A method for decoupling spatial heterogeneity and nonlinear effects based on dual-model visualization translation
By employing a dual-model visualization translation method, the spatial heterogeneity and nonlinear effects of the built environment are decoupled, solving the problem of data feature confusion in existing studies and enabling precise commuting behavior analysis and quantitative support for urban transportation planning strategies.
Patent Information
- Application Number
- CN202611242998.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing research, when exploring the impact of the built environment on commuting behavior, has failed to effectively distinguish the differentiated impacts of the residential end and the employment end, confused the data characteristics of residential choice and employment distribution, failed to provide scientific data support for the precise allocation of public transportation resources, and failed to analyze the intertwined effects of nonlinearity and spatial heterogeneity, resulting in systematic biases in the data analysis conclusions.
A dual-model visualization translation approach is adopted to construct a complete data processing chain, including multi-source data standardization processing, commuting data screening and clustering, dual-model collaborative numerical calculation, and effect decoupling quantitative analysis. By independently modeling the MGWR and GBRT models, the spatial heterogeneity and nonlinear effects of the built environment are decoupled, and a quantitative analysis report is generated.
It enables precise analysis of the impact of the built environment, outputs differentiated urban transportation and land use planning strategies adapted to different commuter groups and location characteristics, improves the spatial fairness of urban transportation and public resource allocation, and provides quantitative data-driven support.
Smart Images

Figure CN123117518A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical digital data processing technology, and in particular to a method for decoupling spatial heterogeneity and nonlinear effects based on dual-model visualization translation. Background Technology
[0002] With the rapid advancement of urbanization, the imbalance between urban work and residence spaces has become increasingly severe, leading to a surge in rigid commuting demand for long-distance and cross-regional travel. This, in turn, results in extremely low commuting efficiency of public transportation during peak hours (such as severe tidal congestion on rail transit and high empty-load rates on regular buses). To alleviate this problem, TOD-based planning has become mainstream. However, existing inventions still face many technical bottlenecks when exploring the impact of the built environment on commuting behavior: existing studies often treat commuting as a homogeneous flow, and the traditional data processing models used cannot effectively distinguish the differentiated impacts of the built environment on the residential end (occurrence) and the employment end (attraction), confusing the "push" of residential choices with the "pull" of employment distribution, thus failing to provide scientific data support for the precise allocation of public transportation resources.
[0003] While existing inventions have partially addressed the issue of the impact of the built environment on commuting behavior, they still suffer from numerous key shortcomings and technical bottlenecks: (1) Existing research treats commuting as a homogeneous flow, ignoring the essential asymmetry of the built environment in shaping residential and employment locations. It fails to distinguish the differences in the built environment between the starting point (residential location) and the ending point (employment location) of commuting, and confuses the two types of data characteristics of residential choice and employment distribution, thus failing to provide accurate targeted data support for optimizing the work-life balance.
[0004] (2) With the widespread application of large-scale smart card data, it has become a consensus in academia that commuters are by no means a single homogeneous group. With the emergence of diverse commuter groups—from "rigid long-distance commuters who are highly dependent on the subway" to "flexible short-distance commuters who rely on regular buses"—different commuter clusters often show significant differentiation in their sensitivity to the built environment in their place of residence and workplace. Does the urban form centered on rail transit, while benefiting long-distance commuters, marginalize short-distance commuters who are mainly engaged in local activities and highly dependent on public transportation? Quantitatively analyzing the differentiated response data characteristics of different groups to urban space is crucial for designing inclusive TOD governance strategies.
[0005] (3) Existing mainstream quantitative analysis data processing paradigms all have inherent technical limitations. Machine learning data processing algorithms, such as Gradient Boosting Regression Tree (GBRT), suffer from opaque internal decision-making logic and cannot fully handle the inherent spatial heterogeneity of the data. Spatial statistical data processing algorithms, such as Multi-Scale Geographically Weighted Regression (MGWR), are unable to effectively extract potential nonlinear relationships and threshold effects from the data. Existing technologies have not yet achieved deep integration of the two types of models, and cannot simultaneously and accurately analyze the intertwined effects of nonlinearity and spatial heterogeneity, making it difficult to comprehensively and objectively reveal the complex impact mechanism of the built environment on public transportation commuting demand.
[0006] (4) Most existing data processing schemes focus on exploring the spatial heterogeneity or nonlinear threshold effect of the built environment. They have not yet achieved data-level separation and joint quantification of the characteristics of the two types of effects. It is impossible to clearly determine whether the observed spatial variation is a real location-dependent preference or an external manifestation of a potential nonlinear density threshold. This can easily lead to misjudgment of "scale dependence" and "location dependence". Relying solely on a single data processing model can lead to systematic biases in the data analysis conclusions. It is impossible to fully quantify the complex data characteristics of the two types of effects that are intertwined.
[0007] Overall, existing analytical methods cannot simultaneously decouple the spatial heterogeneity and nonlinear intertwined effects of the built environment's impact, nor can they accurately reveal the asymmetric impact mechanism of the built environment on residential and employment site selection, making it difficult to meet the actual needs of refined urban transportation planning and work-life space optimization in the new era. Summary of the Invention
[0008] This invention provides a method for decoupling spatial heterogeneity and nonlinear effects based on dual-model visualization translation, and constructs a full-process data processing chain covering multi-source data standardization processing, commuting data screening and clustering, dual-model collaborative numerical calculation, effect decoupling quantitative analysis, and planning parameter output. First, by directly utilizing integrated public transport and subway travel chain data, bypassing the complex underlying card-swiping data preprocessing stage, rapid access and standardized processing of public transport commuting characteristic data are achieved. Second, a two-step clustering method is used to process travel data, extracting effective commuter data records and spatial information of residence and workplace. Based on spatiotemporal characteristic indicators, commuter groups are segmented, overcoming the shortcomings of traditional research that treats commuters as a homogeneous group. Subsequently, a 1000m×1000m rectangular grid is used as a unified spatial analysis and modeling unit. A dual-model analysis framework integrating multi-scale geographic weighted regression (MGWR) and gradient boosting regression tree (GBRT) is built, independently modeling residential and employment locations respectively, simultaneously capturing the spatial heterogeneity and nonlinear effects of built environment impact, achieving decoupling and visual translation of the two types of composite effects. Finally, through multi-dimensional comparative analysis across ends and groups, a quantitative analysis report on the asymmetric impact of the built environment on public transport commuting demand is generated, outputting differentiated urban transportation and land use planning strategies adapted to different commuter groups and different location characteristics. This invention provides a complete and practical technical approach for the accurate analysis of the correlation between the built environment and commuting behavior.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for decoupling spatial heterogeneity and nonlinear effects based on dual-model visual translation includes: S1. Obtain the spatiotemporal trajectory data of the moving object, extract the anonymous identifier, spatial coordinates of trajectory points and corresponding timestamps from it, and construct a standardized vector dataset; S2. Construct a two-layer data filtering algorithm based on spatiotemporal density clustering and spatial density clustering to extract spatiotemporal data records of regularly moving objects from a standardized vector dataset, and output the starting and ending spatial centroids of the regularly moving objects. S3. Based on the centroid of the starting space and the centroid of the ending space, extract the multidimensional spatiotemporal behavioral features of each regular moving object, perform clustering operations on the feature data of the regular moving objects, and generate multiple movement pattern data clusters. S4. Using a rectangular grid of a preset size as a unified spatial analysis unit, spatially associate the starting and ending spatial centroids with the rectangular grid, statistically analyze the distribution intensity of target events and the feature vector of spatial covariates within each grid unit, and construct a modeling dataset. S5. Based on the modeling dataset, a multi-scale geographically weighted regression model and a gradient boosting regression tree model are independently constructed for the starting and ending ends, respectively, forming a dual-model parallel modeling framework. The output includes the local regression coefficients, nonlinear dependencies, threshold feature parameters, variable importance ranking, and dependency curves output by the additivity interpretation method for each spatial covariate. S6. Calculate the first derivative of the dependency curve to obtain the marginal effect change data of each spatial covariate. Perform cross-validation and trend comparison between the local regression coefficients and the marginal effect change data, and output the dominant effect type data of each spatial covariate. S7. Based on the dominant effect type data, extract the variable importance ranking, dominant effect type, threshold feature parameters and local regression coefficient spatial distribution matrix of each spatial covariate at the start and end points, quantify the differential driving parameters of each mobile mode data cluster at the start and end points, output the response difference data between each mobile mode data cluster, and generate a feature vector set for spatiotemporal data effect separation. S8. Based on the feature vector set and dominant effect type data, generate a structured spatial configuration intervention parameter set.
[0010] In summary, the present invention has at least the following beneficial effects: This invention proposes an analytical framework for visual translation of the MGWR-GBRT dual-model system. Through cross-validation and trend comparison of the results of the two models, it achieves precise decoupling of spatial heterogeneity and nonlinear intertwined effects in the impact of the built environment, effectively solves the problem of conclusion bias caused by single-model analysis, and fills the methodological gap in existing research. This invention employs a data processing design that uses independent modeling for both residential and employment locations. It enables precise calculation and quantitative output of the parameters that affect the differentiating impacts of the same built environment variable on residential and employment location. It generates a data feature map of the asymmetric impact of the built environment on public transportation commuting demand, providing quantitative data-targeted support for TOD strategy optimization and improvement of work-life balance. This invention classifies commuters based on their spatiotemporal travel patterns, systematically analyzes the differentiated responses of different commuter groups to the built environment, and can output inclusive planning strategies that adapt to the needs of different groups, effectively improving the fairness and applicability of urban transportation and land use policies. This invention breaks through the limitations of traditional research, which can only offer macro-level policy recommendations. It can directly output a work-residential space intervention map containing specific spatial coordinate grids and quantitative facility allocation thresholds (such as a basic guarantee threshold of 10 facilities / km2). Through these digital spatial configuration parameters, it provides robust quantitative technical support for urban micro-renewal and precise TOD planning.
[0011] This invention, by locating the highly sensitive areas of different heterogeneous data clusters to specific spatial anchor points, can generate planning and control schemes that directionally increase the floor area ratio of enterprises in specific peripheral sub-center grids to intercept long-distance commutes. Simultaneously, it marks the spatial location of the residential security red line for data clusters of short-distance, disadvantaged public transport commuters in the core area. Through this spatial parameter intervention based on underlying big data, extreme tidal commuting data can be effectively reduced, fundamentally optimizing the public transport commuting structure and greatly improving the spatial equity of urban transportation and public resource allocation.
[0012] The method of this invention completes the full-process verification based on real urban public transportation smart card data. Compared with traditional linear models, single spatial models, and single machine learning models, it has significant advantages in terms of goodness of fit, explanatory power of mechanism, and applicability of planning, and has strong practical application value and promotion potential. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of a method for decoupling spatial heterogeneity and nonlinear effects based on dual-model visualization translation.
[0014] Figure 2 The diagram shows the visual translation and mechanism decoupling of the distance from the city center to the residential end for cluster 0.
[0015] Figure 3 This is a visualization diagram of the distance from the city center to the residential area for cluster 1, showing the mutual translation and mechanism decoupling between the two models.
[0016] Figure 4 This is a visualization diagram of the distance from the city center of cluster 2 at the residential end, showing the mutual translation and mechanism decoupling between the two models.
[0017] Figure 5 The distance from the city center for cluster 0 is visualized at the working end as a dual-model translation and mechanism decoupling diagram.
[0018] Figure 6 The diagram shows the bi-model visualization and mechanism decoupling of the distance from the city center of cluster 1 at the working end.
[0019] Figure 7 The diagram shows the bi-model visualization and mechanism decoupling of the distance from the city center of cluster 2 at the working end.
[0020] Figure 8 The visualization of the distance to the nearest subway station for cluster 0 at the residential end is shown in the diagram of the dual-model translation and mechanism decoupling.
[0021] Figure 9 This is a visualization diagram of the distance to the nearest subway station for cluster 1 at the residential end, showing the mutual translation and mechanism decoupling between the two models.
[0022] Figure 10This is a visualization diagram of the distance to the nearest subway station in cluster 2 at the residential end, showing the mutual translation and mechanism decoupling between the two models.
[0023] Figure 11 The distance to the nearest subway station for cluster 0 is visualized in the dual-model translation and mechanism decoupling diagram at the working end.
[0024] Figure 12 The diagram shows the bi-model visualization and mechanism decoupling of the distance to the nearest subway station for cluster 1 at the working end.
[0025] Figure 13 The diagram shows the bi-model visualization and mechanism decoupling of the distance to the nearest subway station for cluster 2 at the working end.
[0026] Figure 14 A visualization of the dual-model translation and mechanism decoupling of the commercial facility density at the residential end for cluster 0.
[0027] Figure 15 The visualization of the commercial facility density in cluster 1 at the residential end is shown in the diagram of the dual-model translation and mechanism decoupling.
[0028] Figure 16 The visualization of the commercial facility density in cluster 2 at the residential end is shown in the diagram of the dual-model translation and mechanism decoupling.
[0029] Figure 17 The visualization of the inter-model translation and mechanism decoupling diagram of the educational facility density at the working end for cluster 0.
[0030] Figure 18 The visualization of the dual-model translation and mechanism decoupling diagram of the educational facility density at the working end for cluster 1.
[0031] Figure 19 The visualization of the inter-model translation and mechanism decoupling of the educational facility density in cluster 2 at the working end is shown in the figure. Detailed Implementation
[0032] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0033] like Figure 1 As shown, this embodiment provides a method for decoupling spatial heterogeneity and nonlinear effects based on dual-model visual translation, including the following steps: A method for decoupling spatial heterogeneity and nonlinear effects based on dual-model visual translation includes: S1. Obtain the spatiotemporal trajectory data of the moving object, extract the anonymous identifier, spatial coordinates of trajectory points and corresponding timestamps from it, and construct a standardized vector dataset; S2. Construct a two-layer data filtering algorithm based on spatiotemporal density clustering and spatial density clustering to extract spatiotemporal data records of regularly moving objects from a standardized vector dataset, and output the starting and ending spatial centroids of the regularly moving objects. S3. Based on the centroid of the starting space and the centroid of the ending space, extract the multidimensional spatiotemporal behavioral features of each regular moving object, perform clustering operations on the feature data of the regular moving objects, and generate multiple movement pattern data clusters. S4. Using a rectangular grid of a preset size as a unified spatial analysis unit, spatially associate the starting and ending spatial centroids with the rectangular grid, statistically analyze the distribution intensity of target events and the feature vector of spatial covariates within each grid unit, and construct a modeling dataset. S5. Based on the modeling dataset, a multi-scale geographically weighted regression model and a gradient boosting regression tree model are independently constructed for the starting and ending ends, respectively, forming a dual-model parallel modeling framework. The output includes the local regression coefficients, nonlinear dependencies, threshold feature parameters, variable importance ranking, and dependency curves output by the additivity interpretation method for each spatial covariate. S6. Calculate the first derivative of the dependency curve to obtain the marginal effect change data of each spatial covariate. Perform cross-validation and trend comparison between the local regression coefficients and the marginal effect change data, and output the dominant effect type data of each spatial covariate. S7. Based on the dominant effect type data, extract the variable importance ranking, dominant effect type, threshold feature parameters and local regression coefficient spatial distribution matrix of each spatial covariate at the start and end points, quantify the differential driving parameters of each mobile mode data cluster at the start and end points, output the response difference data between each mobile mode data cluster, and generate a feature vector set for spatiotemporal data effect separation. S8. Based on the feature vector set and dominant effect type data, generate a structured spatial configuration intervention parameter set for downstream display terminals to perform data visualization mapping.
[0034] In some implementations, step S2 specifically includes: S21. Extract the first component of the spatial coordinates of the starting point in the individual trajectory chain, the second component of the spatial coordinates, and the starting time to form a three-dimensional feature space. Execute the spatiotemporal density clustering algorithm and mark the cluster with the earliest average starting time as the starting cluster. S22. Based on the spatial coordinates of the termination point in the individual trajectory chain, perform spatial density clustering to generate termination clusters. Mark the termination clusters whose centroid distance from the starting cluster is less than a spatial threshold as termination arrival areas. The starting clusters and termination arrival areas jointly define the potential starting spatial range. S23. Extract the set of trajectory endpoints from the starting cluster and the set of trajectory starting points to the destination area of the termination point as candidate points, perform spatial density clustering, mark the largest cluster as the main termination cluster, and retain the objects that are successfully associated with the main termination cluster as regular moving objects.
[0035] In some implementations, step S3 specifically includes: The system extracts the standard deviation of start time, the proportion of activities during peak hours, and the average activity time from the time dimension, and extracts the average moving distance, the symmetrical round-trip rate, and the number of active anchor points from the spatial dimension, thus forming a six-dimensional quantitative feature system. After standardizing the six-dimensional quantitative feature system, the K-Means clustering algorithm is input. The number of clusters is determined by the elbow rule. Based on the data distribution centroid of each cluster in each dimension of the six-dimensional quantitative feature system, the regular moving objects are marked as the first moving pattern data cluster, the second moving pattern data cluster, and the third moving pattern data cluster.
[0036] In some implementations, step S4 specifically includes: S41. Generate a rectangular grid layer covering the entire area based on the geographical boundary of the study area, and spatially overlay and associate the starting spatial centroid and the ending spatial centroid with the rectangular grid layer, and statistically analyze the generation intensity of the starting end event and the attraction intensity of the ending end event in each grid cell. S42. Using the rectangular grid as the measurement unit, extract the distribution density of interest points, the entropy value of spatial element mixing degree, the shortest spatial distance from the grid center to the core of the region, and the shortest spatial distance from the grid center to the nearest transportation hub, as spatial covariate feature vectors. S43. Normalize the intensity of the starting event, the attraction intensity of the ending event, and the feature vector of the spatial covariate to complete the spatial attribute coupling on the same grid scale and construct the modeling variable dataset. S44. Perform a variance inflation factor test on the modeling variable dataset, detect and remove independent variable data whose variance inflation factor values are greater than a preset threshold, and calculate the global Moran index of each spatial covariate to verify the spatial autocorrelation characteristics. S45. The intensity of the starting event and the intensity of the ending event in each grid cell are standardized according to the area of the grid cell to obtain the density of the starting event and the density of the ending event in each grid cell.
[0037] In some implementations, step S5 specifically includes: Using the starting event density of each mobile mode data cluster within each grid cell as the starting dependent variable, the ending event density of each mobile mode data cluster within each grid cell as the ending dependent variable, and the spatial covariate feature vector as the independent variable, based on the modeling dataset, the fitting training of the multi-scale geographic weighted regression model and the gradient boosting regression tree model is carried out independently for the starting dataset and the ending dataset respectively. By using a multi-scale geographic weighted regression model, the local regression coefficients and spatial distribution matrix of each spatial covariate are output, and the spatial heterogeneity characteristics of the influence of spatial covariates on the spatiotemporal patterns of moving objects are extracted. Gradient boosting regression tree model is used to extract the nonlinear influence features of spatial covariates on the behavior of moving objects. The model output is interpreted by combining the additive interpretation method, and the nonlinear dependencies of each spatial covariate, threshold feature parameters, variable importance ranking, and dependency curves output by the additive interpretation method are obtained.
[0038] In some implementations, step S6 specifically includes: S61. Convert the local regression coefficients of the multi-scale geographical weighted regression model into a scatter plot with spatial covariate values on the x-axis and local regression coefficients on the y-axis, and fit a polynomial trend curve. S62. Calculate the first derivative of the dependency curve output by the gradient boosting regression tree model combined with the additivity interpretation method to obtain the marginal effect change data of each spatial covariate. S63. Superimpose and compare the polynomial trend curve with the curve of the marginal effect change data, and output the dominant effect type data of each spatial covariate based on the curve matching degree and coefficient dispersion. When the multi-scale geographically weighted regression coefficients exhibit high dispersion under the same variable values, while the curve of the marginal effect change data remains relatively flat, the dominant effect type of the output spatial covariate is spatial heterogeneity-dominated; when the polynomial trend curve and the curve of the marginal effect change data show high similarity, the dominant effect type of the output spatial covariate is nonlinear effect-dominated; when both of the above characteristics are satisfied, the dominant effect type of the output spatial covariate is a coexistence of spatial heterogeneity and nonlinear effects.
[0039] In some implementations, step S7, which involves extracting the variable importance ranking, dominant effect type, threshold feature parameters, and spatial distribution matrix of local regression coefficients for each spatial covariate at the starting and ending ends, specifically includes: The variable importance ranking, the dominant effect type, the threshold feature parameter, and the spatial distribution matrix of the local regression coefficients are all derived from the output results of the dual-model parallel modeling in step S5 and the dominant effect type data in step S6. By comparing the importance ranking, dominant effect type, effective action threshold parameter, and spatial distribution matrix of local regression coefficients of the same spatial covariate at the start and end points, the differential characteristic data of the direction and intensity of the influence of the variable between the start and end points are quantified, which serve as the core data content of the feature vector set for effect separation.
[0040] In some implementations, step S8 specifically includes: For spatial covariates whose output is dominated by nonlinear effects, extract and output their effective threshold parameters; For spatial covariates whose output is dominated by spatial heterogeneity, locate and output a set of raster spatial coordinates where local regression coefficients show a high positive effect; Based on the effective action threshold parameter and the grid spatial coordinate set, a spatial configuration intervention map is generated as a quantitative data reference for the structured spatial configuration intervention parameter set.
[0041] In some implementations, step S8, generating a structured set of spatial configuration intervention parameters, specifically includes: For spatial covariates whose output is dominated by nonlinear effects, a first type of data labeling parameters are generated based on the effective action threshold parameter. For spatial covariates whose output is dominated by spatial heterogeneity, a second type of data labeling parameters are generated based on the set of raster spatial coordinates that exhibit high positive effects from the local regression coefficients. Based on the response parameter difference data between each mobile mode data cluster extracted in step S7, and combined with the differential driving factor data of each mobile mode data cluster obtained by quantification in step S7, a third type of data annotation parameters are generated. The first type of data annotation parameters, the second type of data annotation parameters, and the third type of data annotation parameters are integrated to form a structured set of spatial configuration intervention parameters for downstream display terminals to perform data visualization mapping.
[0042] In some implementations, step S4, before constructing the modeling dataset, further includes performing a variance inflation factor test on the spatial covariate feature vectors, detecting and removing independent variable data with variance inflation factor values greater than a preset threshold, and calculating the global Moran index of each spatial covariate to verify the spatial autocorrelation characteristics.
[0043] Taking urban public transportation commuting as an example, the technical concept of this invention is as follows: In this embodiment, an urban public transportation commuting scenario is used as a typical application of the method of the present invention for illustration. It should be understood that the data processing method proposed in this invention is universal, and its application is not limited to this. For ease of description, the general terms in this invention and the specific application terms in this embodiment are explained as follows: The spatiotemporal trajectory data of mobile objects specifically refers to public transportation commuting travel chain data; the anonymity identifier specifically refers to passenger anonymity identifiers; the spatial coordinates of trajectory points specifically refer to the spatial coordinates of the origin and destination points; the regular mobile objects specifically refer to regular commuters; the starting spatial centroid specifically refers to the spatial coordinates of the residence; the ending spatial centroid specifically refers to the spatial coordinates of the workplace; the starting departure cluster specifically refers to the main residence departure cluster; the ending arrival area specifically refers to the residential area of the drop-off point; the main ending cluster specifically refers to the main workplace cluster; the multidimensional spatiotemporal behavioral characteristics specifically refer to multidimensional spatiotemporal travel characteristics; the mobility pattern data cluster specifically refers to the commuter data cluster; the first mobility pattern data cluster specifically refers to the rigid long-distance data cluster; the second mobility pattern data cluster specifically refers to the balanced regular data cluster; the third mobility pattern data cluster specifically refers to the flexible short-distance data cluster; the target event distribution intensity specifically refers to residential density and employment density; the spatial covariate feature vector specifically refers to the built environment explanatory variables; the starting end Specifically, the data is categorized as follows: residential end; employment end; residential location dataset; employment location dataset; residential location dataset; employment location dataset; residential generation intensity; employment attraction intensity; point of interest distribution density; land use mixture entropy value; urban core; rail transit station; feature vector set for spatiotemporal data effect separation; quantitative analysis report on the asymmetric impact of the built environment on public transportation commuting demand; structured spatial configuration intervention parameter set; differentiated urban transportation and land use planning strategies; spatial configuration intervention map; work-residence space configuration intervention map; data annotation parameters; planning strategy content; downstream display terminal; display device for showcasing quantitative analysis report and planning strategies; and spatial distribution matrix.
[0044] In this embodiment, the spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation includes the following steps: S1. Constructing a Public Transportation Commuting Analysis Dataset: To focus on the analysis of the asymmetric influence mechanism of the built environment on the distribution of work and residence spaces, this invention uses high-quality integrated urban public transport and subway commuting travel chain data, which has undergone systematic preprocessing, as the input benchmark. In this embodiment, the system directly reads this integrated travel chain data and extracts a fully structured record containing the passenger's anonymous ID, the boarding time and coordinates of the origin, and the alighting time and coordinates of the destination.
[0045] S2. Extraction of effective commuter data and spatial data: Based on spatiotemporal density clustering (ST-DBSCAN) and spatial density clustering (DBSCAN), a two-layer data filtering algorithm is constructed to extract data records of regular commuters from integrated travel chain data, extract the spatial coordinates of their approximate residence and employment location, and filter the commuting travel records of effective commuters. S3. Spatiotemporal Pattern Clustering of Commuter Groups: Construct a six-dimensional quantitative feature system encompassing temporal flexibility, spatial span, and network complexity, and perform parameter optimization clustering. Build a spatiotemporal feature index system for commuting in both temporal and spatial dimensions, execute the K-Means clustering algorithm, and generate three types of commuter clusters, named "Balanced Regular Commuters," "Rigid Long-Distance Commuters," and "Flexible Short-Distance Commuters," respectively. S4. Constructing a spatial analysis unit adapted to the scale of urban blocks: Based on the 15-minute living circle scale of Chinese cities and the geometric characteristics of typical rectangular blocks, a 1000m×1000m rectangular grid is selected as a unified spatial analysis unit, and a modeling variable system including the intensity of job and residence endpoints and built environment factors is constructed accordingly. S5. Construction of Modeling Variable System: Define the dependent variables as residential density and employment density of three types of commuter clusters. Referring to the 5D framework of built environment, select density, diversity, destination accessibility, and public transportation distance as four dimensions as built environment explanatory variables. Clarify the definition, calculation method and statistical characteristics of each variable. Use a 1000m×1000m rectangular grid as a unified spatial analysis and modeling unit. Calculate the values of dependent and independent variables of each grid unit through point of interest (POI) data and geographic information system data to construct a formal modeling dataset. S6, MGWR-GBRT Dual-Model Visualization and Translation with Automatic Decoupling and Quantitative Analysis of Effects: This framework integrates spatial statistical algorithms and machine learning for interpretability, separating "location dependence" and "scale dependence." Independent models are constructed for residential and employment locations, creating an integrated multi-scale geographic weighted regression (MGWR)-gradient boosting regression tree (GBRT) dual-model analysis framework encompassing parallel modeling, interpretation and cross-validation, and decoupling and quantitative analysis. S7. Quantitative analysis of the asymmetric mechanism of the built environment's impact on residential and employment locations: Compare the variable importance, dominant effect type, threshold characteristics, and spatial distribution differences of the same built environment variable in the residential and employment models; analyze the differentiated driving factors and functional anchors of different commuter groups in residential and employment choices; and generate a quantitative analysis report on the asymmetric impact of the built environment on public transportation commuting demand. S8. Differentiated spatial intervention parameters and TOD quantitative planning scheme generation: Abandoning traditional subjective manual planning experience, this system uses quantitative analysis to obtain data on the asymmetric impact of housing and employment, and decouples the effects to quantify the model output into quantitative spatial configuration data. For built-up environment variables dominated by nonlinear effects, the system automatically extracts and outputs their effective threshold parameters (such as the effective connection radius of rail transit and the minimum density threshold for basic medical facilities). For variables dominated by spatial heterogeneity, the system locates and outputs a set of grid spatial coordinates with high marginal pull effects. Based on these quantitative parameters and coordinate sets, a high-resolution intervention map of work-housing space configuration is generated, providing quantitative data references for project implementation.
[0046] Preferably, step S1 includes the following steps: S11. Data Access: Read the entire chain of bus and subway commuting data after the underlying processing. This data has removed duplicate records, abnormal timestamps and invalid card swipe samples with missing key fields. S12. Travel Feature Analysis: The system automatically analyzes various connection modes defined in the travel chain, such as "subway-bus", "bus-subway" and "bus-bus", and extracts stable origin and destination (OD) geographical location information data in each travel chain. S13. Construction of the work-residence analysis database: Based on the unique passenger identifier (user card swipe ID), the full travel trajectory of the same user within the analysis period is aggregated to form a public transportation commuting trajectory database with spatiotemporal continuity, providing a high-quality data foundation for subsequent work-residence center data extraction.
[0047] Preferably, step S2 includes the following specific steps: S21, Spatiotemporal 3D Clustering Detection of Residential Departure Cluster Data (H1): Extract the longitude, latitude, and departure time data of the boarding point in the individual travel chain to form a 3D feature data space, execute the ST-DBSCAN algorithm, and mark the cluster with the earliest average departure time as the main residential departure cluster data. S22. Spatial two-dimensional clustering detection of destination cluster data (H2): Based on the latitude and longitude coordinates of individual drop-off points, DBSCAN clustering is performed to generate stable destination cluster data. Destination cluster data whose centroid distance from H1 is less than the spatial threshold are marked as drop-off point residential area data (H2). Potential residential area data (H) are jointly defined by H1 and H2. S23. Work-Residence Logical Association and Workplace Data (W) Extraction: Extract the trip destination data set starting from H1 and the trip origin data set arriving at H2 as candidate data points, perform high-intensity DBSCAN spatial clustering, mark the largest cluster as the main workplace cluster data (W), and retain only the user data records that have successfully associated with valid workplace cluster data as regular commuter data records.
[0048] Preferably, step S3 includes the following specific steps: S31. Quantification of Spatiotemporal Travel Characteristics Indicators: Extract departure time standard deviation, peak hour travel ratio, and average travel time from the time dimension; extract average commuting distance, symmetrical travel rate (reflecting the "home-work-home" closed loop ratio), and number of active stations from the spatial dimension. S32. Parameter optimization calculation based on the elbow rule: Calculate the sum of squared errors (SSE) under different K values, detect the significant "inflection point" data location of SSE decay, and calculate the optimal number of clusters as 3. S33. Heterogeneous Group Segmentation and Profile Generation: Based on the distribution center of indicators, commuters are divided into "rigid long-distance type" (high symmetrical travel rate, fixed point-to-point mode), "balanced regular type" (regularity and flexibility coexist) and "flexible short-distance type" (low travel tolerance, reliance on regular public transportation).
[0049] Preferably, step S4 includes the following specific steps: S41. Based on the geographical boundary of the study area, generate a 1000m×1000m rectangular raster layer covering the entire area; spatially correlate the work-residence coordinates extracted in S2 with the raster layer, and statistically analyze the residential generation intensity and employment attraction intensity within each raster unit to achieve standardized conversion from individual commuting trajectories to spatially continuous raster data.
[0050] S42. Multi-source built environment feature measurement and attribute coupling: Using a 1000m rectangular grid as the measurement unit, the density of multi-source POI points, the land use mixing entropy value, and the shortest spatial distance from the center of each grid to the urban core area and the nearest rail transit station are extracted and calculated simultaneously; normalization processing is performed to complete the high-dimensional coupling of spatial attributes between the extracted dependent variable (commuting generation / attraction intensity) and the independent variable (built environment features) at the grid scale; S43. Pre-testing of statistical robustness and evaluation of spatial applicability of variables: The variance inflation factor (VIF) test is performed on the coupled raster variable dataset to automatically detect and remove redundant independent variable data with VIF values greater than a preset threshold, so as to reduce the data interference of multicollinearity on the explanatory power of the model; the global Moran's index (Moran's I) of each variable is calculated simultaneously to verify the spatial autocorrelation characteristics of the variables under 1000m raster cells, determine the spatial regression applicability of the data in the MGWR model, and provide a statistically significant data foundation for subsequent decoupling and quantitative analysis of asymmetric effects.
[0051] Preferably, step S5 includes the following specific steps: S51. Defining and calculating dependent variables: The dependent variables are defined as residential density and employment density of three types of commuting data clusters. Residential density is the number of residential data records extracted from each grid cell of a specific commuting data cluster, standardized by the cell area (unit: records / square kilometer). Employment density is the number of employment data records extracted from each grid cell of a specific commuting data cluster, standardized by the cell area. S52. Selection and definition of explanatory variables: Referring to the 5D framework of built environment, four categories of explanatory variables are selected: density dimension (7 indicators including population density, enterprise density, commercial density, etc.), diversity dimension (land use mixing entropy value calculated based on POI data), destination accessibility dimension (distance from grid centroid to city center), and public transport distance dimension (density of bus stops within the grid and distance from grid centroid to the nearest subway station). S53. Variable standardization calculation: Based on the optimal spatial analysis unit of 1000m×1000m rectangular grid, the dependent and independent variables of each grid unit are calculated using point of interest (POI) data and geographic information system (GIS) data to ensure that the data format is uniform and the magnitude is comparable. S54. Modeling Dataset Construction: Integrate the variable values of all grid cells to form a formal modeling dataset containing dependent variables, explanatory variables, and spatial coordinate information, which will be used for subsequent parallel modeling of the two models.
[0052] Preferably, step S6 includes the following specific steps: S61, Parallel Modeling with MGWR-GBRT Dual Models: For each dataset of residential and employment locations, the MGWR and GBRT models are fitted and trained simultaneously. The MGWR model outputs the local regression coefficients and spatial distribution maps of each explanatory variable to capture the spatial heterogeneity of the built environment's influence on commuting spatiotemporal patterns. The GBRT model is used to analyze the nonlinear influence of built environment variables on commuting behavior. The SHAP additive interpretation method is combined to interpret the model output and obtain the nonlinear dependencies, threshold effects, and effective ranges of each variable. S62. Dual-model cross-validation and effect decoupling: The local regression coefficients of the MGWR model are converted into coefficient-variable scatter plots and fitted with a polynomial trend curve. At the same time, the first derivative of the SHAP dependency curve of the GBRT model is calculated to quantify the rate of change of marginal effect. The two trend curves are superimposed and compared. Based on the curve matching degree and coefficient dispersion, the dominant effect types of each built environment variable are decoupled and distinguished. S63. When the MGWR coefficient exhibits high dispersion under the same variable values, while the SHAP derivative curve remains relatively flat, it is believed that the influence mechanism of this variable on commuting demand is dominated by spatial heterogeneity. When the trend curve of the MGWR coefficient and the SHAP derivative curve show high similarity, it is believed that the apparent spatial heterogeneity of this variable is actually driven by a potential nonlinear relationship. When both of the above characteristics are satisfied at the same time, it is believed that the influence of this variable is a coexistence of spatial heterogeneity and nonlinear effects.
[0053] Preferably, step S7 includes the following specific steps: S71. Clarify the basis of analysis: Based on the parallel modeling and effect decoupling quantification results of the MGWR-GBRT dual model, extract the model output results of each built environment variable from the residential and employment sides, including the ranking of variable importance, the type of dominant effect, the threshold characteristic parameters, and the spatial distribution map of local regression coefficients. S72. Cross-end effect comparative analysis: Compare the importance, dominant effect type, effective action threshold and spatial distribution characteristics of the same built environment variable in the residential density model and the employment density model, and quantify the differences between the two in terms of influence direction and influence intensity. S73. Cross-data cluster driving factor analysis: For three types of commuting data clusters—balanced conventional type, rigid long-distance type, and flexible short-distance type—the core driving parameters of the built environment and the spatial data characteristics of functional anchor points in residential and employment choices are calculated for each data cluster. The differences in response parameters between each data cluster are quantitatively extracted.
[0054] Preferably, step S8 includes the following specific steps: S81. Quantitative Analysis of Differentiated Influence Mechanisms: By comparing the full-dimensional analysis results of the same built environment variable in the residential density model and the employment density model, we accurately calculate and quantify the differential mechanism parameters of the variable's influence on commuters' residential location and employment location, and determine its specific influence direction, influence intensity, dominant effect type, and effective action threshold. S82. Analysis of the basis for planning strategy formulation: Integrate the quantitative analysis results of the asymmetric impact mechanism of residence and employment, the determination results of the dominant effect types of various built environment variables, and combine the spatiotemporal travel characteristics data of different commuting data clusters and the spatial development attribute data of different locations to sort out the core basis data for the formulation of differentiated planning strategies. S83. Generation of Differentiated Planning Strategies by Category: For variables dominated by nonlinear effects, formulate universal urban transportation and land use planning strategies based on threshold feature values; for variables dominated by spatial heterogeneity, formulate spatially oriented planning strategies based on location characteristics; and for different commuting data clusters, formulate inclusive planning strategies that adapt to their housing and employment choices. S84. Implementation of Planning Strategies: Integrate planning strategies categorized by type, group, and location to form a differentiated urban transportation and land use planning strategy system that is adapted to different commuter groups and location characteristics, and highly compatible with the asymmetric impact mechanism of the built environment, providing practical decision support for optimizing urban commuting and improving the work-life balance.
[0055] In addition, to achieve the above objectives, the present invention also provides the following technical solution: a spatial heterogeneity and nonlinear effect decoupling and quantification processing system based on dual-model visualization translation, including a dataset construction and processing module, a commuter data extraction and clustering processing module, a dual-model collaborative computing processing module, an effect decoupling and quantification analysis processing module, and a planning strategy generation and output module.
[0056] Preferably, the dataset construction and processing module is used to collect urban public transportation smart card data, complete data cleaning, travel chain data parsing, processing and integration, and construct a standardized analysis dataset; The commuter data extraction and clustering module is used to process travel data through a two-step clustering method, extract valid commuter data records and their residential and employment spatial data, and complete the clustering and classification of commuter group data based on commuting spatiotemporal characteristic indicators. The dual-model collaborative computing processing module is used to construct the MGWR-GBRT dual-model collaborative computing framework, simultaneously complete the parallel numerical calculation of the residential and employment sides, and extract the spatial heterogeneity and nonlinear effect characteristic data of the built environment impact respectively. The effect decoupling and quantification analysis processing module is used to complete the cross-validation and effect decoupling and quantification analysis of the dual model output data, calculate and output the dominant effect type data of each variable, and quantify the asymmetric characteristic data of the impact of the built environment on residential and employment location selection. The planning strategy generation output module is used to generate differentiated urban transportation and land use planning strategy data that are adapted to different commuting data clusters and different location characteristics, based on the asymmetric impact quantification analysis data and the effect decoupling quantification results.
[0057] In one embodiment, the method of the present invention is put into practice using actual AFC card swiping data of a city's urban rail transit system. This example collects AFC card swiping data of the city's subway and bus systems for the entire month of November 1, 2023, as well as spatial location data of subway stations (264 stations), bus stops (2049 stations), and Points of Interest (POI) data of the study area.
[0058] (1) Construct a public transportation commuting analysis database: In this example, the database information includes travelers' AFC card swipe data for buses and subways, spatial distribution data of bus and subway stations, and data on points of interest in the study area.
[0059] AFC (Automatic Fare Collection) data was collected from subway and regular bus smart cards throughout November 2023 in the study area, totaling 31,578,947 records. The AFC data records basic travel information for public transportation users, including entry time and station, and exit time and station. Examples of some data record fields are shown in Table 1. Table 1. Examples of AFC data structures (partial) ; (2) Data preprocessing: Due to the massive volume of raw data recorded by the AFC system, some redundant data, including anomalies, missing data, or logical conflicts, is inevitably generated during data collection, transmission, and storage. To ensure the accuracy of subsequent spatiotemporal travel pattern mining and clustering modeling for public transportation commuters, rigorous validity and consistency checks must be performed on the raw data. Specific processing steps are as follows: ① Field integrity and format cleaning: Check whether the card number, transaction time, transaction site and latitude and longitude coordinate format of all records are consistent, and remove invalid travel records with missing fields (null values) or garbled formats; at the same time, remove completely duplicate redundant records caused by system failure.
[0060] ② Time logic verification: A normal single trip should satisfy the progressive nature of the time sequence. For logical conflicts caused by system anomalies, erroneous data where the card swipe exit (disembarkation) time is earlier than or equal to the entry (boarding) time is filtered out.
[0061] ③ Travel time threshold filtering: Based on the actual operation of urban public transportation, calculate the actual travel time for each record. Remove extreme abnormal samples with excessively short travel times (<60s, usually due to accidental card swiping or system malfunction) and those exceeding the normal travel limit (>18000s, approximately 5 hours, usually due to the system forcibly settling the transaction due to forgetting to swipe the card to exit the station or overnight data).
[0062] Spatial logic and boundary verification: First, records where the geographical coordinates of the origin and destination are located outside the boundary of the study area in Hangzhou are removed; second, to exclude cross-street behavior for non-travel purposes or cancellation of the ride after entering the station, invalid travel records where the entry and exit stations are the same (i.e., the spatial coordinates of the origin and destination are the same) are removed.
[0063] Abnormally high-frequency card number removal: Considering that some card swipe records may originate from bus / subway employee cards, maintenance and testing cards, or courier delivery personnel, their travel frequency and spatial characteristics differ from those of regular residents. This study sets a threshold for the highest daily travel frequency (e.g., more than 20 rides in a single day), and abnormal card numbers exceeding this threshold, along with their entire daily travel records, are removed.
[0064] 3) Extraction of effective commuter data records and spatial data: A two-layer data clustering and filtering algorithm combining spatiotemporal density clustering (ST-DBSCAN) and spatial density clustering (DBSCAN) was implemented, along with work-residence logical association processing, to extract 15,025,731 records of regular commuting trips from 31,578,947 trip chain records. The specific implementation process is as follows: The first layer (residence departure cluster data detection): Extracts data from each user's travel records, including the longitude, latitude, and departure time of the boarding (or arrival) point, and performs ST-DBSCAN spatiotemporal density clustering. The system automatically selects the cluster with the earliest average departure time and marks it as "Main Residential Departure Cluster Data (H1)" (i.e., potential station residential area data). If the user's clustering result contains only noise points, it is marked as lacking consistent spatiotemporal travel pattern characteristics and removed as a non-commuter data record.
[0065] Second layer (destination clustering and logical matching of work and residence): Complete definition of residence (H): Based solely on the longitude and latitude data of the drop-off point, DBSCAN spatial clustering is performed on all user drop-off records to generate multiple high-frequency, stable destination clusters. The system calculates the distance between the geographic centroid of these destination clusters and H1, and marks destination clusters with a distance less than a given threshold as "residence departure cluster data (H2)" (i.e., drop-off station residential area data). H1 and H2 together spatially define the user's "potential residence data (H)".
[0066] Workplace Data (W) Extraction and Commuter Data Screening: The system extracts the set of destination data (denoted as D) from H1 and the set of origin data (denoted as O) from H2 as candidate workplace data points. High-intensity DBSCAN spatial clustering is then performed on these two candidate data point sets, and the largest and densest cluster is marked as the user's "primary workplace cluster data (W)". Finally, the system retains only user data records successfully associated with valid workplace cluster data (W) as regular commuter data records and extracts their corresponding dual-end spatial coordinate data for subsequent modeling.
[0067] (4) Spatiotemporal pattern clustering of commuter groups: For the extracted commuter data records, multidimensional spatiotemporal travel feature data is further extracted and unsupervised clustering analysis is performed. The specific processing flow is as follows: Multidimensional spatiotemporal feature quantification extraction: Traversing the travel trajectory of each effective commuter, six core quantitative indicators are automatically calculated from both temporal and spatial dimensions. Among them, the temporal dimension extracts: standard deviation of departure time (measures the flexibility of travel time), proportion of peak-hour travel (measures the rigid dependence on morning and evening tidal traffic), and average travel time; the spatial dimension extracts: average commuting distance (measures cross-regional mobility), symmetrical travel rate (calculates the proportion of complete "home-workplace-home" closed-loop trajectory to the total number of active days), and number of active stations (assess the spatial network complexity and route fixation).
[0068] K-Means Automatic Clustering and Parameter Optimization: The above six feature data are standardized and then fed into the K-Means clustering algorithm. The system iterates through different K values (from K=1 to K=10) and calculates the sum of squared errors (SSE) for each clustering. Based on the elbow rule curve plotted by the system, the system automatically detects the location of significant "inflection point" data points where SSE decays, thus calculating the optimal number of clusters as 3.
[0069] After the algorithm converges, the system automatically assigns semantic labels to the three clusters based on the centroid of their data distribution across various metrics and outputs the group percentages. Rigid long-distance data cluster (accounting for 48%): System calculations show that this data cluster has the highest average commuting distance, peak-hour travel ratio (median close to 1.0), and symmetrical travel rate (close to 0.8). Its spatial mobility data exhibits an extremely rigid "point-to-point" distribution characteristic, and data records show a high dependence on the rail transit system for cross-regional commuting.
[0070] Balanced and regular data cluster (39%): The values of various indicators in this data cluster are in the middle range. Its time distribution data shows a gentle bimodal pattern and significant delayed travel characteristics; while maintaining a high work-residence round-trip rate, it involves a large number of active site data, indicating that its travel data has a certain degree of flexibility and undertakes some non-commuting activity records.
[0071] The flexible short-distance data cluster (13%) has the shortest average commuting distance and one-way travel time. Its departure time standard deviation is the largest among the three groups, exhibiting extremely high time flexibility (evening rush hour starts significantly earlier). Regarding transportation preference data, this data cluster is the only one categorized by the system as primarily using regular public transportation (usage rate >50%) as its commuting mode.
[0072] (5) Construction of spatial units adapted to urban block scale and statistical robustness pretest: To establish a unified and reliable benchmark for measuring the spatial distribution of work and residence, this embodiment adopts an engineering grid standard that conforms to urban planning scales and implements strict statistical compliance control over spatial variables. The specific implementation steps are as follows: Spatial analysis unit standardization and intensity conversion: Based on the public transportation connection scale of the "15-minute living circle" in China's megacities and the typical checkerboard rectangular street block geometry, the system directly generates a 1000m×1000m standard rectangular raster layer covering the entire study area in the background. The effective work-residence coordinate points extracted in step S2 are spatially overlaid and correlated with this raster layer. The residential generation intensity and employment attraction intensity contained in each raster unit are statistically calculated, thereby achieving the standardized conversion from individual discrete commuting travel chains to spatially continuous raster intensity data.
[0073] Multi-source built environment feature measurement and spatial coupling: Using a defined 1000m rectangular grid as the unified measurement unit, the system synchronously extracts and calculates the density of multi-source POIs (such as commercial, educational, and medical facilities), land use mix entropy, and the shortest spatial path distance from each grid center to the urban core area (CBD) and the nearest rail transit station. After performing Min-Max normalization, the extracted dependent variable (dual-end commuting intensity) and independent variable (built environment features) are coupled in a high-dimensional spatial dimension at the same grid scale to construct a complete modeling variable dataset.
[0074] To ensure the reliability of subsequent dual-model modeling, the system performs rigorous statistical pre-tests on the coupled raster variable set. First, the system automatically performs a variance inflation factor (VIF) test to detect and remove redundant independent variable data with VIF values exceeding a preset safety threshold (e.g., VIF > 10), reducing data interference from multicollinearity on the model's explanatory power. Simultaneously, the system calculates the global Moran's I for each built-up environment characteristic variable to verify whether each variable possesses significant spatial autocorrelation characteristics within a 1000m raster cell. These tests thoroughly establish the spatial regression applicability of the analytical data in the MGWR model, providing a highly statistically significant data foundation for subsequent asymmetric mechanism decoupling.
[0075] (6) Construction of the modeling variable system: The total number of potential residential locations and potential workplaces for the three commuter groups within each grid were calculated separately. After area standardization, the "commuter generation density" (residential end) and "commuter attraction density" (work end) were obtained as dependent variables. Simultaneously, 11 built environment explanatory variables were extracted (such as commercial density, distance to city center, and distance to the nearest subway station). The variance inflation factor (VIF) test showed that the VIF values of all variables were less than 10, ruling out multicollinearity. The Moran's I test showed that all variables were significantly greater than 0, verifying positive spatial correlation. The definitions and descriptive statistics of the built environment explanatory variables are shown in Table 2, and the tests for multicollinearity and spatial autocorrelation are shown in Table 3.
[0076] Table 2. Definitions and Descriptive Statistics of Explanatory Variables for Built Environment ; Table 3. Tests for Multicollinearity and Spatial Autocorrelation ; (7) Construction of the MGWR-GBRT dual-model analysis framework and effect decoupling and quantitative analysis: The aforementioned two-ended dependent and explanatory variables were then fitted into the Multiscale Geographically Weighted Regression (MGWR) model and the Gradient Boosting Regression Tree (GBRT) model, respectively. The following effect decoupling steps were then performed: Extract the SHAP partial dependence curve from the GBRT model output and calculate the slope (i.e., first derivative) of each interval to reflect the rate of change of nonlinear marginal effects.
[0077] The local regression coefficients of each spatial unit output by the MGWR model are extracted. A coefficient-variable scatter plot is drawn with the built-up environment variable value on the x-axis and the local regression coefficients on the y-axis. A local coefficient evolution trend line is generated through polynomial fitting. In this scatter plot, the vertical dispersion of scatter points within the same variable value interval intuitively maps the strength of the spatial heterogeneity of the variable's influence on the dependent variable.
[0078] Visualization and Mutual Translation and Mechanism Decoupling Judgment Rules: The system reconstructs and superimposes the above SHAP first derivative curve, MGWR coefficient scatter points and trend lines in the same dual Y-axis coordinate system, and judges the mechanism decoupling by comparing the dispersion of the scatter points and the trend of the SHAP derivative curve.
[0079] Spatial heterogeneity (location dependence) dominates: If the first derivative curve of SHAP approaches horizontal (the slope is close to 0, that is, the nonlinear scale marginal effect is very weak) within a specific variable range, but the MGWR scatter plot shows a sharp high dispersion in the vertical direction (indicating that the utility of different spatial locations varies greatly under the same density of facilities), then it is believed that the effect of this variable is dominated by the real spatial heterogeneity of the specific location. Nonlinear effects (scale dependence) dominate: If the first derivative curve of SHAP shows significant fluctuations or a steep monotonic slope, and the trend line of the MGWR coefficient is consistent with it, and the MGWR scatter points closely follow the trend line (with very small vertical dispersion), then it is believed that the spatial distribution of the variable is actually driven by a nonlinear scale / density threshold effect. Intertwined effect: If the SHAP first derivative curve has significant fluctuations, and the MGWR scatter points simultaneously exhibit highly vertical discrete characteristics within a specific variable value range, then the influence mechanism of the variable is considered to be dominated by the dual intertwining of nonlinear scale constraints and real spatial heterogeneity.
[0080] (8) Analysis of the two-end asymmetric mechanism: Based on the system's output dual-model visualization decoupling diagram, and by comparing the dispersion of the SHAP first derivative curve and the MGWR local coefficient scatter plots, three typical data characteristics of the asymmetric impact of work-residence relationship are quantified: The intertwined nature of rail transit dependence and group differentiation: such as Figures 8 to 13As shown, the distance to the nearest subway station exhibits a significant nonlinear effect at both the residential and work ends (manifested as a phased change in the red line in the decoupling diagram, generally rising first and then falling within the range of 0-5000 meters, with a clear turning point around approximately 3000 meters). However, at the residential end, the MGWR coefficient scatter points exhibit extremely exaggerated vertical dispersion within the threshold range, and this high variance characteristic is particularly pronounced in "Cluster 1 (rigid long-distance group)". This indicates that the attractiveness of rail transit for a residential location, after satisfying the physical connection distance, is essentially strongly intertwined and dominated by the spatial heterogeneity of specific stations (such as locational added value like school district benefits and commercial area empowerment); while at the work end, the scatter point dispersion converges rapidly, indicating that the dependence of the workplace on the subway is more inclined towards the pure distance constraint of being able to reach the station immediately upon exiting.
[0081] Purely nonlinear, layered, asymmetric gravity in city centers: such as Figures 2 to 7 As shown, the distance to the city center (CBD) exhibits a smooth, monotonically decreasing trend along the red line at both the work and residence ends, with the blue MGWR scatter plots showing extreme convergence (closely following the trend line, with minimal variance). This confirms that the variable is dominated by a purely nonlinear scale dependence (physical distance) and is largely unaffected by local locational factors. Its asymmetry is reflected in the difference in the shape of the gravitational radius: high commuting demand at the work end is tightly clustered in the core area and drops sharply (employment polarization), while high demand at the residential end spreads more gently to the outer urban areas (suburbanization).
[0082] Pure spatial heterogeneity of local facility density in job-housing differentiation: such as Figures 14 to 19 As shown, taking the commercial density at the residential end and the educational density at the workplace as examples, the decoupling diagram shows that the first derivative of SHAP (red line) globally approaches zero, indicating extremely weak nonlinear scale marginal effects. However, the MGWR scatter plot exhibits a dramatic divergence covering both positive and negative aspects. This suggests that such local living / public facilities are dominated by extremely strong pure spatial heterogeneity. Commercial facilities of the same density provide a strong convenience pull (positive coefficient) in some grids at the residential end, while in other grids, noise and congestion transform into a push effect of NIMBY (negative coefficient). Similarly, educational facilities at the workplace exhibit completely fragmented employment attractiveness depending on the specific school type (e.g., primary and secondary schools vs. university towns).
[0083] (9) Implementation and output of planning strategies: Based on the above decoupling conclusions, the system outputs targeted and refined urban spatial intervention strategies: A dual-end asymmetric TOD development strategy to address the intertwining effect: Given the extremely low tolerance for connectivity on the employment side, a strong core agglomeration should be strictly enforced around hub stations, rigidly limiting high-density commercial and business office land within a 500-800 meter core walking radius of the station. This should be achieved by creating high-density underground and covered walkways to prevent commuter attrition due to inconvenient last-mile connections. Given that the tolerance for connectivity on the residential side is as high as 1.5-2 kilometers and is dominated by location value, priority should be given to high-frequency starting stations for rigid long-distance commuters. By optimizing the micro-circulation of connections (such as community shuttle buses and dedicated non-motorized vehicle lanes), the station's coverage area can be substantially expanded. Furthermore, affordable rental housing should be strategically allocated around high-value nodes with significantly positive local MGWR (Mobility Management Response) coefficients.
[0084] For the gradient concentric layer dispersal strategy dominated by nonlinearity: Based on the concentric layer characteristics of CBD attractiveness being dominated by physical distance (nonlinear decay), it is recommended to forcibly allocate secondary commercial and office land in the critical concentric layer where the main center's gravity decays (such as the 15-20 km outer nodes where the slope of the red line in the decoupling diagram gradually decreases), and to precisely plan the outer sub-centers for employment, thereby intercepting the rigid long-distance tidal commuter flow pouring into the central urban area; at the same time, strictly control the approval of new pure residential land in the inner concentric layer of the CBD with strong gravity.
[0085] A location-specific micro-renewal strategy is adopted for areas dominated by spatial heterogeneity: A unified assessment of facility allocation across the entire region is completely abandoned. For heterogeneous variables such as residential and commercial development, the system identifies grid areas with negative MGWR (Not In My Backyard) coefficients (significant NIMBY effect), implementing a strategy of reducing quantity and increasing quality, strictly limiting the addition of large commercial complexes, and instead supplementing street-level pedestrian spaces and green spaces. For grids with positive local pull coefficients, land use controls are relaxed, encouraging the exploration of vertical space mixed development models of "ground-floor shops + upper-level residences / offices," using micro-renewal methods to promote local job-housing balance.
Claims
1. A method for decoupling spatial heterogeneity and nonlinear effects based on dual-model visual translation, characterized in that, include: S1. Obtain the spatiotemporal trajectory data of the moving object, extract the anonymous identifier, spatial coordinates of trajectory points and corresponding timestamps from it, and construct a standardized vector dataset; S2. Construct a two-layer data filtering algorithm based on spatiotemporal density clustering and spatial density clustering to extract spatiotemporal data records of regularly moving objects from a standardized vector dataset, and output the starting and ending spatial centroids of the regularly moving objects. S3. Based on the centroid of the starting space and the centroid of the ending space, extract the multidimensional spatiotemporal behavioral features of each regular moving object, perform clustering operations on the feature data of the regular moving objects, and generate multiple movement pattern data clusters. S4. Using a rectangular grid of a preset size as a unified spatial analysis unit, spatially associate the starting and ending spatial centroids with the rectangular grid, statistically analyze the distribution intensity of target events and the feature vector of spatial covariates within each grid unit, and construct a modeling dataset. S5. Based on the modeling dataset, a multi-scale geographically weighted regression model and a gradient boosting regression tree model are independently constructed for the starting and ending ends, respectively, forming a dual-model parallel modeling framework. The output includes the local regression coefficients, nonlinear dependencies, threshold feature parameters, variable importance ranking, and dependency curves output by the additivity interpretation method for each spatial covariate. S6. Calculate the first derivative of the dependency curve to obtain the marginal effect change data of each spatial covariate. Perform cross-validation and trend comparison between the local regression coefficients and the marginal effect change data, and output the dominant effect type data of each spatial covariate. S7. Based on the dominant effect type data, extract the variable importance ranking, dominant effect type, threshold feature parameters and local regression coefficient spatial distribution matrix of each spatial covariate at the start and end points, quantify the differential driving parameters of each mobile mode data cluster at the start and end points, output the response difference data between each mobile mode data cluster, and generate a feature vector set for spatiotemporal data effect separation. S8. Based on the feature vector set and dominant effect type data, generate a structured spatial configuration intervention parameter set.
2. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 1, characterized in that, Step S2 specifically includes: S21. Extract the first component of the spatial coordinates of the starting point in the individual trajectory chain, the second component of the spatial coordinates, and the starting time to form a three-dimensional feature space. Execute the spatiotemporal density clustering algorithm and mark the cluster with the earliest average starting time as the starting cluster. S22. Based on the spatial coordinates of the termination point in the individual trajectory chain, perform spatial density clustering to generate termination clusters. Mark the termination clusters whose centroid distance from the starting cluster is less than a spatial threshold as termination arrival areas. The starting clusters and termination arrival areas jointly define the potential starting spatial range. S23. Extract the set of trajectory endpoints from the starting cluster and the set of trajectory starting points to the destination area of the termination point as candidate points, perform spatial density clustering, mark the largest cluster as the main termination cluster, and retain the objects that are successfully associated with the main termination cluster as regular moving objects.
3. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 1, characterized in that, Step S3 specifically includes: The system extracts the standard deviation of start time, the proportion of activities during peak hours, and the average activity time from the time dimension, and extracts the average moving distance, the symmetrical round-trip rate, and the number of active anchor points from the spatial dimension, thus forming a six-dimensional quantitative feature system. After standardizing the six-dimensional quantitative feature system, the K-Means clustering algorithm is input. The number of clusters is determined by the elbow rule. Based on the data distribution centroid of each cluster in each dimension of the six-dimensional quantitative feature system, the regular moving objects are marked as the first moving pattern data cluster, the second moving pattern data cluster, and the third moving pattern data cluster.
4. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 1, characterized in that, Step S4 specifically includes: S41. Generate a rectangular grid layer covering the entire area based on the geographical boundary of the study area, and spatially overlay and associate the starting spatial centroid and the ending spatial centroid with the rectangular grid layer, and statistically analyze the generation intensity of the starting end event and the attraction intensity of the ending end event in each grid cell. S42. Using the rectangular grid as the measurement unit, extract the distribution density of interest points, the entropy value of spatial element mixing degree, the shortest spatial distance from the grid center to the core of the region, and the shortest spatial distance from the grid center to the nearest transportation hub, as spatial covariate feature vectors. S43. Normalize the intensity of the starting event, the attraction intensity of the ending event, and the feature vector of the spatial covariate to complete the spatial attribute coupling on the same grid scale and construct the modeling variable dataset. S44. Perform a variance inflation factor test on the modeling variable dataset, detect and remove independent variable data whose variance inflation factor values are greater than a preset threshold, and calculate the global Moran index of each spatial covariate to verify the spatial autocorrelation characteristics. S45. The intensity of the starting event and the intensity of the ending event in each grid cell are standardized according to the area of the grid cell to obtain the density of the starting event and the density of the ending event in each grid cell.
5. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 1, characterized in that, Step S5 specifically includes: Using the starting event density of each mobile mode data cluster within each grid cell as the starting dependent variable, the ending event density of each mobile mode data cluster within each grid cell as the ending dependent variable, and the spatial covariate feature vector as the independent variable, based on the modeling dataset, the fitting training of the multi-scale geographic weighted regression model and the gradient boosting regression tree model is carried out independently for the starting dataset and the ending dataset respectively. By using a multi-scale geographic weighted regression model, the local regression coefficients and spatial distribution matrix of each spatial covariate are output, and the spatial heterogeneity characteristics of the influence of spatial covariates on the spatiotemporal patterns of moving objects are extracted. Gradient boosting regression tree model is used to extract the nonlinear influence features of spatial covariates on the behavior of moving objects. The model output is interpreted by combining the additive interpretation method, and the nonlinear dependencies of each spatial covariate, threshold feature parameters, variable importance ranking, and dependency curves output by the additive interpretation method are obtained.
6. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 1, characterized in that, Step S6 specifically includes: S61. Convert the local regression coefficients of the multi-scale geographical weighted regression model into a scatter plot with spatial covariate values on the x-axis and local regression coefficients on the y-axis, and fit a polynomial trend curve. S62. Calculate the first derivative of the dependency curve output by the gradient boosting regression tree model combined with the additivity interpretation method to obtain the marginal effect change data of each spatial covariate. S63. Superimpose and compare the polynomial trend curve with the curve of the marginal effect change data, and output the dominant effect type data of each spatial covariate based on the curve matching degree and coefficient dispersion. When the multi-scale geographically weighted regression coefficients exhibit high dispersion under the same variable values, while the curve of the marginal effect change data remains relatively flat, the dominant effect type of the output spatial covariate is spatial heterogeneity-dominated; when the polynomial trend curve and the curve of the marginal effect change data show high similarity, the dominant effect type of the output spatial covariate is nonlinear effect-dominated; when both of the above characteristics are satisfied, the dominant effect type of the output spatial covariate is a coexistence of spatial heterogeneity and nonlinear effects.
7. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 1, characterized in that, In step S7, the importance ranking, dominant effect type, threshold feature parameters, and spatial distribution matrix of local regression coefficients of each spatial covariate at the starting and ending ends are extracted, specifically including: The variable importance ranking, the dominant effect type, the threshold feature parameter, and the spatial distribution matrix of the local regression coefficients are all derived from the output results of the dual-model parallel modeling in step S5 and the dominant effect type data in step S6. By comparing the importance ranking, dominant effect type, effective action threshold parameter, and spatial distribution matrix of local regression coefficients of the same spatial covariate at the start and end points, the differential characteristic data of the direction and intensity of the influence of the variable between the start and end points are quantified, which serve as the core data content of the feature vector set for effect separation.
8. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 7, characterized in that, Step S8 specifically includes: For spatial covariates whose output is dominated by nonlinear effects, extract and output their effective threshold parameters; For spatial covariates whose output is dominated by spatial heterogeneity, locate and output a set of raster spatial coordinates where local regression coefficients show a high positive effect; Based on the effective action threshold parameter and the grid spatial coordinate set, a spatial configuration intervention map is generated as a quantitative data reference for the structured spatial configuration intervention parameter set.
9. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 8, characterized in that, In step S8, generating a structured set of spatial configuration intervention parameters specifically includes: For spatial covariates whose output is dominated by nonlinear effects, a first type of data labeling parameters are generated based on the effective action threshold parameter. For spatial covariates whose output is dominated by spatial heterogeneity, a second type of data labeling parameters are generated based on the set of raster spatial coordinates that exhibit high positive effects from the local regression coefficients. Based on the response parameter difference data between each mobile mode data cluster extracted in step S7, and combined with the differential driving factor data of each mobile mode data cluster obtained by quantification in step S7, a third type of data annotation parameters are generated. The first type of data annotation parameters, the second type of data annotation parameters, and the third type of data annotation parameters are integrated to form a structured set of spatial configuration intervention parameters.
10. The spatial heterogeneity and nonlinear effects decoupling method based on dual-model visual translation according to claim 1, characterized in that, In step S4, before constructing the modeling dataset, a variance inflation factor test is performed on the spatial covariate feature vectors to detect and remove independent variable data with variance inflation factor values greater than a preset threshold, and the global Moran index of each spatial covariate is calculated to verify the spatial autocorrelation characteristics.
