Subway peripheral non-commuter destination selection prediction method based on built environment

Abstract

The invention discloses a subway peripheral non-commuter destination selection prediction method based on a built environment, and the method specifically comprises the steps: data collection: selecting a research region in a target city, and collecting the related built environment data in a subway station water collection region; feature selection: adding variables influencing non-commuter according to the built environment variables in combination with local special conditions; model construction: adopting a random forest algorithm to establish a prediction model; sHAP analysis: analyzing the output of the random forest model by applying an SHAP technology, and determining the importance sequence and the action mechanism of each built-up environment factor for the destination selection of the non-commuter; and result interpretation and prediction: determining each built environment factor numerical matrix of the new scene, and calculating to obtain a final destination selection result of each non-commuting trip individual in the new scene. The method can accurately understand and predict the destination selection behavior of the non-commuter around the subway station, thereby providing powerful support for urban planning and traffic management.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban planning and traffic engineering, and in particular relates to a method for predicting destination selection of non-commuters around a subway based on a built environment. Background Art

[0002] The subway system is not only a common means of transportation for daily commuters in the city, but also plays an important role in attracting a large number of non-commuters. In the areas around subway stations, the construction of high-density unique streetscapes, entertainment venues, and daily life service facilities can attract a large number of non-commuters. These non-commuting activities include shopping, tourism, leisure and entertainment, and medical services, which transform the areas around subway stations into vibrant urban centers. Although some studies have shown that there is a nonlinear relationship between walking preference and the built environment (Tu et al., 2021; Tao et al., 2023), these studies have not specifically explored the travel behavior of non-commuters. Current research often focuses on the travel behavior of commuters, while the behavioral characteristics of non-commuters and their impact on destination choice have not been fully understood (Zhang et al., 2014; Tilahun et al., 2016; Dill and McNeil, 2023).

[0003] In the process of rapid urban development, how to effectively integrate subway stations with the surrounding built environment has become a challenge. Due to the complex development of the built environment and the diversification of user needs, unsuccessful integration will lead to functional mismatch, reduce the accessibility of travel destinations, and affect the travel experience (Yang et al., 2016; Wang et al., 2022). These problems hinder the full realization of urban vitality.

[0004] Traditional statistical models such as linear regression and multi-class logistic regression have limitations in dealing with complex nonlinear relationships and cannot well capture the complex interactions between built environment characteristics and non-commuters' destination choices (Prasad et al., 2006; Redo et al., 2012). Therefore, a more flexible and powerful method is needed to reveal the interactions between these variables. Summary of the invention

[0005] Based on the problems and shortcomings in the above background technology, the present invention provides a method for predicting destination selection of non-commuters in the subway periphery based on the built environment.

[0006] The invention provides a method for predicting destination selection of non-commuters in the subway periphery based on built environment, comprising the following steps:

[0007] Step 1: Data collection: Select a study area within the target city and collect relevant built environment data within the subway station catchment area, including but not limited to environmental data, point of interest (POI) data, and traffic data; in addition, it is necessary to obtain travel records and destination selection information of respondents who take the subway for non-commuting trips.

[0008] Step 2: Feature selection: Identify important built environment variables based on the literature review and preliminary analysis, and add variables that may have an impact on non-commuters based on local special circumstances.

[0009] Step 3: Model construction: Use the random forest algorithm to build a prediction model; use the Scikit-learn module in the Python programming language to implement the random forest technology, and divide the data set selected by the travelers into a training set and a test set; set the model parameters as the number of decision trees and the number of variables considered for each split, and do not limit the maximum tree depth; then, perform model training and verification, use the training set data to train the random forest model, and evaluate the performance of different model variants through the out-of-bag error rate; after determining the optimal model, use the test set for verification to ensure that the model has good generalization ability.

[0010] Step 4: SHAP analysis: Apply SHAP technology to analyze the output of the random forest model and clarify the importance ranking and mechanism of each built environment factor on the destination choice of non-commuters.

[0011] Step 5: Interpretation and prediction of results: Determine which built environment elements are most or least attractive to non-commuters based on the SHAP value; adjust the corresponding facility layout or service supply based on the actual scenario to determine the numerical matrix of each built environment factor in the new scenario; substitute the above model coefficient results into the new built environment factor matrix to calculate the final destination selection results of each non-commuting travel individual under this new scenario.

[0012] Furthermore, step 1 is specifically as follows:

[0013] S1.1: Identify the study area:

[0014] Specific subway stations in the target cities were selected as research objects, and a research area with a radius of 1000 meters was delineated with each station as the center.

[0015] S1.2: Field investigation and data acquisition:

[0016] Environmental data: Information about the built environment collected through field measurements, street view maps, and high-resolution satellite imagery; including but not limited to sidewalk width, green area, average number of floors in a building, and building density.

[0017] POI data: records the type, location and number of various points of interest in the study area.

[0018] Traffic data: Statistics on the number of bus stops and their distribution, as well as road grade information.

[0019] S1.3: Questionnaire design and distribution:

[0020] Questionnaire content: A structured paper questionnaire was designed, divided into two parts; the first part collected the demographic information of the respondents, including but not limited to age, gender, and family composition; the second part recorded the travel characteristics of the respondents, including but not limited to the departure and destination, road length, travel purpose, and travel duration.

[0021] Questionnaire distribution: Face-to-face interviews were conducted by randomly selecting participants near the selected metro stations; ensuring that the questionnaires were evenly distributed on weekdays and weekends to obtain comprehensive travel pattern data; the selected dates should exclude extreme weather conditions.

[0022] S1.4: Data collation and screening:

[0023] First, data entry is performed, and the collected questionnaire data is entered into a spreadsheet or database; second, the entered data is carefully screened and verified, and invalid or incomplete answers are eliminated to ensure the validity and accuracy of the data; finally, sample characteristic analysis is performed, and descriptive statistical analysis is performed on the demographic characteristics of the valid sample to ensure that the sample is representative.

[0024] S1.5: Geographic Information System Applications:

[0025] Use GIS software to process the collected spatial data, including POI locations, road networks, and bus stop distribution; create a map representation of the study area to intuitively display the distribution of different types of built environment elements and points of interest.

[0026] S1.6: Comprehensive dataset construction:

[0027] All collected data are integrated into a unified dataset, including built environment variables, POI data, traffic data, and travel records of non-commuters; data are preprocessed, including missing value filling, outlier detection and processing, and data standardization to ensure that the data quality meets the requirements of subsequent modeling.

[0028] Furthermore, step 2 is specifically as follows:

[0029] S2.1: Literature review and preliminary screening:

[0030] We reviewed relevant research literature to understand which built environment factors have been shown to have a significant impact on non-commuters’ destination choices, including but not limited to road length, building coverage, green area, sidewalk width, average number of floors in buildings, building plot coverage, road length, land use mix, number of bus stops, and density of various POIs.

[0031] S2.2: Data Exploratory Analysis:

[0032] Descriptive statistical analysis was performed on the collected data, including calculating the maximum, minimum, standard deviation and mean of each variable to understand the basic distribution of the data; the correlation between the built environment variables was evaluated by calculating the Pearson or Spearman correlation coefficient, and highly correlated variables were identified to avoid multicollinearity problems; box plots or scatter plots were used to visually display the distribution of each variable and its relationship with other variables to help identify outliers and potential important features.

[0033] S2.3: Final feature list:

[0034] Based on the above results, determine the final list of features for modeling; these features should have good predictive ability and no serious collinearity between each other; record the definition, source and role of each selected feature in the model in detail to facilitate subsequent interpretation and application.

[0035] Furthermore, step 3 is specifically as follows:

[0036] S3.1: Data preparation:

[0037] Data splitting, divide the collected data set into training set and test set, using 75% of the data as the training set and 25% of the data as the test set; this can be achieved through the `train_test_split` function in Python's Scikit-learn library.

[0038] Features are separated from labels. Features X and target variables y are separated from the dataset. The features include built environment factors and the target variable is the destination choice of non-commuters.

[0039] S3.2: Model initialization:

[0040] Select the random forest algorithm and use the `RandomForestClassifier` or `RandomForestRegressor` class in the Scikit-learn library to create a random forest model instance, depending on whether the target variable is a classification or regression problem; set the initial parameters and set the initial parameters of the random forest model according to the literature and preliminary experiments, including the number of decision trees n_estimators, the number of variables considered for each split max_features, and the maximum tree depth max_depth.

[0041] S3.3: Model training:

[0042] Fitting model: Use the training set data to train the random forest model, that is, call the `fit` method; this process will generate multiple decision trees based on the training data and make predictions through the integration of these trees.

[0043] Hyperparameter tuning: Use cross-validation techniques to search for the best hyperparameter combination to further optimize model performance.

[0044] S3.4: Model Evaluation:

[0045] Out-of-bag error rate evaluation: During the training process, the error rate of out-of-bag (OOB) samples is calculated as an indicator of the generalization ability of the model; if the OOB error rate is low, it indicates that the model has good generalization ability.

[0046] Test set verification: Use the test set data to verify the model and calculate evaluation indicators such as accuracy, recall, and F1 score to ensure that the model not only performs well on the training set, but can also make accurate predictions on unseen data.

[0047] Furthermore, step 4 is specifically as follows:

[0048] S4.1: Model calculation:

[0049] Apply SHAP analysis: Use SHAP values ​​to interpret the output of the random forest model and understand how much each feature affects the model’s predictions.

[0050] A significant enhancement introduced by SHAP lies in its ability to depict explanations through a linear model framework that can be expressed mathematically in equations (1) and (2); this is known as the additive variable attribution approach, which aggregates individual effects in addition to the intercept to illuminate the explanatory dynamics of the model.

[0051]

[0052] In the formula, g is the framework of the explanatory model, the number of input variables is represented by P, the variable z represents the presence of the corresponding variable 1 or the absence of 0, the symbol Φ0 is the constant term, Φ j Represents the attribute value assigned to variable j.

[0053]

[0054] where S represents a subset of the variables used, the variable p represents the total number of variables, and the function g(z')(S∪{j}) represents the output value of the model given the combination of variables in the subset S; regarding the concept of weight, for p individual variables, given their sequence, the permutation of these variables is equal to p! combinations; when the position of a particular variable j is fixed, the remaining permutations involving other variables are given by |S|! (p-|S|-1)!, where |S| is the cardinality of the subset S, indicating the number of variables it contains.

[0055] Draw SHAP diagrams: Create SHAP summary diagrams and dependency diagrams to visually display the importance of each feature and its relationship pattern with the target variable.

[0056] S4.2: Global explanation:

[0057] SHAP summary plot: Draw a SHAP summary plot to show the importance ranking of all features and their overall impact on the model output; this can be achieved through shap.summary_plot(shap_values,X_test); in the summary plot, each point represents a sample, the color represents the actual value of the feature, and the horizontal axis represents the SHAP value, that is, the impact of the feature on the prediction result of the sample.

[0058] Mean Absolute SHAP Value: The mean absolute SHAP value is calculated for each feature to quantify its average effect size on the model output.

[0059] S4.3: Local interpretation:

[0060] SHAP dependency graph: For the selected key features, a SHAP dependency graph is drawn to show the relationship between the feature value and its corresponding SHAP value.

[0061] Individual sample explanation: Select a specific sample and generate a visualization of its SHAP value to understand the specific contribution of each feature in the sample to its prediction result; this can be achieved by shap.force_plot(explainer.expected_value,shap_values[0,:],X_test.iloc[0,:]), where shap_values[0,:] and X_test.iloc[0,:] correspond to the SHAP value and feature value of the first sample, respectively.

[0062] S4.4: Interpretation of results:

[0063] Based on the SHAP summary plot and the average absolute SHAP value, we determined which built environment factors had the greatest impact on non-commuters' destination choice; analyzed the nonlinear relationship between characteristics and destination choice through SHAP dependency plots; and examined whether SHAP values ​​showed interactions between certain characteristics, especially the change patterns observed under different combinations of characteristics.

[0064] Furthermore, step 5 is specifically as follows:

[0065] S5.1: Summarize key findings:

[0066] Main influencing factors: Based on the analysis of SHAP values, the most important built environment factors are identified.

[0067] Non-linear relationship revelation: For variables that show significant non-linear relationships, describe the changing trend of their impact on destination selection as the variable value changes.

[0068] S5.2: Interpretation of results:

[0069] Explain why certain factors have a greater impact on destination choice, emphasizing the importance of accessibility and aesthetic factors; discuss those factors that show context-dependent effects and show that they may have different effects in different contexts.

[0070] S5.3: Case studies:

[0071] Specific case analysis: Select representative cases to analyze in detail how built environment factors work together to influence non-commuters’ destination choices; use maps and charts to intuitively display the key variables and their impacts in these cases; and extract quantitative values ​​of each built environment element based on the calculated interactive influence relationship mechanism and the characteristics of the built environment elements in the cases.

[0072] S5.4: Suggestions and Feedback:

[0073] Urban planning recommendations: Based on the research results, specific recommendations are provided to urban planners, including but not limited to increasing green space, optimizing bus station layout, and adjusting building density to improve the attractiveness of areas around subway stations.

[0074] Traffic Management Strategy: Propose improvement measures for traffic management departments, including improving road design and enhancing the connectivity of public transportation services to enhance the travel experience of non-commuters.

[0075] Commercial layout optimization: Provide location suggestions to merchants to help them find the best location around subway stations to attract more non-commuters.

[0076] The beneficial technical effects of the present invention are:

[0077] (1) Improved prediction accuracy: By combining the random forest model with SHAP analysis, we can more accurately identify the key built environment factors that affect non-commuters’ destination choices. This method goes beyond traditional linear statistical models and reveals the complex nonlinear relationships between variables, thereby improving prediction accuracy.

[0078] (2) Optimizing urban planning and design: After identifying which built environment elements are most or least attractive to non-commuters, urban planners can adjust the development strategy of areas around subway stations accordingly, such as increasing green areas and properly setting up bus stops, so as to attract more non-commuters and enhance regional vitality.

[0079] (3) Enhanced accessibility and user experience: Based on the research results of this invention, transportation management departments and urban developers can better understand how to enhance the accessibility of subway stations and their surrounding areas by improving physical layout and service supply, and provide non-commuters with a more convenient and comfortable travel experience.

[0080] (4) Supporting sustainable urban development: A deeper understanding of the behavior patterns of non-commuters can help promote the implementation of a more efficient and attractive public transportation system, thereby reducing the use of private cars and promoting a green and low-carbon urban development model.

[0081] (5) Guiding business decisions: For merchants, understanding what type of built environment is more popular with non-commuters can help them make more informed choices when selecting locations, maximize store exposure and customer flow, and achieve better economic benefits.

[0082] In summary, the present invention not only provides a new perspective and technical means for studying the behavior of non-commuters using the subway, but also brings practical application value to multiple fields such as urban planning, traffic management, and commercial operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 The theoretical framework diagram of the method for predicting destination selection of non-commuters in the periphery of a subway station according to the present invention.

[0084] Figure 2 A flow chart is constructed for the random forest model of the present invention.

[0085] Figure 3 This is a diagram of the SHAP value analysis results of the present invention.

[0086] Figure 4 Schematic diagram showing the influence of the key variable (road length) of the present invention.

[0087] Figure 5It is a schematic diagram showing the influence of the key variable (average number of floors in a building) of the present invention.

[0088] Figure 6 It is a schematic diagram of the impact of the key variable (greening area) of the present invention.

[0089] Figure 7 It is a schematic diagram of the impact of the key variables (tourist attractions POI) of the present invention.

[0090] Figure 8 This is the predicted distribution map of non-commuting destinations of the present invention. DETAILED DESCRIPTION

[0091] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0092] The principle of the method for predicting the destination selection of non-commuters around subway based on built environment is as follows: Figure 1 As shown, the following steps are included:

[0093] Step 1: Data collection: Select a study area within the target city and collect relevant built environment data within the subway station catchment area, including but not limited to environmental data, point of interest (POI) data, and traffic data; in addition, it is necessary to obtain the travel records and destination selection information of respondents who take the subway for non-commuting trips.

[0094] Step 2: Feature selection: Identify important built environment variables based on the literature review and preliminary analysis, and add variables that may have an impact on non-commuters based on local special circumstances.

[0095] Step 3: Model construction: Use the random forest algorithm to build a prediction model; use the Scikit-learn module in the Python programming language to implement the random forest technology, and divide the data set selected by the travelers into a training set and a test set; set the model parameters as the number of decision trees and the number of variables considered for each split, and do not limit the maximum tree depth; then, perform model training and verification, use the training set data to train the random forest model, and evaluate the performance of different model variants through the out-of-bag error rate; after determining the optimal model, use the test set for verification to ensure that the model has good generalization ability.

[0096] Step 4: SHAP analysis: SHAP (SHapley Additive exPlanations) technology is used to analyze the output of the random forest model to clarify the importance ranking and mechanism of each built environment factor on the destination choice of non-commuters. This step can reveal the complex nonlinear relationship between variables and provide global and local explanations.

[0097] Step 5: Interpretation and prediction of results: Determine which built environment elements are most or least attractive to non-commuters based on the SHAP value; adjust the corresponding facility layout or service supply in combination with the actual scenario to determine the numerical matrix of each built environment factor in the new scenario; substitute the above model coefficient results into the new built environment factor matrix to calculate the final destination selection results of each non-commuter travel individual in this new scenario. This prediction result can help improve the built environment around the subway, enhance the non-commuter experience, and promote the improvement of regional vitality.

[0098] Furthermore, step 1 is specifically as follows:

[0099] S1.1: Identify the study area:

[0100] Specific subway stations in the target cities were selected as research objects, and a research area with a radius of 1,000 meters was defined with each station as the center. The selected subway stations should have a large number of non-commuting passers-by and have diverse built environment characteristics.

[0101] S1.2: Field investigation and data acquisition:

[0102] Environmental data: collect information about the built environment through field measurements, street view maps, and high-resolution satellite imagery (such as Baidu Satellite Network Map); including but not limited to sidewalk width, green area, average number of floors in a building, and building density.

[0103] POI data: records the type, location and number of various points of interest in the study area.

[0104] Traffic data: Statistics on the number and distribution of bus stops, as well as road grades (main roads and branch roads), etc., in order to understand how public transportation services affect the travel decisions of non-commuters.

[0105] S1.3: Questionnaire design and distribution:

[0106] Questionnaire content: A structured paper questionnaire was designed, divided into two parts; the first part collected the demographic information of the respondents, including but not limited to age, gender, and family composition; the second part recorded the travel characteristics of the respondents, including but not limited to the departure and destination, road length, travel purpose, and travel duration.

[0107] Questionnaire distribution: Face-to-face interviews were conducted by randomly selecting participants near the selected metro stations; ensuring that the questionnaires were evenly distributed on weekdays and weekends to obtain comprehensive travel pattern data; the selected dates should exclude extreme weather conditions.

[0108] S1.4: Data collation and screening:

[0109] First, data entry is performed, and the collected questionnaire data is entered into a spreadsheet or database; second, the entered data is carefully screened and verified, and invalid or incomplete answers are eliminated to ensure the validity and accuracy of the data; finally, sample characteristic analysis is performed, and descriptive statistical analysis is performed on the demographic characteristics of the valid sample to ensure that the sample is representative.

[0110] S1.5: Geographic Information System Applications:

[0111] Use GIS software to process the collected spatial data, including POI locations, road networks, and bus stop distribution; create a map representation of the study area to intuitively display the distribution of different types of built environment elements and points of interest.

[0112] S1.6: Comprehensive dataset construction:

[0113] All collected data are integrated into a unified dataset, including built environment variables, POI data, traffic data, and travel records of non-commuters; data are preprocessed, including missing value filling, outlier detection and processing, and data standardization to ensure that the data quality meets the requirements of subsequent modeling.

[0114] Furthermore, step 2 is specifically as follows:

[0115] S2.1: Literature review and preliminary screening:

[0116] We reviewed relevant research literature to understand which built environment factors have been shown to have a significant impact on non-commuters’ destination choices, including but not limited to road length, building coverage, green area, sidewalk width, green area, average number of building floors, building plot coverage, road length, land use mix, number of bus stops, and density of various POIs.

[0117] S2.2: Data Exploratory Analysis:

[0118] Descriptive statistical analysis was performed on the collected data, including calculating the maximum, minimum, standard deviation and mean of each variable to understand the basic distribution of the data; the correlation between the built environment variables was evaluated by calculating the Pearson or Spearman correlation coefficient, and highly correlated variables were identified to avoid multicollinearity problems; box plots or scatter plots were used to visually display the distribution of each variable and its relationship with other variables to help identify outliers and potential important features.

[0119] S2.3: Final feature list:

[0120] Based on the above results, determine the final list of features for modeling; these features should have good predictive ability and no serious collinearity between each other; record the definition, source and role of each selected feature in the model in detail to facilitate subsequent interpretation and application.

[0121] Furthermore, step 3 is specifically as follows:

[0122] S3.1: Data preparation:

[0123] Data splitting, divide the collected data set into training set and test set, using 75% of the data as the training set and 25% of the data as the test set; this can be achieved through the `train_test_split` function in Python's Scikit-learn library.

[0124] Features are separated from labels. Features X and target variables y are separated from the dataset. The features include built environment factors and the target variable is the destination choice of non-commuters.

[0125] S3.2: Model initialization:

[0126] Select the random forest algorithm and use the `RandomForestClassifier` or `RandomForestRegressor` class in the Scikit-learn library to create a random forest model instance, depending on whether the target variable is a classification or regression problem; set the initial parameters and set the initial parameters of the random forest model according to the literature and preliminary experiments, including the number of decision trees (n_estimators), the number of variables considered for each split (max_features), and the maximum tree depth (max_depth).

[0127] S3.3: Model training:

[0128] Fitting model: Use the training set data to train the random forest model, that is, call the `fit` method; this process will generate multiple decision trees based on the training data and make predictions through the integration of these trees.

[0129] Hyperparameter tuning: Use cross-validation techniques (such as GridSearchCV or RandomizedSearchCV) to search for the best hyperparameter combination to further optimize model performance. You can try to adjust key parameters such as `n_estimators`, `max_features`, `max_depth`, etc.

[0130] S3.4: Model Evaluation:

[0131] Out-of-Bag error rate evaluation: During the training process, the error rate of out-of-bag (OOB) samples is calculated as an indicator of the generalization ability of the model; if the OOB error rate is low, it indicates that the model has good generalization ability.

[0132] Test set verification: Use the test set data to verify the model and calculate evaluation indicators such as accuracy, recall, and F1 score to ensure that the model not only performs well on the training set, but can also make accurate predictions on unseen data.

[0133] Furthermore, step 4 is specifically as follows:

[0134] S4.1: Model calculation:

[0135] Apply SHAP analysis: Use SHAP values ​​to interpret the output of the random forest model and understand the degree to which each feature affects the model’s predictions. This step can help identify which built environment factors are most critical to non-commuters’ destination choices.

[0136] A significant enhancement introduced by SHAP lies in its ability to depict explanations through a linear model framework that can be expressed mathematically in equations (1) and (2); this is known as the additive variable attribution approach, which aggregates individual effects in addition to the intercept to illuminate the explanatory dynamics of the model.

[0137]

[0138] In the formula, g is the framework of the explanatory model, the number of input variables is represented by P, the variable z represents the presence of the corresponding variable 1 or the absence of 0, the symbol Φ0 is the constant term, Φ j Represents the attribute value (Shapley value) assigned to variable j.

[0139]

[0140] where S represents a subset of the variables used, the variable p represents the total number of variables, and the function g(z')(S∪{j}) represents the output value of the model given the combination of variables in the subset S; regarding the concept of weight, for p individual variables, given their sequence, the permutation of these variables is equal to p! combinations; when the position of a particular variable j is fixed, the remaining permutations involving other variables are given by |S|! (p-|S|-1)!, where |S| is the cardinality of the subset S, indicating the number of variables it contains.

[0141] This paper utilizes the SHAP package in the Python programming language to elucidate the foundations of the random forest model. The SHAP package includes an improved version of the TreeExplainer algorithm, which is designed to calculate Shapley values ​​faster. By applying this approach, global and local output patterns of the random forest model are explained. Globally, the scheme provides a comprehensive overview, highlighting the relative importance of elements related to the built environment. By classifying population groups, the tool provides local insights into nonlinear correlations and illustrates the complex meanings and interactions within the model.

[0142] Draw SHAP graphs: Create SHAP summary graphs and dependency graphs to visually display the importance of each feature and its relationship pattern with the target variable. This helps to gain a deeper understanding of nonlinear effects and interactions between features.

[0143] S4.2: Global explanation:

[0144] SHAP summary plot: Draw a SHAP summary plot to show the importance ranking of all features and their overall impact on the model output; this can be achieved through shap.summary_plot(shap_values,X_test); in the summary plot, each point represents a sample, the color represents the actual value of the feature, and the horizontal axis represents the SHAP value, that is, the impact of the feature on the prediction result of the sample.

[0145] Mean Absolute SHAP Value: The mean absolute SHAP value is calculated for each feature to quantify its average effect size on the model output. This helps identify the most important built environment factors.

[0146] S4.3: Local interpretation:

[0147] SHAP dependency graph: For selected key features, a SHAP dependency graph is plotted to show the relationship between the feature value and its corresponding SHAP value. This can help understand the nonlinear relationship between the feature and the target variable.

[0148] Individual sample explanation: Select a specific sample and generate a visualization of its SHAP value to understand the specific contribution of each feature in the sample to its prediction result; this can be achieved by shap.force_plot(explainer.expected_value,shap_values[0,:],X_test.iloc[0,:]), where shap_values[0,:] and X_test.iloc[0,:] correspond to the SHAP value and feature value of the first sample, respectively.

[0149] S4.4: Interpretation of results:

[0150] Based on the SHAP summary plot and the average absolute SHAP value, we determined which built environment factors had the greatest impact on non-commuters' destination choice; analyzed the nonlinear relationship between characteristics and destination choice through SHAP dependency plots; and examined whether SHAP values ​​showed interactions between certain characteristics, especially the change patterns observed under different combinations of characteristics.

[0151] Furthermore, step 5 is specifically as follows:

[0152] S5.1: Summarize key findings:

[0153] Main influencing factors: Based on the analysis of SHAP values, the most important built environment factors are identified.

[0154] Non-linear relationship revelation: For variables that show significant non-linear relationships, describe the changing trend of their impact on destination selection as the variable value changes.

[0155] S5.2: Interpretation of results:

[0156] Explain why certain factors have a greater impact on destination choice, emphasizing the importance of accessibility and aesthetic factors; discuss those factors that show context-dependent effects and show that they may have different effects in different contexts.

[0157] S5.3: Case studies:

[0158] Specific case analysis: Select representative cases to analyze in detail how built environment factors work together to influence non-commuters’ destination choices; use maps and charts to intuitively display the key variables and their impacts in these cases; and extract quantitative values ​​of each built environment element based on the calculated interactive influence relationship mechanism and the characteristics of the built environment elements in the cases.

[0159] S5.4: Suggestions and Feedback:

[0160] Urban planning recommendations: Based on the research results, specific recommendations are provided to urban planners, including but not limited to increasing green space, optimizing bus station layout, and adjusting building density to improve the attractiveness of areas around subway stations.

[0161] Traffic Management Strategy: Propose improvement measures for traffic management departments, including improving road design and enhancing the connectivity of public transportation services to enhance the travel experience of non-commuters.

[0162] Commercial layout optimization: Provide location suggestions to merchants to help them find the best location around subway stations to attract more non-commuters.

[0163] Example:

[0164] Study subjects:

[0165] Two adjacent subway stations in the central area of ​​Chengdu, People's Park Station and Kuanzhai Alley Station, were selected as the research objects. These stations are not only located in the core area of ​​the city, but also have rich and diverse built environment characteristics, tourist attractions and other facilities and environments that attract non-commuters. They are eligible survey areas and can provide more comprehensive variables for the model.

[0166] Data Collection:

[0167] Environmental data: Information such as sidewalk width, green area, average number of floors in buildings, and building plot coverage was collected through field measurements and high-resolution satellite images (e.g., Baidu Satellite Network Map at a scale of 1:1500 in the summer of 2023).

[0168] As shown in Table 1. The calculation of land use mixture degree L is shown in formula (3).

[0169]

[0170] In formula (1), n ​​is the land use type, n is the total number of different land use types, S n is the building area of ​​land use type n, in km 2 , S is the total building area of ​​all land use types in the study area, unit is km 2 Land use is divided into ten categories: commercial, daily services (government / postal / other), business, residential, catering, entertainment and culture, retail, school, open space (park / square) and others.

[0171] Table 1 Built environment elements

[0172]

[0173] POI data: records the location and number of various points of interest in the study area, including catering facilities, tourist attractions, hotels, bank ATMs, museums, shops, parking lots, daily life facilities, entertainment venues, etc.

[0174] Traffic data: Statistics on the number and distribution of bus stops, as well as road grades.

[0175] The collected spatial data were processed using GIS software to create a map representation of the study area. All data were integrated into a unified data set and data standardization was performed, as shown in Table 2.

[0176] Table 2 Statistics of built environment elements

[0177]

[0178]

[0179] Questionnaire design: A structured paper questionnaire was designed. The questionnaire was divided into two different parts: the first part aimed to collect the socio-demographic information of the respondents, including age, gender, family composition, frequency of visits to the neighborhood, and concern about crowding when choosing a destination. Participants were asked to rate their concerns about crowding using a 5-point Likert scale (from "very worried" to "not worried"). The second part investigated the characteristics of the trip itself, recording the departure and destination, road length, trip purpose, trip duration, etc., and representing the destination in a map graphical manner. Participants were specifically asked to recall and describe in detail one or more trip experiences starting from or going to a nearby subway station.

[0180] Questionnaire Collection Method: Data collection was systematically conducted through face-to-face interviews. This process ensured direct interaction with the respondents, thus facilitating more accurate data capture.

[0181] Implementation Team: Data collection was completed by a team of 12 specially trained investigators. The survey was completed between June and August 2023.

[0182] Sample selection: Participants were randomly selected near two designated subway stations (People’s Park Station and Kuanzhai Alley Station) and had travel experience in these two subway stations. Their purpose of travel was non-commuting, ensuring the diversity and representativeness of the sample.

[0183] Data volume: A total of 921 individual responses were collected. After careful screening and verification, 871 of them were determined to be valid responses, which makes the questionnaire validity rate reach about 95%.

[0184] Implementation schedule: In order to fully understand travel patterns at different time periods, interviews were conducted on both weekdays and weekends.

[0185] The data were entered into a spreadsheet and carefully screened and verified to ensure data validity and accuracy.

[0186] Example of questionnaire data:

[0187] Sample characteristics: Table 3 shows the demographic characteristics of the participants, such as age, gender, and frequency of visiting the neighborhood.

[0188] Table 3 Questionnaire sample statistics

[0189]

[0190]

[0191] Of the 871 respondents, 112 recalled making more than one trip to and from the subway station, with a total of 1,143 trips recorded. Figure 2 A map representation of the distribution of tourist destinations within the specified study area is shown. It is noteworthy that there is a clear concentration of destinations near the two identified stations, which include tourist attractions and places characterized by a significantly higher density of POIs. This concentration is further emphasized by the dense clustering marked in red. The comprehensive range of destinations within the study area emphasizes their attractiveness for travelers and enriches the heterogeneity of the study sample.

[0192] Model calculation:

[0193] The specific modeling steps are as follows: Figure 2 shown.

[0194] Data Split: The dataset is divided into training set (75%) and testing set (25%).

[0195] Feature and label separation: Separate the features (X) and target variables (y) from the dataset.

[0196] Model initialization: Use the RandomForestClassifier class in the Scikit-learn library to create a random forest model instance, setting parameters to n_estimators = 200, max_features = 2, max_depth = None.

[0197] Model training: The random forest model is trained using the training set data and hyperparameters are tuned using cross-validation techniques such as GridSearchCV.

[0198] Model evaluation: Calculate the out-of-bag error rate and evaluation indicators on the test set to ensure that the model has good generalization ability.

[0199] Results:

[0200] The SHAP value analysis results are as follows Figure 3 As shown, it shows the degree of influence of different features on model prediction and the mode of action (positive or negative). Figure 4 , Figure 5 , Figure 6 , Figure 7 It shows the nonlinear impact of certain factors, such as road length, average number of floors of buildings, green area, tourist attractions POI and other key variables on non-commuters' destination choice.

[0201] Non-commuting travel destination forecast:

[0202] Taking Chengdu Tianfu Square subway station as the prediction target area, repeat the above steps to obtain the built environment factor value. Then use the random forest model results to perform prediction calculations and obtain the predicted distribution data of the destination of the local non-commuting travel crowd. Figure 8 shown.

[0203] Prediction target area

[0204] The prediction target area is the Tianfu Square subway station in Chengdu. Tianfu Square Station is located in the center of Chengdu, surrounded by abundant tourist attractions, commercial facilities and entertainment venues, and is a typical hot spot for non-commuters.

[0205] Predictive Calculation

[0206] The built environment factor values ​​around Tianfu Square subway station were input into the random forest model.

[0207] The model is used to perform predictive calculations and obtain the destination distribution data of non-commuting people.

[0208] Results

[0209] Figure 8 The heat map of the predicted distribution of destinations for non-commuters is shown. This map intuitively shows the attractiveness of different areas to non-commuters.

[0210] Suggested applications of prediction results:

[0211] Urban planning suggestions: Increase green space, optimize bus stop layout, and adjust building density to improve the attractiveness of areas around subway stations.

[0212] Traffic management strategies: Improve road design, enhance the connectivity of public transportation services, and improve the travel experience of non-commuters.

[0213] Commercial layout optimization: Provide location suggestions to merchants to help them find the best location around subway stations to attract more non-commuters.

Claims

1. A method for predicting destination selection of non-commuters in the subway area based on built environment, characterized in that: The following steps are involved: Step 1: Data collection: Select a research area in the target city and collect relevant built environment data within the subway station catchment area, including but not limited to environmental data, point of interest (POI) data, and traffic data; in addition, it is necessary to obtain the travel records and destination selection information of respondents who take the subway for non-commuting trips; Step 2: Feature selection: Identify important built environment variables based on the literature review and preliminary analysis, and add variables that may have an impact on non-commuters based on local special circumstances; Step 3: Model construction: Use the random forest algorithm to build a prediction model; use the Scikit-learn module in the Python programming language to implement the random forest technology, divide the data set selected by the travelers into a training set and a test set; set the model parameters as the number of decision trees and the number of variables considered for each split, and do not limit the maximum tree depth; then, perform model training and validation, use the training set data to train the random forest model, and evaluate the performance of different model variants through the out-of-bag error rate; after determining the optimal model, use the test set for validation to ensure that the model has good generalization ability; Step 4: SHAP analysis: Apply SHAP technology to analyze the output of the random forest model to clarify the importance ranking of each built environment factor on the destination choice of non-commuters and its mechanism of action; Step 5: Interpretation and prediction of results: Determine which built environment elements are most or least attractive to non-commuters based on the SHAP value; adjust the corresponding facility layout or service supply based on the actual scenario to determine the numerical matrix of each built environment factor in the new scenario; substitute the above model coefficient results into the new built environment factor matrix to calculate the final destination selection results of each non-commuting travel individual under this new scenario.

2. The method for predicting destination selection of non-commuters in the subway periphery based on built environment according to claim 1, characterized in that: The step 1 is specifically as follows: S1.1: Identify the study area: Specific subway stations in the target cities were selected as research objects, and a research area with a radius of 1,000 meters was defined with each station as the center; S1.2: Field investigation and data acquisition: Environmental data: collect information about the built environment through field measurements, street view maps, and high-resolution satellite imagery; including but not limited to sidewalk width, green area, average number of floors in a building, and building density; POI data: records the type, location and number of various points of interest in the study area; Traffic data: statistics on the number and distribution of bus stops, as well as road grade information; S1.3: Questionnaire design and distribution: Questionnaire content: A structured paper questionnaire was designed, divided into two parts; the first part collected the demographic information of the respondents, including but not limited to age, gender, and family composition; the second part recorded the travel characteristics of the respondents, including but not limited to the departure and destination, road length, travel purpose, and travel duration; Questionnaire distribution: Randomly select participants near the selected subway stations for face-to-face interviews; ensure that the questionnaires are evenly distributed on weekdays and weekends to obtain comprehensive travel pattern data; the selected dates should exclude extreme weather conditions; S1.4: Data collation and screening: First, data entry is performed, and the collected questionnaire data is entered into a spreadsheet or database; second, the entered data is carefully screened and verified, and invalid or incomplete answers are eliminated to ensure the validity and accuracy of the data; finally, sample characteristic analysis is performed, and descriptive statistical analysis of the demographic characteristics of the valid samples is performed to ensure that the samples are representative; S1.5: Geographic Information System Applications: Use GIS software to process the collected spatial data, including POI locations, road networks, and bus stop distribution; create a map representation of the study area to visually display the distribution of different types of built environment elements and points of interest; S1.6: Comprehensive dataset construction: Integrate all collected data into a unified dataset, including built environment variables, POI data, traffic data, and travel records of non-commuters; Preprocess the data, including missing value filling, outlier detection and processing, and data standardization, to ensure that the data quality meets the requirements of subsequent modeling.

3. The method for predicting destination selection of non-commuters in the subway periphery based on built environment according to claim 1, characterized in that: The step 2 is specifically as follows: S2.1: Literature review and preliminary screening: Review relevant research literature to understand which built environment factors have been shown to have a significant impact on non-commuters' destination choices, including but not limited to road length, building coverage, green area, sidewalk width, average number of floors in buildings, building plot coverage, land use mix, number of bus stops, and density of various POIs; S2.2: Data Exploratory Analysis: Conduct descriptive statistical analysis on the collected data, including calculating the maximum, minimum, standard deviation and mean of each variable to understand the basic distribution of the data; evaluate the correlation between the built environment variables by calculating the Pearson or Spearman correlation coefficient, identify highly correlated variables and avoid multicollinearity problems; use box plots or scatter plots to visually display the distribution of each variable and its relationship with other variables, and help identify outliers and potential important features; S2.3: Final feature list: Based on the above results, the final list of features used for modeling is determined; These features should have good predictive power and no serious collinearity between each other; The definition, origin, and role of each selected feature in the model are documented in detail to facilitate subsequent interpretation and application.

4. The method for predicting destination selection of non-commuters in the subway periphery based on built environment according to claim 1, characterized in that: The step 3 is specifically as follows: S3.1: Data preparation: Data splitting: Divide the collected data set into training set and test set, using 75% of the data as the training set and 25% of the data as the test set. This can be achieved through the `train_test_split` function in Python's Scikit-learn library. Separation of features and labels: separating features X and target variables y from the dataset; features include built environment factors, and the target variable is the destination choice of non-commuters; S3.2: Model initialization: Select the Random Forest algorithm and create a Random Forest model instance using the `RandomForestClassifier` or `RandomForestRegressor` class from the Scikit-learn library, depending on whether the target variable is a classification or regression problem; Set the initial parameters, according to the literature and preliminary experiments, to set the initial parameters of the random forest model, including the number of decision trees n_estimators, the number of variables considered for each split max_features, and the maximum tree depth max_depth; S3.3: Model training: Fitting model: Use the training set data to train the random forest model, that is, call the `fit` method; this process will generate multiple decision trees based on the training data and make predictions through the integration of these trees; Hyperparameter tuning: Use cross-validation techniques to search for the best hyperparameter combination to further optimize model performance; S3.4: Model Evaluation: Out-of-bag error rate evaluation: During the training process, the error rate of out-of-bag (OOB) samples is calculated as an indicator of the generalization ability of the model. If the OOB error rate is low, it indicates that the model has good generalization ability. Test set verification: Use the test set data to verify the model and calculate evaluation indicators such as accuracy, recall, and F1 score to ensure that the model not only performs well on the training set, but can also make accurate predictions on unseen data.

5. The method for predicting destination selection of non-commuters in the subway periphery based on built environment according to claim 1, characterized in that: The step 4 is specifically as follows: S4.1: Model calculation: Apply SHAP analysis: Use SHAP values ​​to interpret the output of the random forest model and understand the impact of each feature on the model prediction; A significant enhancement introduced by SHAP is its ability to depict explanations through a linear model framework that can be expressed mathematically in equations (1) and (2); this is known as the additive variable attribution approach, which aggregates individual effects in addition to the intercept to illuminate the explanatory dynamics of the model; In the formula, g is the framework of the explanatory model, the number of input variables is represented by P, the variable z represents the presence of the corresponding variable 1 or the absence of 0, the symbol Φ0 is the constant term, Φ j represents the attribute value assigned to variable j; where S represents a subset of the variables used, the variable p represents the total number of variables, and the function g(z')(S∪{j}) represents the output value of the model given the combination of variables in the subset S; regarding the concept of weight, for p individual variables, given their sequence, the permutation of these variables is equal to p! combinations; when the position of a particular variable j is fixed, the remaining permutations involving other variables are given by |S|! (p-|S|-1)!, where |S| is the cardinality of the subset S, indicating the number of variables it contains; Draw SHAP diagrams: Create SHAP summary diagrams and dependency diagrams to intuitively display the importance of each feature and its relationship pattern with the target variable; S4.2: Global explanation: SHAP summary plot: Draw a SHAP summary plot to show the importance ranking of all features and their overall impact on the model output; this can be achieved through shap.summary_plot(shap_values,X_test); in the summary plot, each point represents a sample, the color represents the actual value of the feature, and the horizontal axis represents the SHAP value, that is, the impact of the feature on the prediction result of the sample; Mean Absolute SHAP Value: Calculate the mean absolute SHAP value of each feature to quantify its average effect on the model output; S4.3: Local interpretation: SHAP dependency graph: For the selected key features, draw a SHAP dependency graph to show the relationship between the feature value and its corresponding SHAP value; Individual sample explanation: Select a specific sample and generate a visualization of its SHAP value to understand the specific contribution of each feature in the sample to its prediction result; this can be achieved by shap.force_plot(explainer.expected_value,shap_values[0,:],X_test.iloc[0,:]), where shap_values[0,:] and X_test.iloc[0,:] correspond to the SHAP value and feature value of the first sample respectively; S4.4: Interpretation of results: Based on the SHAP summary graph and the average absolute SHAP value, determine which built environment factors have the greatest impact on non-commuters’ destination choice; The nonlinear relationship between features and destination choice was analyzed through SHAP dependency plots; the SHAP values ​​were checked to see if they showed any interactions between features, especially the variation patterns observed under different feature combinations.

6. The method for predicting destination selection of non-commuters in the subway periphery based on built environment according to claim 1, characterized in that: The step 5 is specifically as follows: S5.1: Summarize key findings: Main influencing factors: Based on the analysis of SHAP values, the most important built environment factors are identified; Non-linear relationship revelation: For variables that show significant non-linear relationships, describe the changing trend of their impact on destination selection as the variable value changes; S5.2: Interpretation of results: Explain why certain factors have a greater impact on destination choice, emphasizing the importance of accessibility and aesthetic factors; discuss those factors that show context-dependent effects, showing that they may have different effects in different contexts; S5.3: Case studies: Specific case analysis: Select representative cases to analyze in detail how built environment factors work together to influence non-commuters’ destination choices; use maps and charts to intuitively display the key variables and their impacts in these cases; extract quantitative values ​​of each built environment factor based on the calculated interactive influence relationship mechanism and the characteristics of the built environment factors in the cases; S5.4: Suggestions and Feedback: Urban planning recommendations: Based on the research results, provide specific recommendations to urban planners, including but not limited to increasing green space, optimizing bus station layout, and adjusting building density to improve the attractiveness of areas around subway stations; Traffic Management Strategy: Propose improvement measures for traffic management departments, including improving road design and enhancing the connectivity of public transportation services to enhance the travel experience of non-commuters; Commercial layout optimization: Provide location suggestions to merchants to help them find the best location around subway stations to attract more non-commuters.

Images