An Interpretable Machine Learning Model-Based Study of the Relationship between Built Environment and Residents' Daily Travel Distance

Abstract

Urban sustainability and reduced carbon emissions have become critical challenges in contemporary urban development, with residents' daily travel distances significantly impacting both the environmental and social aspects of city life. Despite extensive research on travel behavior, Crossroads Density: Number of Crossroads within 500m the complex relationship between the built environment and travel behavior remains inadequately understood due to the limitations of traditional linear analyses. To address this gap, this study introduces an analytical approach grounded in an interpretable machine learning framework to investigate the pivotal environmental factors influencing residents' daily travel distances and to understand their underlying mechanisms. Initially, a variety of machine learning regression models were constructed to evaluate their predictive efficacy regarding residents' daily travel distances. Among these, the CatBoost model demonstrated the most accurate fit, highlighting the robustness of the methodology. Subsequently, the SHAP (SHapley Additive exPlanations) methodology was employed to ascertain the significance of each feature's contribution, thereby elucidating the extent and non-linear dynamics of various built environmental indicators on travel distances. The findings indicate that the POI (Point of Interest) diversity index, the proportion of main roads, the proximity to the city center, and other factors significantly influence residents' travel distances, exhibiting observable interactive effects. Importantly, these results underscore the implications of the built environment on travel behavior, offering actionable insights for urban planners. By focusing on the built environment's impact on travel distances, this study's interpretable machine learning framework delves into the multi-factorial non-linear correlations, providing a valuable tool for enhancing urban spatial planning, reducing travel distances, lowering traffic emissions, and thereby contributing to the sustainable development of cities.

Introduction

Different urban areas often have different shares of vehicle patterns (Ton, Bekhor et al., 2020) and different driving distances travelled (Scheiner, Frank et al., 2024). In urban planning, by determining and utilizing the characteristics of the residential built environment that promote shorter travel distances (Mondal, 2023), it can reduce travel distances and promote sustainable mode choices (Nikitas, Tsigdinos et al., 2021). This can further reduce emissions from urban transportation. Additionally, a shorter daily travel distance for residents can not only reduce energy consumption and environmental load, but also help to improve quality of life (Guo, Yang et al., 2023). A shorter travel distance means less investment of time and money costs, which helps improve the activity radius and convenience of life for residents. At the same time, shorter travel distances can reduce exposure to air pollution (Næss, Stefansdottir et al., 2021). Therefore, exploring the key factors affecting residents' daily travel distance is important for building a livable and sustainable urban environment.

Existing studies show that the built environment around urban residential areas has a significant impact on the traveling distance (Wang, C., Wang et al., 2022; Zhang, Yu, Wang et al., 2023).This research aims to analyze the relationship between the built environment and travel distances by achieving several objectives:

1.To understand the motivations and considerations influencing residents' travel choices and to explore how various built environments attract residents and affect their travel distances.

2.To investigate the interactions between different built environments and their effects on residents' travel distances.

3.By analyzing the relationship between daily travel distances and the built environment, this research aims to improve urban planning strategies.

Understanding the relationship between the built environment and residents' daily travel distances necessitates first identifying the key built environment variables that significantly influence travel behavior. This process involves a thorough review of existing literature to assess how these variables impact travel patterns. The use of interpretable machine learning techniques is essential for unravelling the complexities of this relationship, as it enables researchers to transparently examine the interactions among various built environment factors. Furthermore, interpretable machine learning provides a valuable framework for exploring the intricate connections between the built environment and travel distances. This approach not only elucidates how different built environment features shape travel behavior but also offers critical insights to assist urban planners in making informed decisions.

The rest of the paper is organized as follows: the next section reviews the existing literature on the relationship between the built environment and daily travel distances; Section 3 describes the methodology and data sources used in this study; Section 4 presents the results of our analysis of the relationship between built environment variables and daily travel distances; and Section 5 discusses the policy implications of our findings and potential limitations of the study.

Literature Review

The built environment, including street layout, land use, distribution of service facilities, and transportation infrastructure, has a profound impact on residents' travel behavior. The concept that the built environment profoundly influences individual travel behavior is now widely recognized (Fang, Homma et al., 2024). Scholars have introduced the "5D" theoretical framework, which categorizes the built environment into five distinct dimensions: Density, Diversity, Design, Distance to Transit, and Destination Accessibility. This framework has established a robust theoretical basis for research in this domain. Studies have shown that the characteristics of the built environment around a residential area, such as the degree of commercial facility agglomeration, significantly affect residents' travel distance. The average travel distance of residents tends to be longer in large cities with less pronounced clustering of facilities (Hipp, Lee et al., 2022). This may be related to the attractiveness of urban centers and the dispersion of suburbs (Indradjati, 2024). In addition, the accessibility of destinations, such as the distance to public transportation stops and the connectivity of walking paths, are key factors influencing travel distances (Zhang, Yiwen, Zhou et al., 2022). The influence of the built environment has been shown to be more important than socio-demographic factors in several studies (Yang, Zhou et al., 2024). For example, spatial mismatches, such as the distance between workplace and residence, as well as over-sized and inaccessible streets, can lead to an increase in travel distances and a rise in carbon emissions from travel (Yang, Zhou et al., 2024). The existence of these problems not only increases the travel costs of residents, but also poses a challenge to the sustainable development of cities. However, daily travel distance is often used as a dependent variable in research, and few studies have directly explored the factors influencing residents' daily travel distance. In addition, there is a lack of in-depth analysis of the interaction effects of the influencing factors and nonlinear relationships.

When analyzing the relationship between the built environment and daily travel characteristics, researchers can use a variety of methods. These include quantile regression models based on linear regression assumptions, such as use of multilayer linear regression models (Wang, Q., Liu et al., 2023) in exploring the impact of built environment on travel well-being ; logit models are commonly used to analyze the probabilistic impact of built environment factors on travel behavior, such as the study from Wu and Liu (2023) of the built environment choice of commuting mode decision-making for different housing types of spatial units; and the geographically weighted regression (GWR) model and its variant models can be used to consider spatial heterogeneity,a study (Liu, J., Wang et al., 2024) use of the model to analyze the factors affecting the use of shared bicycles in the built environment of cities, and the application of these models enhances our understanding of the motivations behind residents' travel choices (Cao, Yang et al., 2024). With the advantage of machine learning methods in multivariate fitting without the need to adhere to the assumption of linear relationships between variables, in recent years machine learning based methods such as categorical decision trees (Liu, T. and Ding, 2024)), and random forests have been applied to study such problems among others such as a study (Tamim Kashifi, Jamal et al., 2022) proposed a systematic machine learning framework for better understanding of mode choice decisions of travellers. They used LightGBDT model and SHAP analysis to improve the interpretability of the model and found that travel distance, traveler's age and annual income, number of cars/bikes owned, and travel density were the predictors that significantly influenced travel mode decisions. CatBoost (Wang, Y., Yao et al., 2024) made several optimizations on top of GBDT, especially in dealing with category-based features and missing values. with significant improvements. However, although machine learning achieves better results in terms of goodness-of-fit, the interpretability of variables still requires in-depth research (Ji, Li et al., 2019).The SHAP attribution analysis employs the Shapley value in game theory as an explanatory measure, as a study demonstrates (Li, Z., 2022) the utility of SHAP in real datasets and provides the opportunity to flexibly model, interpret, and visualize complex geographic phenomena and processes. At the same time, it has the advantages of strong global and local interpretability, fair distribution of variable contribution and excellent visualization effect, which make up for the poor interpretability defects of tree-based models (such as decision tree, advancing forest, etc.).

Upon reviewing the existing literature on the built environment and travel behavior, several research gaps are identified. Initially, there is a scarcity of studies that explicitly adopt daily travel distance as the dependent variable. Subsequently, there is a dearth of research examining the influence of the built environment on daily travel distances. Lastly, the study uniquely explores the interplay among different aspects of the built environment to evaluate their collective impact on travel behavior. To address these gaps, a machine learning regression model is employed to scrutinize the influence of the built environment on daily trips. Additionally, an explanatory machine learning model is utilized to dissect the varied effects of the built environment and to highlight key indicators that are pivotal to sustainable urban development. This approach aims to provide a comprehensive understanding of how the built environment shapes travel behavior and to identify areas for intervention that can promote more sustainable travel patterns.

Methodology

Technology roadmap

In the subsequent sections, the methodological approach and procedural steps for discerning the determinants that influence the extent of travel distance are delineated. The schematic representation of the modeling framework and analytical methodology is depicted in Figure 1. This figure sequentially illustrates the initial phase of outlining the empirical data procured, followed by an exposition on data preparation, selection, and pre-processing techniques. Thereafter, the construction of the machine learning model is detailed, along with the stratification of training and testing datasets. A comparative analysis of various machine learning models was conducted and rigorous evaluation protocols were executed to ensure the reliability of the findings. This included the assessment of performance metrics and the development of models that offer a high degree of interpretability.

Figure 1. The conceptual framework of this study

Study area and data

This study has designated the primary urban district of Fuzhou City as the focal research area, which has been further segmented into 775 distinct sub-areas for analysis. The dataset utilized in this research encompasses a full week of mobile phone signaling data collected in Fuzhou during the month of April, 2023. The weather during the data collection period was consistently sunny, with average temperatures ranging from 14 to 27 degrees Celsius, conditions that are conducive to travel and thus provide a robust context for assessing the influence of the built environment on daily travel patterns. Mobile phone signaling data Data obtained through China Unicom, the data is cleaned and does not contain personal information but only location point and interval time. The data whose starting point is within the scope of the study are screened out, and the excessive abnormal data are removed, and a total of 4.92 million pieces of data without sample expansion are obtained. The average daily travel distance for each residential point was derived from a clustering analysis that considered both residential and base station data. Population statistics were sourced from the sixth national census conducted by China. The POI data and the building area profile data were obtained through the Amap Application Programming Interface. The study incorporates 11 categories of POI, which include catering services, public facilities, companies, science and educational institutions, residential areas, life services, government and social organizations, accommodation services, transportation infrastructure, and sports, leisure, and shopping amenities. Furthermore, road network data were procured from OpenStreetMap (OSM), while information regarding the main roads and other thoroughfares was sourced from the public dataset provided by the Fuzhou municipal government.

Figure 2. Scope of the study

Selection of community built environment variables

The 5D dimensions—Density, Diversity, Design, Distance to Transit, and Destination Accessibility provide a comprehensive theoretical underpinning for studies within this field. These dimensions offer a multifaceted view of the built environment's characteristics and their subsequent impact on personal travel decisions. The Density dimension primarily addresses the spatial distribution and demographic concentration; Diversity encompasses the extent of land use mix and the balance between employment and residential areas; Design pertains to the micro-level elements such as the intricacies of street networks and the quality of pedestrian environments; while Distance to Transit and Destination Accessibility are closely linked to the ease of access to public transportation infrastructure. The judicious selection of quantifiable indicators for these dimensions, followed by their integration into an analytical model, can not only provide an accurate depiction of the built environment's current state within the study area but also elucidate the mechanisms by which environmental attributes shape travel behaviors. For example, a study (Li, P., Chen et al., 2024) took the metro system in Xi'an as the research object, constructed built environment indicators based on 5D elements, and found that the contribution of distance from the center of the city to the metro travel distance is the highest in the built environment .

In view of this, a series of built environment indicators are selected under the guidance of the “5D” theoretical framework and in accordance with the specific needs of the study. The selection is based on the previous research results and emphasizes the representativeness, accessibility and computational convenience of the indicators.The selection elements and criteria are shown in Table 1. Furthermore, the decision to limit the number of indicators was intentional, aimed at maintaining a balance between model complexity and interpretability. While other relevant variables, such as income levels and modes of transportation, could enrich the analysis, the choice to focus solely on the most pertinent indicators of the built environment allows for a more concentrated investigation of its impact. This approach provides a clear understanding of how the built environment influences travel behavior. The selected indicators cover different aspects of the five dimensions, ensuring they accurately reflect the real situation of the built environment in the study area. Additionally, to represent the built environment elements within the community context, a 500-meter buffer centered on the community was chosen to measure the built environment, providing reliable data support for analyzing the relationship between the built environment and travel behavior (Teixeira, Barbosa et al., 2023).

Table 1. Built environment variables and their meanings

Category	Symbol	Meaning	Definition	Reference
Density	Pop	Population Density	Population within a 500m radius	Ma, Duan et al. (2024)
	Building	Building Density	Area of buildings within a 500m radius	Han, Sun et al. (2020)
Diversity	POI_index	POI Mixture index	Shannon entropy of POI within 500m	Lv and Pan (2022)
Design	Crossroads	Crossroads Density	Number of Crossroads within 500m	Han, Sun et al. (2020)
	B_ratio	Main and Side Road Ratio	Ratio of main to side roads within 500m	Han, Sun et al. (2020)
Distance to transit	Bus_stop	Bus Stop Density	Number of bus stops within 500m	Ma, Duan et al. (2024)
Destination Accessibility	C_distance	Distance to City Center	Distance to the city center	Li, P., Chen et al. (2024)
	D_distance	Distance to living facilities	Sum of the distances to the nearest types of living facilities	Li, P., Chen et al. (2024)
Travel Attributes		Travel Distance	Mobile signaling data

Among them, the distance to living facilities includes three POI: living facilities, shopping facilities and catering facilities; the travel distance is the earth distance; Shannon entropy, the most commonly used diversity measurement in assessing mixed use. It can be used to evaluate the mixture degree of land use (Yue, Zhuang et al., 2017), the POI mixture index is calculated as follows:

Where S is the proportion of diversity pi type i in the entire POI class.

Interpretable machine learning

The adoption of machine learning regression methods in this study is motivated by their capacity to model complex, non-linear relationships between built environment factors (X) and travel distance (Y). Traditional linear regression approaches often impose the assumption of linearity, which may not accurately reflect the underlying dynamics in real-world scenarios. Machine learning techniques excel at capturing intricate patterns and interactions within high-dimensional data, allowing for a more nuanced understanding of how various built environment indicators influence travel behavior.

Category Boosting (CatBoost), an enhancement of the Gradient Boosting Decision Trees (GBDT) algorithm, was developed by the Yandex team to address certain limitations inherent in traditional gradient boosting tree models. CatBoost is engineered to deliver faster and more efficient machine learning models. It introduces several optimizations beyond the standard GBDT, particularly focusing on the enhanced management of categorical features and the treatment of missing values. The GBDT approach is an ensemble learning technique that amalgamates multiple decision trees with gradient boosting. Within CatBoost, each decision tree is trained to correct the errors made by its predecessors and is also informed by the subsequent trees. This iterative process aims to reduce the model's loss function, thereby refining the predictive capabilities of the model. CatBoost also extends GBDT by introducing optimizations that notably improve the handling of extreme values and outliers. Additionally, it addresses the issue of overfitting through the incorporation of a regularization term. The judicious selection of parameters can further enhance the model's generalization performance. In the context of this study, five parameters are identified as pivotal to the efficacy of the CatBoost model: Learning Rate: This parameter increases the model's robustness by adjusting the weights at each iteration. Maximum Depth: It specifies the maximum level of tree branching, with higher values potentially leading to overfitting. Sample Sampling Ratio: This involves a random subset of the training data, applied prior to tree growth. A lower ratio can mitigate overfitting but risks underfitting if set too low. Number of Trees: It represents the number of boosting iterations the model will undergo. Regularization Parameter: Higher values increase the model's conservatism, reducing the likelihood of overfitting. For the purposes of this study, 80% of the trip data records were randomly allocated to a training dataset, with the remainder reserved for testing. By fine-tuning these parameters, the model's complexity and generalization capacity can be optimally balanced, leading to improved predictive outcomes.

In this research, the CatBoost algorithm has been chosen for the construction of the predictive model, with a comparative analysis against LightGBM, XGBoost, and Support Vector Regression (SVR). The decision to include these algorithms stems from several considerations: firstly, their status as classical machine learning algorithms with broad applications and mature theoretical foundations in their respective domains. Secondly, their recognized excellence in performance on regression problems positions them as strong comparators to CatBoost. Moreover, their reliability and stability have been substantiated through extensive practical applications. When compared to LightGBM and XGBoost, CatBoost presents an advantage by providing automated and optimized methods for handling categorical features, which simplifies the model training process. Conversely, while SVR excels at managing nonlinear relationships, it often requires more complex tuning and may encounter computational and memory constraints with large-scale datasets. CatBoost's automated approach to categorical features and missing values eliminates the need for manual one-hot encoding or imputation, thereby streamlining feature engineering. This approach not only bolsters the model's robustness and stability but also ensures high predictive performance despite the presence of imperfect data. Furthermore, SHAP has been integrated to interpret the predictions of the machine learning models. SHAP is an innovative framework that assigns a fair contribution metric to each feature's value, grounded in game theory's Shapley values. It offers a quantifiable measure of each feature's impact on the prediction, thereby clarifying the model's decision-making process. SHAP acts as a conduit between the often-competing objectives of precision and interpretability in machine learning. Within this study, SHAP elucidates the extent to which each feature within the CatBoost model contributes to the prediction outcomes, enhancing the model's interpretability and overall clarity.

Analysis of Travel Behavior Patterns in Relation to Build Environment Features

Model comparison

Model comparisons were conducted to establish a benchmark for selecting the optimal model to address a specific problem. By evaluating the performance of various models on identical tasks, it becomes feasible to discern their respective strengths and limitations, thereby guiding the choice of the most suitable model. LightGBM, XGBoost, and SVR have been selected as comparative models to CatBoost for several compelling reasons. An exhaustive examination of the parameters within the gradient boosting tree algorithm is undertaken to ascertain the most effective hyperparameter configuration. As depicted in Table 2, any parameters not explicitly listed are assumed to be retained at their default settings.

Table 2. Machine learning hyperparameter Tuning

Hyperparameter Name	CatBoost	XGBoost	LightGBM
Learning Rate	0.68	0.04		0.274
Number of Trees	800	780		-
Max Depth	6	4		-
Regularization	6	0.02		-
Min Child Weight	-	1		-
Subsample	-	0.55		0.99
Gamma	-	0.03		-
Feature Fraction	-	0.65		0.446
Number of Leaves	-	-		186

The comparative analysis is presented in the table, which evaluates the performance of four prevalent machine learning models on a specific predictive task. The assessment is quantified using three principal metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²). The RMSE metric is particularly sensitive to large discrepancies between predicted and actual values, placing a higher penalty on such errors. In contrast, the MAE metric evenly weights all errors without differentiating based on magnitude. The R² metric, conversely, measures the proportion of variance in the dependent variable that is predictable from the independent variables, thus reflecting the model's goodness of fit. The empirical findings indicate that the CatBoost model outperforms its counterparts in this predictive endeavor. Specifically, the CatBoost model demonstrated an RMSE of 0.840, an MAE of 0.597, and an R² score of 0.736. These results are superior to those of the other models across all three evaluation metrics, signifying its robustness and efficacy in this particular application.

Table 3. Machine learning evaluation metrics

Evaluation Metrics	CatBoost	XGBoost	LightGBM		SVR
RMSE	0.840	0.874		0.884		0.999
MAE	0.597	0.603		0.595		0.646
R2	0.736	0.713		0.707		0.626

Characteristic importance

Within the scope of this study, the SHAPframework was employed to attain a more profound comprehension of the individual features' contributions to the prediction of daily life travel distances. Figure 3 delineates the marginal contributions of various features to the model's predictions, with the understanding that the collective contributions of all features integrate to a total of 100%. The findings from this analysis indicate that the POI mixture index exerts the most substantial influence on the prediction of daily life travel distances among the evaluated features. In comparison, the attributes of population density and building density have a relatively minor impact on the predictive model's outcomes.

Figure 3. Feature importance plot

The SHAP importance map offers a detailed depiction of the interplay between features and daily travel distances. The horizontal axis of the map represents the range of sample feature values, progressing from low to high values as one moves from left to right. The vertical axis corresponds to the Shapley values, which quantify the impact of each feature. A positive Shapley value indicates a beneficial effect on the prediction, while a negative value suggests a detrimental one. The graph effectively illustrates the extent and direction of the influence that varying features exert on the final predictive outcomes under different input conditions. It also provides a nuanced view of the impact of each feature on the variability of daily travel distances across different instances. As depicted in the figure, the POI mixture index (referred to as 'poi_index') exhibits a positive correlation with daily travel distance. Contrary to intuition, an increase in the POI mixture index is associated with a decrease in daily travel distance. This inverse relationship suggests that a higher POI mixture index, indicative of greater accessibility to services and facilities, may lead to reduced travel requirements for residents. The ratio of main to side roads (denoted as 'b_ratio') manifests an inverted U-shaped relationship with daily travel distance. This implies the existence of an optimal balance in the road network configuration that minimizes daily travel distances. Urban planning should aim to strike this balance, avoiding both excessive concentration and dispersion of resources to enhance travel efficiency. The distance to the city center and the distance to living facilities (referred to as 'c_distance' and 'd_distance', respectively) are negatively correlated with daily travel distance. This correlation indicates that greater distances from these central and residential amenities correspond to longer daily travel distances. Strategies to reduce the need for long-distance travel, such as the provision of nearby essential services and facilities, could be beneficial. The density of bus stops ('bus_stop') and bus stations demonstrate a positive correlation with daily travel distance. An increase in the density of these transportation nodes is linked to shorter daily travel distances, suggesting that a well-developed public transportation system can significantly curtail the travel distances of residents, particularly when bus stops are evenly and frequently distributed. Lastly, the density of crossroads ('crossroads') is inversely related to daily travel distance. A higher density of intersections correlates with shorter travel distances, indicating that a well-connected road network can facilitate more efficient travel patterns.

Figure 4. Individual feature importance plots

Feature interaction effects

Following an analysis of the primary effects that individual characteristics exert on daily travel distances, the study delved further to investigate potential interaction effects. This additional exploration aimed to elucidate the nonlinear dynamics at play. An interaction effect is said to occur when the influence of one variable is contingent upon the state of another. The study meticulously examined all possible pairs of features for evidence of interaction, quantifying the extent to which each pair interplayed in relation to daily travel distances. A null intersection value for a feature pair signifies the absence of any interactive effect between them. The graphical representation of these interaction effects is presented in the accompanying figure. Numerical values within the figure denote the degree of influence on the model's outcomes, with larger numbers signifying a more pronounced impact on the model's predictions. The intensity of these effects is further visualized through the heat map, where lighter-shaded squares correspond to features with more substantial effects. Notably, the interaction between the POI mixture index and the distance to the city center, as well as the main-to-side road ratio in conjunction with the stem-to-stern ratio, are highlighted by larger numerical values and lighter color blocks, indicating significant interaction effects. Conversely, other groups of features exhibit relatively minor effects on the model's predictive outcomes, as reflected by the darker squares allocated to these features in the heat map. This visual contrast aids in discerning the relative significance of various feature interactions in shaping the model's predictions.

Figure 5. Heat map of feature interaction effect values

Advancing Sustainable Urban Development: Insights from Built Environment Analysis

The research has essentially achieved its objectives, including understanding the motivations and considerations influencing residents' travel choices and exploring how various built environments attract residents and affect their travel distances. It also investigates the interactions between different built environments and their effects on travel distances, while analyzing the relationship between daily travel distances and the built environment to improve urban planning strategies. This comprehensive approach provides a robust framework for examining the complexities of urban travel behavior.

Employing the SHAP framework, the analysis reveals that the POI mixture index is the most significant predictor of daily travel distances. It shows that the POI mixture index has the greatest influence on predicting daily travel distances among all features, while the impacts of population density and building density are comparatively less significant. This aligns with the findings of Tamim Kashifi, Jamal et al. (2022), who emphasized the role of built environment features in travel mode choice analysis, a factor that extends to our examination of travel distances. Our research findings build upon the observation that the diversity and mixed use of urban features may significantly impact residents' daily travel distances.

The investigation of interaction effects uncovers important relationships between different built environments, particularly the interplay between the POI mixture index and the distance to the city center, as well as the ratio of main to side roads. These insights resonate with Schläpfer, Dong et al. (2021), who identified patterns in human mobility influenced by urban structure. Our results suggest that enhancing street network connectivity and increasing the POI mixture index can significantly reduce travel distances, providing actionable insights for urban planners.

Moreover, the study establishes that the CatBoost-based analytical framework outperforms LightGBM, XGBoost, and SVM models in predicting daily travel distances. This underscores the necessity of utilizing robust analytical tools to accurately assess the effectiveness of built environment characteristics. The notable contribution of the POI mixture index—accounting for over 20% of the prediction variance—further emphasizes the need to consider diverse urban factors when designing transportation policies.

These findings resonate with the work of Mo, Xu et al. (2023), which discusses advancements in travel behavior prediction using large language models, thereby validating the importance of interpretable frameworks in understanding travel dynamics. Furthermore, this research addresses a critical gap in existing literature, as noted by Chen, Zhao et al. (2022), who pointed out the lack of comprehensive studies linking urban design to travel behaviors in varying contexts.

Despite these contributions, this research has its limitations. The selection of variables was constrained, potentially overlooking other relevant factors that could influence travel behavior. Additionally, the reliance on observational data limits the ability to draw causal inferences, as the inherent complexities of travel behavior may not be fully captured. Future research could benefit from integrating big data sources with survey methodologies, allowing for a more comprehensive understanding of the motivations behind travel choices and enabling the exploration of a wider range of variables. This approach would enhance the validity of the findings and provide richer insights into the dynamics of urban travel behavior.

Conclusion

In conclusion, our research reinforces the significance of multidimensional urban characteristics in informing transportation policies and urban planning strategies. Future studies should explore these relationships across different urban settings and consider the implications of policy interventions in optimizing urban forms to foster sustainable travel behaviors. By building on the findings of this study, researchers and practitioners can contribute to more vibrant and accessible urban environments.

Author Contributions

Conceptualization, Boyuan Huang: Writing- original draft, Methodology, Formal analysis. Qiuyi Zhang: Writing -review & editing Methodology, Funding acquisition, Formal analysis, Conceptualization. Peifeng Yang: Writing -review & editing, Validation. All authors have read and agreed to the published version of the manuscript.

Ethics Declaration

The authors declare that they have no conflicts of interest regarding the publication of the paper.

Acknowledgments

This research was supported by the National Natural Science Foundation of China (No.42201225) and the Youth Fund of the Ministry of Education for Humanities and Social Sciences (No.22YJC840041). Publishing is also supported by the Open Fund of Key Laboratory of Ecology and Energy Saving Study of Dense Habitat, Ministry of Education (No.20230104) and Key Laboratory of Monitoring, Evaluation and Early Warning of Territorial Spatial Planning Implementation, Ministry of Natural Resources (LMEE-KF2023002).

Funding Statement

This research was supported by the National Natural Science Foundation of China (No.42201225) and the Youth Fund of the Ministry of Education for Humanities and Social Sciences (No.22YJC840041). Publishing is also supported by the Open Fund of Key Laboratory of Ecology and Energy Saving Study of Dense Habitat, Ministry of Education (No.20230104)and Key Laboratory of Monitoring, Evaluation and Early Warning of Territorial Spatial Planning Implementation, Ministry of Natural Resources (LMEE-KF2023002).

References

•  Cao,  F.,  Yang,  J., et al. (2024). "What Promotes Residents' Health? Identifying the Direct and Indirect Impacts of Community Environment on Residents' Health". International review for spatial planning and sustainable development, 12(1), 117-134. doi: http://doi.org/10.14246/irspsd.12.1_117.
•  Chen,  L.,  Zhao,  L., et al. (2022). "Investigating the Spatiotemporal Pattern between the Built Environment and Urban Vibrancy Using Big Data in Shenzhen, China". Computers, Environment and Urban Systems, 95, 101827. doi: https://doi.org/10.1016/j.compenvurbsys.2022.101827.
•  Fang,  C.,  Homma,  R., et al. (2024). "Impact of the Built Environment on Leisure Running: A Comparative Study of Weekdays and Weekends in Kumamoto City, Japan". International review for spatial planning and sustainable development, 12(4), 77-97. doi: http://doi.org/10.14246/irspsd.12.4_77.
•  Guo,  L.,  Yang,  S., et al. (2023). "Examining the Nonlinear Effects of Residential and Workplace-Built Environments on Active Travel in Short-Distance: A Random Forest Approach". International Journal of Environmental Research and Public Health, 20(3), 1969. doi: https://doi.org/10.3390/ijerph20031969.
•  Han,  B.,  Sun,  D., et al. (2020). "Classification of Urban Street Networks Based on Tree-Like Network Features". Sustainability, 12(2), 628. doi: https://doi.org/10.3390/su12020628.
•  Hipp , J. R.,  Lee,  S., et al. (2022). "Employment Deconcentration and Spatial Dispersion in Metropolitan Areas: Consequences for Commuting Patterns". Cities, 131, 103947. doi: https://doi.org/10.1016/j.cities.2022.103947.
•  Indradjati , P. N. (2024). "Walking Choice in Medium-Sized and Small Cities". International review for spatial planning and sustainable development, 12(3), 139-160. doi: http://doi.org/10.14246/irspsd.12.3_139.
•  Ji, S.,  Li,  J., et al. (2019). "Survey on Techniques, Applications and Security of Machine Learning Interpretability". Journal of Computer Research and Development (in Chinese), 56(10), 2071-2096. doi: http://doi.org/10.7544/issn1000-1239.2019.20190540.
•  Li, P.,  Chen,  X., et al. (2024). "Study on the Nonlinear Relationship between Subway Construction Environment and Station Travel Distance Based on Xgboost-Shap". Journal of Railway Science and Engineering, 21(04), 1624-1633. doi: http://doi.org/10.19713/j.cnki.43-1423/u.T20231070.
•  Li, Z. (2022). "Extracting Spatial Effects from Machine Learning Model Using Local Interpretation Method: An Example of Shap and Xgboost". Computers, Environment and Urban Systems, 96, 101845. doi: https://doi.org/10.1016/j.compenvurbsys.2022.101845.
•  Liu,  J.,  Wang,  M., et al. (2024). "Riding Towards a Sustainable Future: An Evaluation of Bike Sharing’s Environmental Benefits in Xiamen Island, China". Geography and Sustainability, 5(2), 276-288. doi: https://doi.org/10.1016/j.geosus.2024.01.002.
•  Liu,  T. and  Ding,  C. (2024). "Revisiting Built Environment and Travel Behavior: A Natural Experiment Accounting for Residential Self-Selection". Journal of Transport Geography, 115, 103794. doi: https://doi.org/10.1016/j.jtrangeo.2024.103794.
•  Lv, X. and  Pan , t. (2022). "Take Shanghai Songjiang New City as an Example". Urban Traffic (in Chinese), 20(04), 99-110. doi: https://doi.org/10.13813/j.cn11-5141/u.2022.0024.
•  Ma, S.,  Duan,  C., et al. (2024). "Measurement and Evolution Law of Traffic Carbon Emission under Mixed-Line Environment". Environmental Sciences, 45(11), 6403-6411. doi: https://doi.org/10.13227/j.hjkx.202312108.
•  Mo, B.,  Xu,  H., et al. (2023). "Large Language Models for Travel Behavior Prediction". arXiv preprint arXiv:2312.00819. doi: https://doi.org/10.48550/arXiv.2312.00819.
•  Mondal,  B. (2023). "Analysing the Consequences of Progressive Urban Spatial Cycles to Evaluate Urban Land Use Policy". Environment and Planning B: Urban Analytics and City Science, 50(7), 1722-1738. doi: https://doi.org/10.1177/23998083221140884.
•  Næss,  P.,  Stefansdottir,  H., et al. (2021). "Residential Location and Travel in the Reykjavik Capital Region". Sustainability, 13(12), 6714. doi: https://doi.org/10.3390/su13126714.
•  Nikitas,  A.,  Tsigdinos,  S., et al. (2021). "Cycling in the Era of Covid-19: Lessons Learnt and Best Practice Policy Recommendations for a More Bike-Centric Future". Sustainability, 13(9), 4620. doi: https://doi.org/10.3390/su13094620.
•  Scheiner,  J.,  Frank,  S., et al. (2024). "In Search of Causality in the Relationship between the Built Environment and Travel Behaviour. On the Challenges of Planning and Realising an Ambitious Mixed-Methods Panel Travel Survey among Relocating Households in Germany". Progress in Planning, 182, 100820. doi: https://doi.org/10.1016/j.progress.2023.100820.
•  Schläpfer,  M.,  Dong,  L., et al. (2021). "The Universal Visitation Law of Human Mobility". Nature, 593(7860), 522-527. doi: https://doi.org/10.1038/s41586-021-03480-9.
•  Tamim Kashifi,  M.,  Jamal,  A., et al. (2022). "Predicting the Travel Mode Choice with Interpretable Machine Learning Techniques: A Comparative Study". Travel Behaviour and Society, 29, 279-296. doi: https://doi.org/10.1016/j.tbs.2022.07.003.
•  Teixeira , I. P.,  Barbosa,  J.  Paulo  d. A. S., et al. (2023). "Built Environments for Physical Activity: A Longitudinal Descriptive Analysis of Sao Paulo City, Brazil". Cities & Health, 7(1), 137-147. doi: https://doi.org/10.1080/23748834.2022.2127173.
•  Ton,  D.,  Bekhor,  S., et al. (2020). "The Experienced Mode Choice Set and Its Determinants: Commuting Trips in the Netherlands". Transportation Research Part A: Policy and Practice, 132, 744-758. doi: https://doi.org/10.1016/j.tra.2019.12.027.
•  Wang, C.,  Wang, X., et al. (2022). "Influence of Built Environment on Subway Trip Origin and Destination: Insights Based on Mobile Positioning Data". Transportation Research Record, 2676(9), 693-710. doi: https://doi.org/10.1177/03611981221088223.
•  Wang,  Q.,  Liu,  X., et al. (2023). "Does Compact City Shape Matter to Residents' Happiness?". Cities, 141, 104524. doi: https://doi.org/10.1016/j.cities.2023.104524.
•  Wang,  Y.,  Yao,  X., et al. (2024). "Investigating the Effect of Industry-Specific Economic Distance on the Prediction of Intercity Population Movement". Cities, 150, 105047. doi: https://doi.org/10.1016/j.cities.2024.105047.
•  Wu, J. and  Liu,  X. (2023). "Analysis of the Influence of Different Housing Types on the Choice of Commuting Mode". Journal of Jilin University(Engineering and Technology Edition) (in Chinese). doi: https://doi.org/10.13229/j.cnki.jdxbgxb.20230387.
•  Yang,  S.,  Zhou,  L., et al. (2024). "Examining Multiscale Built Environment Interventions to Mitigate Travel-Related Carbon Emissions". Journal of Transport Geography, 119, 103942. doi: https://doi.org/10.1016/j.jtrangeo.2024.103942.
•  Yue,  Y.,  Zhuang,  Y., et al. (2017). "Measurements of Poi-Based Mixed Use and Their Relationships with Neighbourhood Vibrancy". International Journal of Geographical Information Science, 31(4), 658-675. doi: https://doi.org/10.1080/13658816.2016.1220561.
•  Zhang,  Y.,  Wang,  R., et al. (2023). "The Influence of Built-Environment Factors on Connectivity of Road Networks in Residential Areas: A Study Based on 204 Samples in Nanjing, China". Buildings, 13(2), 301. doi: https://doi.org/10.3390/buildings13020301.
•  Zhang,  Y.,  Zhou,  W., et al. (2022). "Effects of the Built Environment on Travel-Related Co2 Emissions Considering Travel Purpose: A Case Study of Resettlement Neighborhoods in Nanjing". Buildings, 12(10), 1718. doi: https://doi.org/10.3390/buildings12101718.