipal Administrative Center (2016—2035), and the boundaries of NTs are based on the Territorial Spatial Planning Documentation (2017—2035) for Changping District (CP), Shunyi District (SY), Tongzhou District (TZ), Daxing District (DX), and Fangshan District (FS).

Population and economic data are sourced from the Beijing Statistical Yearbook (Beijing Municipal Bureau of Statistics, 2020). Transportation infrastructure data are from one of China’s leading mapping service platforms, Amap, and the official website of the Beijing Subway Company. Road network data are obtained from OpenStreetMap. Land use data are from the paper (Gong et al., 2020).

POI data, representing non-geographical points on electronic maps, provide more detailed information compared to traditional land use data, offering a more precise basis for inferring passenger travel purposes. Numerous studies have utilized POI datasets, including malls, schools, banks, and bus stations, to reflect the density and mix of urban functions. In this study, a POIs dataset for Beijing was acquired from Amap in 2019, containing nearly 800,000 data points.

3.2.2 Dependent variables

The dependent variable is the hourly time-accurate subway outbound passenger flow generated from the 800 m × 800 m grid analysis cell. Based on TransBigData (Yu and Yuan, 2022) in the Python library, smart card data can be divided into subway travel data and bus travel data according to the “travel mode (f_mode and t_mode)” field. According to the “card number (card_id)” field and the “time (f_tm and t_tm)” field in ascending order, the OD data of the passengers who traveled by bus or subway are extracted as UM.

Figure 4 shows the urban mobility based on the passenger travel BSC and SSC data, including the magnitude of urban mobility under four calendar events: Holiday, Have to work, Workday, and Weekend. The trends of passenger flow changes for both bus and subway are the same, and the total volume of passenger flow more intuitively displays the apparent line changes among the four calendar events. Urban mobility is the highest on Workdays, significantly lower during Holidays, and the least on Weekends. It is worth noting that during the holiday, urban mobility initially decreases sharply over time and rises slowly on the last day. On Workdays, urban mobility is the lowest on the first day (Have to work) and the highest on the last day (the Workday before the Weekend), with travel demand being more concentrated on the day before the Weekend than on holidays.

Figure 5 illustrates the spatial distribution of UM based on passenger travel SSC data, revealing significant spatial and temporal variations. The flow lines are larger in the west and northeast directions on Holiday (Fig. 5(a)), which is attributed to the tourist flow to Fragrant Hills and the passenger dispersal from Capital Airport within SYNT in the northeast direction. There is a significant increase in UM within the CAC on Have to work, with multiple distinct hotspots (Fig. 5(b)), mainly at four train stations within Beijing. On Have to work, there are fewer hotspots in the city due to the decrease in the number of passengers returning. However, there are still larger commuter and air distribution flows in Xierqi-Wudaokou, CAC-Capital Airport, and CAC-BMC (Fig. 5(c)). Weekend is influenced by the demand for traveling or returning to the home towns, and there are larger flows in Capital Airport and the four train stations in the city (Fig. 5(d)).

3.2.3 Independent variables

CEs had a significant effect on UM (Fig. 5), but there was a threshold effect of spatio-temporal heterogeneity in the effect of BE on UM, and there was an interaction between the two. Therefore, when choosing variables, we stratify CEs as moderating variables rather than directly as independent variables. The dependent variables were categorized into four categories according to the CEs in 3.2.1: Labour Day, Have to work, Workday and Weekend. The independent variables were categorized into six categories: Centrality, Accessibility, Intensity, Diversity, Vitality, and Comfort, according to the characteristics of UM. Variables with Pearson correlation coefficients higher than 0.7 and Variance Inflation Factors (VIFs) more than 10 were considered to have multicollinearity (Qian and Ukkusuri, 2015). After the test for multicollinearity, Parking Density was excluded from the analysis due to extremely high VIF values, and 20 candidate variables were finally identified to study the effect of UM, as shown in Table 1.

Centrality includes Distance to the city center, Degree Centrality, Betweenness Centrality, Closeness Centrality, Eigenvector Centrality, and PageRank (Li et al., 2024), which are used to represent the distance and importance of nodes in the subway network from the city center. Accessibility includes Transfer Lines, Road Density, and Parking Density, reflecting the characteristics of transportation infrastructure services (Wang et al., 2020). Intensity and Diversity include floor area ratio (FAR), building density (BD), functional density (FD), and functional mix (FM). According to the control hierarchy of Beijing subway stations, these are divided into two levels: Regulated Area and Study Area. Regulated Area is defined as a 300 m radius around CAC and BMC stations and a 500 m radius around NTs stations; Study Area is defined as an 800 m radius around CAC and BMC stations and a 1000 m radius around NTs stations (He, 2023). To avoid robust covariance of the inflow and outflow of public transport and subway stations, we constructed the ratio of the passenger flow of public transport in and out of the station within 50 m of the buffer zone of the subway station, whose ratio is Inflow/(Outflow + 1). Thus, vitality includes subway inflow/outflow and bus inflow/outflow. Comfort includes the Peak Hour and Balance Coefficient (Munch and Proulhac, 2023).

3.3 Methodology

3.3.1 Nonlinear relationship analysis model

The Gradient Boosting Decision Tree (GBDT) algorithm is renowned for its effectiveness in revealing complex nonlinear relationships in built environments (Zhang et al., 2024). GBDT is an iterative decision tree algorithm that constructs a set of weak learners and accumulates the results of multiple decision trees as the final prediction output (Friedman, 2001). GBDT excels at modeling complex nonlinear relationships and has good generalization and expression abilities (Yang et al., 2022). It also has good robustness and does not require special preprocessing of data (Cao and Tao, 2023). It is often used to study the nonlinear relationship between BE and travel behavior (Lang et al., 2024; Tao et al., 2020).

The Extreme Gradient Boosting (XGBoost) algorithm is a machine-learning method of GBDT. Compared to traditional GBDT, XGBoost significantly improves speed, memory utilization, and parallelization and is widely used in largescale datasets (C. Chen et al., 2021). The loss function is used to measure the error between predicted and actual values:

\begin{array}{}\text{(1)}& Obj(\theta )={\displaystyle \sum _{i=1}^{n}l(y_i\mathrm{,}{\widehat{y}}_i)}+{\displaystyle \sum _{k=1}^T\Omega (f_k)}\mathrm{.}\end{array}

In Eq. (1), l(y_i,{\widehat{y}}_i) is the error loss function, which measures the difference between observed y_i and predicted values {\widehat{y}}_i; n represents the number of samples used to train the model; Ω(f_k) is a regularization term for model complexity that can suppress model overfitting; T is the number of leaf nodes.

3.3.2 Explainable machine learning methods

The more complex the traditional machine learning method model, the better the predictive performance, but the poorer the interpretability. Ensemble learning algorithms such as GBDT and XGBoost have poor interpretability. At the same time, SHapley Additive exPlans (SHAP) can calculate the marginal contribution of features to the model output, and then explain the black box model from both global and local perspectives (Yang et al., 2022). Therefore, this study adopts SHAP to explain the BE-UM relationship model.

SHAP considers all features as contributors and uses Shapley values from game theory to calculate marginal contributions, constructing an interpretable machine learning model. The calculation formula for the Shapley value is as follows:

\begin{array}{}\text{(2)}& {\varphi }_i={\displaystyle \sum _{S\subseteq N\backslash \left\{i\right\}}\frac{\left|S\right|!(|M|-|S|-1)!}{\left|M\right|!}}\cdot [f(S\cup \left\{i\right\})-f(S)]\mathrm{.}\end{array}

In Eq. (2), ϕ_i is the Shapley value of the i-th feature; S is a subset of features that does not include feature i; M is the total number of features; f(S) indicates that only the features of subset S are used to predict the model output for the sample.

A large number of studies on the nonlinear relationship of BE (Peng et al., 2023) have used partial dependence plots (PDP) to illustrate the influence of independent variables on the dependent variable but lack explanations for local effects (Lei et al., 2024). SHAP analyzes the BE‒UM relationship from the perspective of global feature importance and provides the local interpretability of individual samples and the contribution of each feature to the predicted value. For this, we calculated the average Shapley value of each feature in all samples:

\begin{array}{}\text{(3)}& I_j=\frac{1}{N}{\displaystyle \sum _{i=1}^N\left|{\varphi }_j^{(i)}\right|}\mathrm{.}\end{array}

In Eq. (3), I_j is the importance of feature j; N is the total number of samples; ϕ_j⁽ⁱ⁾ is the Shapley value of feature j corresponding to the i-th sample.

In addition, SHAP dependency graph visualization of feature interaction can be used to explore how BE spatiotemporal heterogeneity affects urban mobility. The calculation of SHAP feature interaction values is as follows:

\begin{array}{}\text{(4)}& {\varphi }_{i\mathrm{,}j}={\displaystyle \sum _{S\subseteq M\backslash \left\{i\mathrm{,}j\right\}}\frac{\left|S\right|!(|M|-|S|-2)!}{2|M-1|!}{\delta }_{ij}(S)}\mathrm{.}\end{array}

In Eq. (4), ϕ_i,j represents the SHAP feature interaction value between different features i and j; δ_ij(S) = f(SU{i,j}) - f(SU{i}) - f(SU{j}) + f(S); The remaining parameters are the same as Eq. (2).

4 Analysis results

4.1 Performance comparison of models

The experiment was conducted in a Python environment, and a preliminary prediction model was established using the XGBoost algorithm based on the variables of CAC and NTs. We used 5-fold cross-validation to evaluate the performance and generalization ability of the model (Tao and Cao, 2023). We employed GridSearch to select the optimal hyperparameters of the model (Lu et al., 2023), as shown in Appendix B. In this study, grid search was used to optimize three key parameters of the prediction model: n_estimators (number of trees) control the complexity of the model, learning_rate (learning rate) controls the step size at each update, and max_depth (maximum depth per tree) controls the complexity and overfitting risk of the model.

OLS (the ordinary least squares) is a linear regression model that assumes a linear relationship between the independent and dependent variables. XGBoost can capture nonlinear relationships. Therefore, it performs better on urban multi-source heterogeneous big data. We use three evaluation metrics, RMSE, MAE, and R², to assess the predictive model. RMSE (Root Mean Squared Error) measures the degree of deviation between predicted values and actual values, with smaller values indicating more accurate model predictions; MAE (Mean Absolute Error) further reflects the difference between the predicted value and the actual value of the model, with smaller values indicating more accurate model predictions; R²(R-Square) represents the explanatory power of the model on the target variable, and the closer the value is to 1, the better the fitting effect of the model. By comparing three evaluation indicators, the study found that the XGBoost model has a higher R², lower MAE, and RMSE (Table 2), indicating that XGBoost is superior to OLS in terms of fitting degree and prediction accuracy and is suitable for analyzing the impact of BE on UM.

4.2 Relative importance of independent variables

Table 3 shows the SHAP importance values for each variable, which reflect the importance of each variable in the overall model prediction. Generally speaking, the higher the SHAP importance value, the more significant the impact of the variable on the prediction results. From a temporal perspective, D0, D4, and D6 have the highest value for Have to Work, D1, D3, and D5 have the highest importance value for Labour Day, and D2 has the highest value for Workday. From a spatial perspective, the FAR of the Study Area has the highest importance value for D0 and D4; Eigenvector Centrality has the highest importance value for D1, D3, and D5; Transfer Lines have the highest importance value for D2, and Degree Centrality has the highest importance value for D6. It can be seen that FAR of Study Area and Eigenvector Centrality play a more important role in predicting UM in most of the cases, especially for Have to work in CAC and Labour Day in NTs importance is greater.

Figure 6 further illustrates the spatiotemporal heterogeneity, aiming to reveal the relative importance differences of each independent variable in BE for prediction. The bar chart on the left side of Fig. 6 represents global importance. A variable with a long bar significantly impacts the overall prediction results. In the BE variable, FAR of Study Area, PageRank, Transfer Lines, and Betweenness Centrality consistently show high importance, suggesting that volumetric and centrality metrics significantly impact UM. In addition to this, Degree Centrality of Study Area and Closeness Centrality for CAC (Fig. 6(a)); Distance to the city center for SYNT (Fig. 6(d)); Road Density for BMC (Fig. 6(e)); YZNT and FSNT’s Closeness Centrality (Fig. 6(f) and 6(h)) are also more influential. BE variables also have significant temporal heterogeneity: e.g., FAR of Study Area becomes more important in Have to work, and Eigenvector Centrality becomes more important in Labour Day, highlighting the fact that in workday, efficiency is more important. The BE variables also have significant temporal heterogeneity: for example, FAR of Study Area becomes more important in Have to work, and Eigenvector Centrality becomes more important in Labour Day, which highlights the dynamic evolution pattern that the efficiency orientation on working days is more influential on urban spatial functions, and the social orientation on holidays is more influential on the structure of the social network, which provides a key basis for the resilient planning and spatiotemporal adaptation of urban design.

The point graph on the right side of Fig. 6 represents local contributions, with each point representing a sample. SHAP values greater than 0 indicate a positive impact on the prediction, while negative values indicate the opposite. For example, FAR of Study Area (Fig. 4(a)) has high red points on the right side, indicating a positive correlation with UM. BD of Regulated Area, Bus Inflow/Outflow, and Subway Inflow/Outflow (Fig. 5(a) and 5(b)) have high-value points on the left side, indicating a negative correlation between CAC and UM.

4.3 Nonlinear relationship of variables

The SHAP dependence plot reveals fine-grained nonlinear relationships both locally and globally, as well as potential threshold effects, thereby aiding researchers in profoundly understanding the complex interactions among variables within the predictive model. Although this study involved 20 variables, the analysis primarily focuses on a subset of important and intriguing variables. Figure 7 selects the SHAP dependence plots for four variables in CAC.

Figures 7(a), 7(e), 7(i), 7(m) depict how Closeness Centrality in the Centrality metric affects UM, revealing nonlinear changes under the four CEs types. Under most calendar events of CAC (Fig. 7(a), 7(e), 7(m)), UM decreases with increasing centrality when Closeness Centrality is less than 0.086; however, UM increases with increasing centrality when Closeness Centrality is greater than the turning point of 0.086. Although there is a slow increase in UM with increasing centrality on workday (Fig. 7(i)) when Closeness Centrality is less than 0.084, the other trends are the same as for other calendar events. For the subway station in a relatively marginal location, its function is mainly residential or single-function area. The population stays and reduces the flow after arrival, which leads to the centrality enhancement to inhibit the UM instead. The centrality enhancement can promote UM growth only when the Closeness Centrality breaks through the critical value of 0.086.

Figures 7(b), 7(f), 7(j), 7(n) depict the relationship between FAR of Study Area and UM in the Intensity metric. Under most calendar events in the CAC (Figs. 7(f), 7(j), 7 (n)), FAR of Study Area and UM show a positive correlation trend. This is the same as the common sense pattern: higher FAR usually means more jobs, commercial facilities, or public services are concentrated and more attractive to foot traffic. However, in Labour Day (Fig. 7(b)), when FAR of Study Area is less than 2.0, UM decreases with the increase of floor area ratio; and UM increases with the increase of floor area ratio only after AR of Study Area is greater than 2.0. This phenomenon reveals that BE and UM present a threshold response mechanism on specific holidays, reflecting the special sensitivity of the behavioral pattern of the Labour Day crowd to the volumetric rate. The areas with low floor area ratios are mostly old neighborhoods and ordinary commercial streets, and even if the FAR of Study Area is raised, it is more likely to bring traffic congestion, exposing the functional vulnerability and spatial failure. Only when the FAR of Study Area breaks through 2.0 can the station area resilience be raised to cope flexibly with the pressure of the holiday crowds.

Figures 7(c), (g), 7(k), 7(o) illustrate the relationship between FD of Study Area and UM in the Diversity metrics, which shows an overall positive correlation trend under the four categories of CEs in the CAC, albeit with some fluctuations. On Labour Day (Fig. 7(c)), when FD of Study Area is less than 2100, UM decreases slowly thereafter; however, when FD of Study Area is greater than 2100, UM increases significantly thereafter. The pattern is similar for Have to work and Weekend (Fig. 7(g), 7(o)), where UM decreases when FD of Study Area When FD of Study Area is smaller than 2300, UM increases; but when FD of Study Area is larger than 2300, UM decreases. On the other hand, in Workday (Fig. 7(k)), UM increases with the FD of Study Area, especially when FD of Study Area is greater than 1750. This phenomenon reveals a calendar event-driven dynamic threshold effect between urban functional density and UM, reflecting the matching mechanism between crowd behavior and spatial function under different CEs. On holidays, the attractiveness of scattered POIs is limited; instead, the crowd is siphoned off by high-density areas, and UM decreases subsequently; only after the POI density reaches the scale effect of 2100 UM increases subsequently. On transfer days and weekends, a small amount of POI growth can satisfy the demand for living and promote UM growth; instead, UM decreases instead of increasing after POI density redundancy. Due to rigid commuting demand, workdays drive a positive correlation between POI and UM.

Figures 7(d), (h), 7(l), 7(p) depict the relationship between Balance Coefficient and UM in the Comfort indicator. Under most calendar events of CAC (Figs. 7(d), 7(h), 7(p)), Balance Coefficient and UM show a negative trend. However, on Workday (Fig. 7(l)), UM increases with Balance Coefficient, and the upward trend is pronounced when Balance Coefficient is 1.5—2.7. In addition, UM also increases slowly with increasing Balance Coefficient in the case where Balance Coefficient is greater than 3.4 on Labor Day (Fig. 7(d)) and in the case where Balance Coefficient is greater than 2.4 in Have to work (Fig. 7(h)). This phenomenon reveals that the peaks and valleys of passenger flows under different CEs and the differences in trip purposes jointly shape the UM pattern. This indicates that the UM decreases in Beijing when Balance Coefficient is high during most hours, i.e., when the subway is congested during peak periods and empty during workdays. A weekday Balance Coefficient of 1.5—2.7 is a moderate peak-to-valley difference, and optimizing weekday peak-period capacity allocation can significantly improve UM. At the same time, the promotion effect is not apparent beyond the threshold range.

Figure 8 shows the SHAP dependence plots for four variables in NTs. Figures 8(a), 8(e), 8(i), 8(m) depict the nonlinear relationship between PageRank and UM in Centrality metrics, and the trends are all that initially UM decreases as PageRank increases, after reaching the turning point, and then increases as PageRank increases. At Have to work (Fig. 8(e)), the turning point is minimized to 0.0014; at Labour Day and Weekend (Fig. 8(a), 8(m)), the turning point is 0.0017; and at Workday, the turning point is 0.0019. The differences in PageRank turning points illustrate the changes in functional demand of the urban network under different CEs. The transfer day reflects a low demand for elastic travel demand for the importance of metro network nodes, moderate demand on holidays and weekends, and a higher threshold is needed on workdays to emphasize the importance of metro network nodes.

Figures 8(b), 8(f), 8(j), 8(n) illustrate how the FAR of Regulated Area in the Intensity metric affects the UM of NTs. On Labour Day and Weekend (Fig. 8(b), 8(n)), the FAR of Regulated Area is positively correlated with the UM, especially when the FAR of Regulated Area is less than 0.20, the subsequent rise in UM is greater. In Have to work (Fig. 8 (f)), the relationship between FAR of Regulated Area and UM is divided into four segments: when FAR of Regulated Area is less than 0.13, UM is negatively correlated; when FAR of Regulated Area is between 0.13 and 0.2, UM is positively correlated; when FAR of Regulated Area is between 0.2 and 0.3, UM is again negatively correlated; when FAR of Regulated Area is greater than 0.3, UM is again positively correlated. On Workday (Fig. 8(j)), FAR of Regulated Area is negatively correlated with UM, and when FAR of Regulated Area is less than 0.2, UM decreases even more. This phenomenon reveals that the FAR of Regulated Area promotes or suppresses UM depending on how well the NTs function matches the CEs scenario. On holidays and weekends, the increase in station area FAR leads to an increase in UM; on transfer days, a temporary mismatch of spatial functions occurs under flexible work, and different FARs have different impacts on UM; and on workdays, the increase in FAR leads to the separation of jobs and residences, and the UM continues to become lower.

Figures 8(c), 8(g), 8(k), 8(o) depict the nonlinear relationship between FD of Study Area and UM for the Diversity indicator. In Labour Day, Have to work, Weekend (Figs. 8 (c), 8(g), 8(o)), when FD of Study Area is less than 700, UM rises aggressively with it, and when FD of Study Area is greater than 700, UM rises slowly or decreases slightly. On Workday (Fig. 8(k)), UM decreases slightly when FD of Study Area is less than 500; UM elevates significantly when FD of Study Area is greater than 500. Boosting the number of POIs on non-working days attracts crowds to gather, and UM rises rapidly; however, saturated POI density causes congestion again, decreasing UM. The low POI density in NTs on workdays is mainly for a single residential function; thus, the UM is lower. This phenomenon reveals that the relationship between POI density and UM is mainly affected by workdays and non-workdays, that non-workdays have a greater demand for the number of POIs, and that a general FD of Study Area of 700 or less can significantly increase the UM of non-workdays.

Figures 8(d), 8(h), 8(l), 8(p) illustrate the relationship between Bus Inflow/Outflow and UM for the vitality metrics, and at all CEs, UM increases as Bus Inflow/Outflow increases, and this tendency is particularly evident when Bus Inflow/Outflow is between 0 and 2. However, on Labour Day, Have to work, Weekend (Figs. 8(d), 8(h), 8(p)), when Bus Inflow/Outflow exceeds 2, UM decreases subsequently. This indicates that the Inflow/Outflow of NTs in Beijing mostly lies between 0 and 2, and the bus and subway show a complementary relationship in most cases. However, on non-working days, excessive concentration of bus service or regional congestion causes people to choose other travel modes or avoid the area, reducing the number of subway outflows and decreasing the UM.

5 Discussion and conclusion

This study systematically deconstructs the nonlinear threshold effect and spatio-temporal heterogeneity of BE-UM through an interpretable machine learning framework, revealing the limitations of traditional linear models. It is found that the UM driving mechanisms of CACs and NTs are significantly differentiated, and there is a clear spatiotemporal threshold for the influence of BE indicators, which provides a quantitative control path to solve the contradiction between peak congestion on workdays and resource idleness on non-workdays. The study advances the development of UM theory. It verifies the necessity of differentiated planning through the closed loop of “model prediction-attribution explanation-policy recommendation,” laying a scientific foundation for transforming the innovative city governance paradigm.

5.1 Discussion

Based on UM pattern recognition, this study interprets the BE-UM model by SHAP dependency Figue, which reveals the asymmetric threshold effect between BE variables and UM, breaking through the traditional linear association assumption. The XGBoost model combined with lattice search optimization shows significant advantages: compared with the OLS model, R² is improved, both MAE and RMSE are reduced, and the five-fold cross-validation confirms its stable generalization ability. It is consistent with research on modifications and improvements to traditional machine learning algorithms (E. Chen et al., 2021; Lei et al., 2024). Smart card data based on passenger travel shows that there are significant spatiotemporal differentiation features of UM, and the resulting analytical framework of CEs and NTs quantifies the differentiated response laws of different NTs under vacation travel, transfer to work, daily commuting, and weekend breaks, which provides a theoretical basis for the regulation of UM under multi-scenarios.

It is found that there is significant spatial and temporal variability in BE impact UM and that the volume ratio and network centrality are the core elements predicting UM, especially in the CAC of Have to work and the NTs of Labour Day. In the CAC of Have to work, the areas with high volume ratios are directly driven by the UM due to the intensification of functionality, which reinforces the efficiency orientation of commuting, and in the NTs of Labour Day NTs, transportation hubs, shopping districts, and other metro stations with high network centrality dominate UM enhancement due to stronger diffusion of social activities. In addition, variables in the Centrality metrics stably influence UM in various CEs, such as Betweenness Centrality, Closeness Centrality, Eigenvector Centrality, and PageRank.

Although several studies have examined the effect of POI density on UM at different times (Chen et al., 2024; Du et al., 2023; Güller, 2025), these studies usually focus on intraday differences and ignore the interaction effects under CEs and different regions. These studies have emphasized the positive correlation between POIs and UM. However, the present study further found that the effect of FD of Study Area on UM is characterized by temporal and spatial differentiation: off-days are more sensitive to the density of POIs and have a greater demand for the number of POIs. The scale effect of CAC with FD of Study Area of 2100 or more and NTs with FD of Study Area of 700 or less can significantly enhance UM.

In addition, the analysis results of Vitality and Comfort indicators show that although NTs have mostly complementary relationships between buses and subways, Beijing’s subway traffic congestion is mostly during peak periods. Subway traffic is lower during weekday peak periods and non-workday periods, and resources are idle. This difference is attributed to different patterns of human activity during weekday peak periods and other CEs.

Our study also emphasizes a general nonlinear threshold effect in the BE-UM relationship. The threshold effect is defined as the fact that the UM characteristics of BE are completely different under different spatial and temporal constraints, which points to the potential pitfalls of the widespread implementation of “one-size-fits-all” land use development and public transportation planning in China’s NTs (Luo et al., 2025). It has long been assumed that a high plot ratio and high-density development patterns in metro station areas can promote UM. However, the results of this study show that station area resilience can only be enhanced, and UM growth is promoted after the FAR of Study Area exceeds 2.0 in CAC. The Centrality metric has the same threshold effect, with Closeness Centrality exceeding 0.086 in CAC and PageRank exceeding 0.0017 in NTs before UM is promoted. 0.0017 before it can promote UM increase. Therefore, we suggest that urban planners consider the thresholds of metro Centrality and station area volume ratio (300 m buffer for city centers and 500 m buffer for NTs) as a reference and reach a certain threshold to significantly promote the travel of crowds while too low or too high thresholds may result in waste of resources and urban congestion.

5.2 Conclusion

UM can alleviate the pressure of traffic tides, improve urban efficiency and convenience, enhance urban resilience and vitality, and thus improve residents’ quality of life and happiness. This study systematically reveals the spatio-temporal divergence law and nonlinear threshold effect of BE on UM through an interpretable machine learning approach that breaks through the traditional linear association assumption. The results show that (1) combining XGBoost and SHAP interpretable framework, the model prediction performance significantly outperforms the traditional OLS, which provides a quantitative tool and theoretical basis for UM regulation under different CEs and NTs. (2) CACs enhance commuting efficiency due to functional intensification and high volumetric ratio; NTs rely on metro network node centrality to promote UM enhancement. Non-working day UM is more sensitive to POI density, and the FD of Study Area thresholds differ significantly, with CAC greater than 2100 and NTs less than 700. (3) Among the core metrics, the volumetric ratio is related to network centrality (e.g., Betweenness Centrality, Closeness Centrality, Eigenvector Centrality, PageRank, etc.) have clear thresholds under different spatio-temporal constraints. The volume ratio of CAC needs to break through 2.0, Closeness Centrality breaks through 0.086, and PageRank of NTs breaks through 0.0017 before UM can be significantly improved. (4) Peak congestion and peak idleness are contradictory in the Beijing metro. The contradiction between peak congestion and peak idleness stems from separating occupations and residences and the difference in activity patterns. Optimizing the planning based on the threshold effect is suggested to avoid “one-size-fits-all” development. These findings advance the development of complex urban systems theory and provide a new data-driven decision-making paradigm for innovative city governance.

Further research can be conducted to deepen the data dimension and spatial generalizability. In terms of data, although the data can meet the research needs, under the more diverse travel demand and fine-grained spatiotemporal pattern mining requirements, other aspects of urban traffic can be considered, such as travel time, travel mode choice, or travel purpose, while also taking into account exogenous variables such as weather factors and emergency events. Regarding spatial generalization, the study focused on the CAC and NTs regions, and subsequent comparative studies between NTs and different regions of the citywide domain can be carried out to verify the model migration capability. Meanwhile, in the future, we can integrate other public transportation travel data, such as shared bicycles, taxis, and ride-hailing vehicles, and other data, such as mobile phone signaling, to break through the limitation of single-swipe card data on characterizing complex travel behaviors.

References

[1] Abideen, Z.U., Sun, X., Sun, C., 2025. The multi-module joint modeling approach: predicting urban crowd flow by integrating spatial—temporal patterns and dynamic periodic relationship. Eng. Appl. Artif. Intell. 141, 109721.

[2] Acevedo, H., Porta, S., 2025. Network structure and urban mobility sustainability: a topological analysis of cities from the urban mobility readiness index. Sustain. Cities Soc. 119, 106076.

[3] Achar, A., Dhivya, B., Anil, K.B., Lelitha, V., 2020. Bus arrival time prediction: a spatial kalman filter approach. IEEE Trans. Intell. Transport. Syst. 21 (3), 1298—1307.

[4] Ariffin, K.S., Salama, M.A., Mustapha, A., Mahdin, H., Lim, D., Mahmoud, M.A., Mohammed, M.A., Al-Rimy, B., Fudzee, M., Yang, T.J., 2021. Review of data fusion methods for real-time and multi-sensor traffic flow analysis. IEEE Access 9, 51258—51276.

[5] Bao, Z., Ng, S.T., Yu, G., Zhang, X., Ou, Y., 2023. The effect of the built environment on spatial-temporal pattern of traffic congestion in a satellite city in emerging economies. Dev. Built Environ. 14, 100173.

[6] Broughton, M., Broughton, H., Yu, L., Hu, L., 2003. New City Development and Construction in the UK. Urban Plan, pp. 78—81 (in Chinese) (12).

[7] Cao, J., Tao, T., 2023. Using machine-learning models to understand nonlinear relationships between land use and travel. Transport. Res. Transport Environ. 123, 103930.

[8] Cervero, R., Kockelman, K., 1997. Travel demand and the 3Ds: density, diversity, and design. Transport. Res. Transport Environ. 2 (3), 199—219.

[9] Chang, X., Wu, J., Yu, J., Liu, T., Yan, X., Lee, D.H., 2024. Addressing COVID-induced changes in spatiotemporal travel mobility and community structure utilizing trip data: an innovative graph-based deep learning approach. Transp. Res. Part A Policy Pract. 180, 103973.

[10] Chen, C., Feng, T., Ding, C., Yu, B., Yao, B., 2021. Examining the spatial-temporal relationship between urban built environment and taxi ridership: results of a semi-parametric GWPR model. J. Transport Geogr. 96, 103172.

[11] Chen, E., Ye, Z., Wu, H., 2021. Nonlinear effects of built environment on intermodal transit trips considering spatial heterogeneity. Transport. Res. Transport Environ. 90, 104613.

[12] Chen, J., Li, W., Zhang, H., Cai, Z., Sui, Y., Long, Y., Shibasaki, R., 2021. GPS data in urban online ride-hailing: a simulation method to evaluate impact of user scale on emission performance of system. J. Clean. Prod. 287.

[13] Chen, M., Cheng, Y., Zhang, S., Chen, K., Wei, J., Hu, X., 2024. Catchment-area delineation approach considering travel purposes for station-level ridership prediction task. Transp. Res. Rec. 2678 (5), 397—415.

[14] Ding, C., Cao, X., Liu, C., 2019. How does the station-area built environment influence metrorail ridership? Using gradient boosting decision trees to identify non-linear thresholds. J. Transport Geogr. 77, 70—78.

[15] Du, Q., Huang, Y., Zhou, Y., Guo, X., Bai, L., 2023. Impacts of a new urban rail transit line and its interactions with land use on the ridership of existing stations. Cities 141, 104506.

[16] Friedman, J.H., 2001. Greedy function approximation: a gradient boosting machine. Ann. Stat. 29 (5), 1189—1232.

[17] Gao, F., Deng, X., Liao, S., Liu, Y., Li, H., Li, G., Chen, W., 2023. Portraying business district vibrancy with mobile phone data and optimal parameters-based geographical detector model. Sustain. Cities Soc. 96, 104635.

[18] Gong, P., Chen, B., Li, X., Liu, H., Wang, J., Bai, Y., et al., 2020. Mapping essential urban land use categories in China (EULUC-China): preliminary results for 2018. Sci. Bull. 65 (3), 182—187.

[19] Güller, C., 2025. Investigating the relationship between built environment and spatiotemporal heterogeneity of metro ridership. Travel Behav. Soc. 40, 101053.

[20] He, K., 2023. Research on integrated planning and control method of rail transit integrated with detailed planning: a case study of Beijing. Urban Dev. Res 30 (8), 11—17 (in Chinese).

[21] Howard, E., 2003. Garden cities of to-morrow. Organ. Environ. 16 (1), 98—107.

[22] Kim, J., Jang, K., Shim, J., 2024. Factors influencing bus-to-subway transfer duration at subway stations: evidence from large-scale smart card data in Seoul. J. Transport Geogr. 120, 103969.

[23] Kim, K., 2020. Effects of weather and calendar events on mode-choice behaviors for public transportation. J. Transport. Eng. Part A Syst. 146 (7), 04020056.

[24] Klomp, R., Borst, C., van Paassen, R., Mulder, M., 2016. Expertise level, control strategies, and robustness in future air traffic control decision aiding. IEEE Trans. Hum.-Mach. Syst. 46 (2), 255—266.

[25] Kong, H., Zhang, X., Zhao, J., 2020. How does ridesourcing substitute for public transit? A geospatial perspective in Chengdu, China. J. Transport Geogr. 86, 102769.

[26] Lang, Y., Sho, K., Eom, S., Nishi, H., Hasegawa, D., Zhao, H., Aoki, T., Zhu, J., Matsuo, K., Masumura, A., 2024. Changes in visitor behaviour across COVID-19 pandemic: unveiling urban visitation dynamics and non-linear relationships with the built environment using mobile big data. Habitat Int. 154, 103216.

[27] Lei, J., He, M., Shi, Z., He, M., Liu, Y., Qian, Q., Qian, H., 2024. How does the built environment affect intermodal transit demand across different spatiotemporal contexts? J. Transport Geogr. 121, 104033.

[28] Li, D., 2014. Coordinated Development of Rail Transit and New Cities from the Perspective of Multi Center in Metropolitan Areas. Tianjin University (in Chinese).

[29] Li, S., Chen, P., Hui, F., Gong, M., 2024. Evaluating urban vitality and resilience under the influence of the COVID-19 pandemic from a mobility perspective: a case study in Shenzhen, China. J. Transport Geogr. 117, 103886.

[30] Li, S., Zhao, P., 2022. Examining commuting disparities across different types of new towns and different income groups: evidence from Beijing, China. Habitat Int. 124, 102558.

[31] Li, Z., Gao, F., Hao, J., Liang, J., Han, C., Tang, J., 2025. Exploring spatiotemporal characteristics of ride-hailing ridership connecting with metro stations: a comparative analysis of holidays, weekdays, and weekends. J. Transport Geogr. 123, 104104.

[32] Lin, Y., Shui, W., Li, Z., Huang, S., Wu, K., Sun, X., Liang, J., 2021. Green space optimization for rural vitality: insights for planning and policy. Land Use Policy 108 (9), 105545.

[33] Liu, X., Wang, M., Qiang, W., Wu, K., Wang, X., 2020. Urban form, shrinking cities, and residential carbon emissions: evidence from Chinese city-regions. Appl. Energy 261, 114409.

[34] Liu, Z., Liu, J., Hu, R., Yang, B., Huang, X., Yang, L., 2023. Calendar events’ influence on the relationship between metro ridership and the built environment: a heterogeneous effect analysis in Shenzhen, China. Tunn. Undergr. Space Technol. 141, 105388.

[35] Lu, C., Gu, J., Lu, W., 2023. An improved attention-based deep learning approach for robust cooling load prediction: public building cases under diverse occupancy schedules. Sustain. Cities Soc. 96, 104679.

[36] Luo, Y., Li, B., Zhang, H., Kang, M., Su, S., 2025. Unlocking the nonlinear TOD-metro ridership relationship: a novel machine learning approach embedding spatiotemporal heterogeneity. J. Transport Geogr. 126, 104222.

[37] Ma, X., Zhang, J., Ding, C., Wang, Y., 2018. A geographically and temporally weighted regression model to explore the spatiotemporal influence of built environment on transit ridership. Comput. Environ. Urban Syst. 70, 113—124.

[38] Meng, Y., Xing, H., 2019. Exploring the relationship between landscape characteristics and urban vibrancy: a case study using morphology and review data. Cities 95, 102389.

[39] Munch, E., Proulhac, L., 2023. Is work hours’ flexibility really a solution to morning peak period congestion? Comparative analysis between Paris and San Francisco. J. Transport Geogr. 113, 103712.

[40] Nidhi, N., Lobiyal, D.K., 2022. Traffic flow prediction using support vector regression. Int. J. Inf. Technol. 14 (2), 1—8.

[41] Ouyang, J., Fan, H., Wang, L., Zhu, D., Yang, M., 2022. Revealing urban vibrancy stability based on human activity time-series. Sustain. Cities Soc. 85, 104053.

[42] Pei, T., Shu, H., Guo, S., Song, C., Chen, J., 2020. Spatial patterns of geographic flow: concepts and classifications. J. Earth Inf. Sci 22 (1), 30—40 (in Chinese).

[43] Peng, B., Zhang, Y., Li, C., Wang, T., Yuan, S., 2023. Nonlinear, threshold and synergistic effects of first/last-mile facilities on metro ridership. Transport. Res. Transport Environ. 121, 103856.

[44] Peng, J., Fu, X., Wu, C., Dai, Q., Yang, H., 2025. Comparative analysis of nonlinear impacts on the built environment within station areas with different metro ridership segments. Travel Behav. Soc. 38, 100898.

[45] Qian, X., Ukkusuri, S.V., 2015. Spatial variation of the urban taxi ridership using GPS data. Appl. Geogr. 59, 31—42.

[46] Sharath, M.N., Patnala, P.K., Mehran, B., Regehr, J., 2024. Effects of Covid-19 pandemic restrictions on zonal transit demand: evidence from a low-density city. Res. Transp. Bus. Manag. 57, 101234.

[47] Sharma, G., Patil, G.R., 2024. Urban spatial structure and equity for urban services through the lens of accessibility. Transp. Policy 146, 72—90.

[48] Shi, F., Yang, B., Li, M., 2023. An improved framework for assessing the impact of different urban development strategies on land cover and ecological quality changes — a case study from Nanjing Jiangbei new area, China. Ecol. Indic. 147, 109998.

[49] Su, H., Qi, Z., Wang, Q., 2025. Impacts of land use characteristics on extreme heat events: insights from explainable machine learning model. Sustain. Cities Soc. 120, 106139.

[50] Tao, S., Dai, T., Guo, Y., Wang, Y., Liu, B., Jiang, H., 2024. How do built environment characteristics influence bus use patterns across neighbourhood types in Beijing? A machine-learning analysis. Travel Behav. Soc. 35, 100756.

[51] Tao, T., Cao, J., 2023. Exploring nonlinear and collective influences of regional and local built environment characteristics on travel distances by mode. J. Transport Geogr. 109, 103599.

[52] Tao, T., Wu, X., Cao, J., Fan, Y., Das, K., Ramaswami, A., 2020. Exploring the nonlinear relationship between the built environment and active travel in the Twin Cities. J. Plann. Educ. Res. 43 (3), 0739456X2091576.

[53] Vainio, T., Karppi, I., Sankala, I., 2024. Perceptions and preferences of experienced security in mass urban events — what’s technology got to do with it? Cities 144, 104626.

[54] Vilhelmson, B., 2007. The use of the car-mobility dependencies of urban everyday life. In: Threats from Car Traffic to the Quality of Urban Life. Emerald Group Publishing Limited, pp. 143—164.

[55] Wang, S., Yu, D., Kwan, M.P., Zheng, L., Miao, H., Li, Y., 2020. The impacts of road network density on motor vehicle travel: an empirical study of Chinese cities based on network theory. Transp. Res. Part A Policy Pract. 132, 144—156.

[56] Wang, Z., Liu, Y., Luo, X., Tong, Z., An, R., 2023. Research on the nonlinear relationship between urban vitality and built environment based on multi-source data — taking the main urban area of Wuhan on weekends as an example. Prog. Geogr. Sci. (Geographical Science Progress, in Chinese) 42 (4), 716—729.

[57] Xiang, Y., Gobi, K.S., Ruiwei, L., 2022. Identify impacting factor for urban rail ridership from built environment spatial heterogeneity. Case Stud. Transp. Policy 10 (2), 1159—1171.

[58] Xu, B., 2022. China’s new towns in controversy: a literature review. J. Plann. Lit. 37 (2), 236—248.

[59] Xu, C., Li, Z., Wang, W., 2016. Short-term traffic flow prediction using a methodology based on autoregressive integrated moving average and genetic programming. Transport 31 (3), 343—358.

[60] Xu, Y., Chen, R., Du, H., Chen, M., Fu, C., Li, Y., 2025. Evaluation of green space influence on housing prices using machine learning and urban visual intelligence. Cities 158, 105661.

[61] Yabe, T., Bueno, B.G.B., Frank, M.R., Pentland, A., Moro, E., 2025. Behaviour-based dependency networks between places shape urban economic resilience. Nat. Hum. Behav. 9, 496—506.

[62] Yang, H., Zhang, Q., Helbich, M., Lu, Y., He, D., Ettema, D., Chen, L., 2022. Examining non-linear associations between built environments around workplace and adults’ walking behaviour in Shanghai, China. Transp. Res. Part A Policy Pract. 155, 234—246.

[63] Yang, L., Yu, B., Liang, Y., Lu, Y., Li, W., 2023. Time-varying and non-linear associations between metro ridership and the built environment. Tunn. Undergr. Space Technol. 132, 104931.

[64] Yang, Y., Sasaki, K., Cheng, L., Liu, X., 2022. Gender differences in active travel among older adults: non-linear built environment insights. Transport. Res. Transport Environ. 110, 103405.

[65] You, M., Guan, C., 2024. Does self-containment of spatial scale and land use function contribute to mitigate urban heat island effects? Lessons from new towns in Shanghai. Land Use Policy 146, 107323.

[66] Yu, Q., Yuan, J., 2022. TransBigData: a Python package developed for transportation spatio-temporal big data processing, analysis and visualization. J. Open Source Softw. 7 (71), 4021.

[67] Zeng, Q., Wu, H., Zhou, L., Huang, G., Li, Y., Dewancker, B.J., 2024. Toward pedestrian-friendly cities: nonlinear and interaction effects of building density on pedestrian volume. J. Transport Geogr. 119, 103954.

[68] Zhang, J., Cheng, X., 2024. Theoretical paradigm and spatial analysis platform of urban factor flow. In: Green Digital Intelligence, Improving Quality and Efficiency — Proceedings of the 2024 Annual Conference on Urban Transport Planning in China. Beijing Jianzhu University, p. 13.

[69] Zhang, P., Dong, Y., Ren, Z., Wang, G., Guo, Y., Wang, C., Ma, Z., 2023. Rapid urbanization and meteorological changes are reshaping the urban vegetation pattern in urban core area: a national 315-city study in China. Sci. Total Environ., 167269

[70] Zhang, Z., Zhao, L., Zhang, M., 2024. Exploring non-linear urban vibrancy dynamics in emerging new towns: a case study of the Wuhan metropolitan area. Sustain. Cities Soc. 112, 105580.

[71] Zhao, L., Zhang, M., Cheng, S., Fang, Y., Wang, S., Zhou, C., 2023. Investigate the effects of urban land use on PM2.5 concentration: an application of deep learning simulation. Build. Environ. 242, 110521.

[72] Zhao, P., Wang, H., Liu, Q., Yan, X.Y., Li, J., 2024. Author correction: unravelling the spatial directionality of urban mobility. Nat. Commun. 15 (1), 5724.

[73] Zhou, X., Yeh, A.G.O., Yue, Y., Li, W., 2022. Residential-employment mixed use and jobs-housing balance: a case study of Shenzhen, China. Land Use Policy 119, 106201.

Rights & permissions

2095-2635/2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd.