Public transport node performance analysis method based on multivariate data and interpretable machine learning

Abstract

The invention discloses a public transport node performance analysis method based on multivariate data and interpretable machine learning, relates to the field of machine learning, and aims to solve the problem of incomplete performance analysis caused by single data source, insufficient spatial analysis granularity and inaccurate matching between individual travel behaviors and public transport node microscopic features in the prior art. Comprising the following steps: S1, collecting and preprocessing multi-source data; s2, generating a main trip mode label through trip chain codes; s3, constructing a bus node space based on 150-meter buffer area superposition and a line density threshold value; s4, counting a section POI proportion, and generating a node feature vector by fusing a location index; s5, associating a travel starting point with a node space through line matching and a distance threshold value, and fusing individual and node features to construct a data set; and S6, quantizing the nonlinear influence of the node function layout on the travel selection by adopting a hierarchical LightGBM model in combination with SHAP explanation. The method has a good application prospect in the field of urban traffic planning.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a method for analyzing the efficiency of public transportation nodes based on multivariate data and interpretable machine learning. Background Art

[0002] Traditional travel decision-making research often relies on questionnaire survey data and macro-level built environment indicators (such as the 5D index) for analysis, which tends to overlook fine-grained differences in individual travel behavior. In the prior art, although there are patents that use interpretable machine learning to analyze the impact of the built environment, such as CN 117874709 A, a method for analyzing the multi-scale impact of the built environment on residents' travel based on interpretable machine learning, and CN 116452016 A, a method and system for analyzing urban traffic structure based on the XGBoost model, these methods lack a single data source, a fixed spatial analysis scale, and a lack of micro-description of the specific functional layout around bus and subway nodes. Furthermore, they do not fully utilize the combined matching of individual travel data and traffic node route information, limiting the accuracy and comprehensiveness of the analysis. Summary of the Invention

[0003] To address the incomplete performance analysis issues of existing technologies due to a single data source, insufficient spatial analysis granularity, and inaccurate matching of individual travel behaviors with the micro-characteristics of public transportation nodes, the present invention provides the following solution: a public transportation node performance analysis method based on multivariate data and interpretable machine learning, comprising the following steps:

[0004] S1. Obtain resident travel survey data, public transportation station and route data, geographic information data, residential and job population data, and point of interest data, and perform missing value processing and coordinate system analysis to obtain a spatially normalized multi-source dataset;

[0005] S2. Calculating travel duration and travel distance based on the spatially normalized multi-source dataset, encoding the entire chain of transportation modes into digital primary travel modes based on preset rules, and obtaining individual travel records;

[0006] S3. Based on the spatially normalized multi-source dataset, construct a public transportation node space by generating a 150-meter buffer zone and merging adjacent areas with a number of routes greater than or equal to 10;

[0007] S4. Based on the public transportation node space and the spatially normalized multi-source dataset, the number and proportion of the four major categories of POIs in the node are statistically analyzed through spatial overlay analysis to obtain a comprehensive feature vector for each node;

[0008] S5. Perform feature matching based on the individual travel records, the public transportation node space, and the comprehensive feature vector of each node to construct a feature matching dataset;

[0009] S6. Based on the feature matching dataset, the key features and their nonlinear effects that affect public transportation travel decisions are identified through hierarchical LightGBM model training and SHAP interpretability analysis.

[0010] Furthermore, in S1, the resident travel survey data includes: family number, personal number, personal travel number, starting point longitude and latitude, destination longitude and latitude, departure time, arrival time, transportation mode for the entire travel chain, age, gender, education level, personal annual income, family annual income, work industry, and number of family vehicles;

[0011] The public transportation station and line data include: bus station name, bus station longitude and latitude, bus line numbers of all bus routes that stop at the bus station, subway station name, subway station longitude and latitude, and subway line number where the subway station is located;

[0012] The geographic information data includes: projection coordinate system information, urban road network information and urban boundary information;

[0013] Residential and job population data include: urban population and job density with a grid accuracy of 250m*250m or urban population and job density with a grid accuracy of 80m*80m;

[0014] The POI data includes: POI name, major category, minor category and latitude and longitude information.

[0015] Furthermore, in S2, the travel duration is calculated by:

[0016] T journey =T arrival -T departure

[0017] Get, where T journey is the total travel time, T arrival is the arrival time of the trip record, T departure The departure time recorded for the trip;

[0018] The travel distance is given by:

[0019]

[0020] Get, where d is the travel distance, x end is the horizontal coordinate of the plane projection of the travel destination, x start is the horizontal coordinate of the plane projection of the starting point of the trip, y end is the plane projection ordinate of the travel destination, y start It is the plane projection vertical coordinate of the starting point of the trip.

[0021] Furthermore, in S2, the individual travel record includes: personal economic and social characteristics, travel information and travel mode.

[0022] Furthermore, S3 is specifically:

[0023] S31. Generate a buffer zone with a radius of 150 meters with each bus stop as the center. The buffer zone contains all bus route information of the bus stop.

[0024] Overlay buffers and calculate the number of bus routes passing through each overlay area. The number of bus routes is calculated by:

[0025]

[0026] Get, where count spatial_join is the number of bus routes in the superimposed buffer zone, count i is the number of lines with a single station participating in the superposition;

[0027] S32, if count spatial_join If ≥10, all the bus stops associated with it constitute a transportation node, and the combined area of the 150-meter buffer zone formed by all the stops is regarded as the transportation node space.

[0028] Furthermore, S4 is specifically:

[0029] S41: spatially superimpose the public transportation node space with the POI data, extract POI information within the public transportation node space, and screen POIs in four categories: catering, shopping, life services, and leisure and entertainment. Specifically, the following steps are performed:

[0030]

[0031] Implementation, where category is catering or shopping or life services or leisure and entertainment, count category is the number of POIs of a certain category, node is all the stations in the traffic node space, POI k is the kth POI point, I is the indicator function;

[0032] S42. Calculate the proportion of each category in the node space by:

[0033]

[0034] Implementation, where ratio category is the proportion of a certain type of POI to all POIs in the node, count total is the total number of all four categories of POIs in the node;

[0035] S43. Extract indicators of the area where the location data overlaps with the node space, and construct a comprehensive feature vector for each node.

[0036] Furthermore, S5 is specifically:

[0037] S51. Determine whether the individual travel record is within the walking influence range of any public transportation node, specifically by:

[0038]

[0039] Implementation, where origin is the latitude and longitude coordinates of the starting point of the individual travel record, node i is the center coordinate index of the i-th traffic node space, d is the plane Euclidean distance calculation formula, Indicates the minimum distance from the starting point to all public transportation nodes. origin-pos The value is 1 or 0;

[0040] S52: First, check each individual travel record to see whether its main travel mode is bus or subway. If not, do not perform subsequent matching processing;

[0041] S52: Filter out candidate nodes from all public transportation nodes that have an intersection with the route number involved in the individual travel record, specifically by:

[0042]

[0043] Implementation, where record.lines is the set of transportation line numbers used in individual travel records, node.lines is the set of line numbers contained in the public transportation node space, and CandidateNodes origin A candidate node that intersects with the route number involved in the individual travel record;

[0044] S54. For the candidate nodes that intersect with the route numbers involved in the individual travel record, calculate the distance between their starting points and the centers of the nodes, specifically by:

[0045] d origin (node)=distance(record.origin,node.center)

[0046] Implementation, where record.origin is the latitude and longitude of the starting point in the same geographical location match, node.center is the latitude and longitude coordinates of the center point of the candidate node, d origin (node) is the distance between its starting point and the center of each node;

[0047] S55. Select the node with the smallest distance by:

[0048]

[0049] Implementation, where is the node with the smallest distance;

[0050] S56, distance judgment, if If the distance is less than 1000 meters, the travel record is considered to match the public transportation node space and the record is:

[0051] indicator origin =1

[0052]

[0053] Otherwise, both are set to 0, and origin_node_id is the unique identifier of the public transportation node matched to the travel origin:

[0054] S57. Integrate the individual travel record and the comprehensive feature vector of each node using the unique identifier of the public transportation node matched to the travel starting point to form a multi-source fused feature matching dataset.

[0055] Furthermore, S6 is specifically:

[0056] S61, use the three-layer LightGBM model;

[0057] At the first level, a binary classification model is used to distinguish between motorized travel and non-motorized travel;

[0058] At the second level, in motorized travel, the binary classification model further distinguishes between public transportation and non-public transportation; in non-motorized travel, the binary classification model further distinguishes between walking and non-motorized travel;

[0059] The third layer uses a binary classification model to distinguish between buses and subways for public transportation and taxis and self-driving cars for non-public transportation.

[0060] S62, directly use the classification accuracy as the training objective function;

[0061] Accuracy by:

[0062]

[0063] get;

[0064] S63. Using the built-in evaluation indicators of LightGBM, set the metric parameter to 'accuracy' to calculate the classification accuracy of the current model after each iteration;

[0065] Cross-validation is used to select the hyperparameter combination that maximizes Accuracy, including learning rate, tree depth, and minimum number of samples;

[0066] The model training of each layer uses accuracy as the main evaluation indicator, and the model finally selected can achieve the highest classification accuracy in its respective classification task;

[0067] S64. Calculate the SHAP value of each input feature using the SHAP method, where the input features include individual travel characteristics, node spatial characteristics, and user profile information;

[0068] S65. Calculate the SHAP interaction value to reveal the positive and negative impact of each functional layout feature under different user profiles and different travel purposes;

[0069] S66. Based on the results of the global SHAP analysis, rationally adjust the scale and spatial layout of public transportation node spaces to adapt them to the needs of the area and achieve intensive and balanced allocation of various functions within the public transportation node area.

[0070] Compared with the existing technology, the beneficial effects of the present invention are: the present invention judges the node space around the user by distance, determines the public transportation nodes that the user actually enters and exits by route matching and distance judgment, captures the impact of the spatial characteristics of public transportation nodes on individual travel behavior, and thus provides more accurate guidance for public transportation node optimization.

[0071] The dataset construction process was refined, with in-depth preprocessing and correlation matching, shifting from macro-statistics to micro-analysis of individual behaviors. Specifically, the dataset: a. Extracts the primary mode of transportation for a single trip (e.g., subway and walking connections, simplified to subway trips); b. Accurately matches each passenger's trip chain to specific public transportation nodes and routes, rather than simply assuming passengers enter and exit from the nearest station.

[0072] The analysis is more refined and targeted. Based on transportation industry knowledge and transportation mode classification, a three-layer LightGBM model was used to construct a three-level judgment: "whether to travel by motorized vehicle, whether to travel by public transportation, and the specific transportation mode." This allows for a detailed analysis of the characteristics of residents' decision-making process at each level. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a flowchart of a public transportation node efficiency analysis method based on multivariate data and interpretable machine learning. DETAILED DESCRIPTION

[0074] Example 1, combined Figure 1 This embodiment describes a method for analyzing the efficiency of public transportation nodes based on multivariate data and interpretable machine learning, comprising the following steps:

[0075] S1. Obtain resident travel survey data, public transportation station and route data, geographic information data, residential and job population data, and point of interest data, and perform missing value processing and coordinate system analysis to obtain a spatially normalized multi-source dataset;

[0076] S2. Calculating travel duration and travel distance based on the spatially normalized multi-source dataset, encoding the entire chain of transportation modes into digital primary travel modes based on preset rules, and obtaining individual travel records;

[0077] S3. Based on the spatially normalized multi-source dataset, construct a public transportation node space by generating a 150-meter buffer zone and merging adjacent areas with a number of routes greater than or equal to 10;

[0078] S4. Based on the public transportation node space and the spatially normalized multi-source dataset, the number and proportion of the four major categories of POIs in the node are statistically analyzed through spatial overlay analysis to obtain a comprehensive feature vector for each node;

[0079] S5. Perform feature matching based on the individual travel records, the public transportation node space, and the comprehensive feature vector of each node to construct a feature matching dataset;

[0080] S6. Based on the feature matching dataset, the key features and their nonlinear effects that affect public transportation travel decisions are identified through hierarchical LightGBM model training and SHAP interpretability analysis.

[0081] Specifically, the purpose of this invention is to propose a spatial efficiency analysis method for urban public transportation nodes based on individual travel utility, focusing on the layout optimization of spatial transportation facilities and service facilities in urban public transportation nodes, and exploring the key factors and marginal utility that affect passengers' travel choices. Using resident travel survey data, urban public transportation station distribution data, and urban point of interest distribution data, and using the SHAP-packaged LightGBM model, we study the patterns of urban travel mode selection, focusing on the configuration scale and layout of public transportation nodes, and providing important decision support for optimization design.

[0082] Furthermore, in S1, the resident travel survey data includes: family number, personal number, personal travel number, starting point longitude and latitude, destination longitude and latitude, departure time, arrival time, transportation mode for the entire travel chain, age, gender, education level, personal annual income, family annual income, work industry, and number of family vehicles;

[0083] The public transportation station and line data include: bus station name, bus station longitude and latitude, bus line numbers of all bus routes that stop at the bus station, subway station name, subway station longitude and latitude, and subway line number where the subway station is located;

[0084] The geographic information data includes: projection coordinate system information, urban road network information and urban boundary information;

[0085] Residential and job population data include: urban population and job density with a grid accuracy of 250m*250m or urban population and job density with a grid accuracy of 80m*80m;

[0086] The POI data includes: POI name, major category, minor category and latitude and longitude information.

[0087] Specifically, S1 aims to integrate multi-source heterogeneous data, eliminate spatiotemporal benchmark differences and format conflicts between data sources, unify spatial analysis scales and coordinate systems, and build a high-precision, multi-dimensional, and spatiotemporally consistent feature engineering foundation, thereby supporting subsequent traffic node spatial feature extraction, individual-node matching, and interpretable machine learning modeling, ensuring accurate correlation analysis from micro-individual decision-making to macro-node functional layout.

[0088] Furthermore, in S2, the travel duration is calculated by:

[0089] T journey =T arrival -T departure

[0090] Get, where T journey is the total travel time, T arrival is the arrival time of the trip record, T departure The departure time recorded for the trip;

[0091] The travel distance is given by:

[0092]

[0093] Get, where d is the travel distance, x end is the horizontal coordinate of the plane projection of the travel destination, x start is the horizontal coordinate of the plane projection of the starting point of the trip, y end is the plane projection ordinate of the travel destination, y start It is the plane projection vertical coordinate of the starting point of the trip.

[0094] Furthermore, in S2, the individual travel record includes: personal economic and social characteristics, travel information and travel mode.

[0095] Specifically, for each travel record, the following rules are used to map it to a digital code based on the content of the entire transportation chain:

[0096] (a) If "subway" or "bus" is included:

[0097] If both exist, set the code to 3 (for "subway and bus")

[0098] Otherwise, record it separately as "subway" (code 1) or "bus" (code 2)

[0099] (b) Otherwise, if it includes “taxi” or “online car-hailing”, the code is 4 (“ride a taxi”).

[0100] (c) Otherwise, if “private car” is included, the code is 5

[0101] (d) Otherwise, if it contains “motorcycle” or “bicycle”, the code is 6

[0102] (e) Otherwise, if “walking” is included, the code is 7

[0103] (f) Otherwise the default code is 8 ("High-speed rail and other")

[0104] Pseudocode example:

[0105]

[0106]

[0107] Furthermore, S3 is specifically:

[0108] S31. Generate a buffer zone with a radius of 150 meters with each bus stop as the center. The buffer zone contains all bus route information of the bus stop.

[0109] Overlay buffers and calculate the number of bus routes passing through each overlay area. The number of bus routes is calculated by:

[0110]

[0111] Get, where count spatial_join is the number of bus routes in the superimposed buffer zone, count i is the number of lines with a single station participating in the superposition;

[0112] S32, if count spatial_join If ≥10, all the bus stops associated with it constitute a transportation node, and the combined area of the 150-meter buffer zone formed by all the stops is regarded as the transportation node space.

[0113] Furthermore, S4 is specifically:

[0114] S41: spatially superimpose the public transportation node space with the POI data, extract POI information within the public transportation node space, and screen POIs in four categories: catering, shopping, life services, and leisure and entertainment. Specifically, the following steps are performed:

[0115]

[0116] Implementation, where category is catering or shopping or life services or leisure and entertainment, count category is the number of POIs of a certain category, node is all the stations in the traffic node space, POI k is the kth POI point, I is the indicator function;

[0117] S42. Calculate the proportion of each category in the node space by:

[0118]

[0119] Implementation, where ratio category is the proportion of a certain type of POI to all POIs in the node, count total is the total number of all four categories of POIs in the node;

[0120] S43. Extract indicators of the area where the location data overlaps with the node space, and construct a comprehensive feature vector for each node.

[0121] Furthermore, S5 is specifically:

[0122] S51. Determine whether the individual travel record is within the walking influence range of any public transportation node, specifically by:

[0123]

[0124] Implementation, where origin is the latitude and longitude coordinates of the starting point of the individual travel record, node i is the center coordinate index of the i-th traffic node space, d is the plane Euclidean distance calculation formula, Indicates the minimum distance from the starting point to all public transportation nodes. origin-pos The value is 1 or 0;

[0125] S52: First, check each individual travel record to see whether its main travel mode is bus or subway. If not, do not perform subsequent matching processing;

[0126] S52: Filter out candidate nodes from all public transportation nodes that have an intersection with the route number involved in the individual travel record, specifically by:

[0127]

[0128] Implementation, where record.lines is the set of transportation line numbers used in individual travel records, node.lines is the set of line numbers contained in the public transportation node space, and CandidateNodes origin A candidate node that intersects with the route number involved in the individual travel record;

[0129] S54. For the candidate nodes that intersect with the route numbers involved in the individual travel record, calculate the distance between their starting points and the centers of the nodes, specifically by:

[0130] d origin (node)=distance(record.origin,node.center)

[0131] Implementation, where record.origin is the latitude and longitude of the starting point in the same geographical location match, node.center is the latitude and longitude coordinates of the center point of the candidate node, d origin (node) is the distance between its starting point and the center of each node;

[0132] S55. Select the node with the smallest distance by:

[0133]

[0134] Implementation, where is the node with the smallest distance;

[0135] S56, distance judgment, if If the distance is less than 1000 meters, the travel record is considered to match the public transportation node space and the record is:

[0136] indicator origin =1

[0137]

[0138] Otherwise, both are set to 0, and origin_node_id is the unique identifier of the public transportation node matched to the travel origin:

[0139] S57. Integrate the individual travel record and the comprehensive feature vector of each node using the unique identifier of the public transportation node matched to the travel starting point to form a multi-source fused feature matching dataset.

[0140] Furthermore, S6 is specifically:

[0141] S61, use the three-layer LightGBM model;

[0142] At the first level, a binary classification model is used to distinguish between motorized travel and non-motorized travel;

[0143] At the second level, in motorized travel, the binary classification model further distinguishes between public transportation and non-public transportation; in non-motorized travel, the binary classification model further distinguishes between walking and non-motorized travel;

[0144] The third layer uses a binary classification model to distinguish between buses and subways for public transportation and taxis and self-driving cars for non-public transportation.

[0145] S62, directly use the classification accuracy as the training objective function;

[0146] Accuracy by:

[0147]

[0148] get;

[0149] S63. Using the built-in evaluation indicators of LightGBM, set the metric parameter to 'accuracy' to calculate the classification accuracy of the current model after each iteration;

[0150] Cross-validation is used to select the hyperparameter combination that maximizes Accuracy, including learning rate, tree depth, and minimum number of samples;

[0151] The model training of each layer uses accuracy as the main evaluation indicator, and the model finally selected can achieve the highest classification accuracy in its respective classification task;

[0152] S64. Calculate the SHAP value of each input feature using the SHAP method, where the input features include individual travel characteristics, node spatial characteristics, and user profile information;

[0153] S65. Calculate the SHAP interaction value to reveal the positive and negative impact of each functional layout feature under different user profiles and different travel purposes;

[0154] S66. Based on the results of the global SHAP analysis, rationally adjust the scale and spatial layout of public transportation node spaces to adapt them to the needs of the area and achieve intensive and balanced allocation of various functions within the public transportation node area.

[0155] Although the present invention has been described with respect to a limited number of embodiments, it will be apparent to those skilled in the art, having benefit of the foregoing description, that other embodiments are contemplated within the scope of the invention thus described. Furthermore, it should be noted that the language used in this specification has been selected primarily for readability and didactic purposes, rather than for the purpose of explaining or limiting the subject matter of the present invention. Consequently, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is intended to be illustrative rather than restrictive of the scope of the invention, which is defined by the appended claims.

Claims

1. A public transportation node efficiency analysis method based on multivariate data and interpretable machine learning, characterized by: The following steps are involved: S1. Obtain resident travel survey data, public transportation station and route data, geographic information data, residential and job population data, and point of interest data, and perform missing value processing and coordinate system analysis to obtain a spatially normalized multi-source dataset; S2. Calculating travel duration and travel distance based on the spatially normalized multi-source dataset, encoding the entire chain of transportation modes into digital primary travel modes based on preset rules, and obtaining individual travel records; S3. Based on the spatially normalized multi-source dataset, construct a public transportation node space by generating a 150-meter buffer zone and merging adjacent areas with a number of routes greater than or equal to 10; S4. Based on the public transportation node space and the spatially normalized multi-source dataset, the number and proportion of the four major categories of POIs in the node are statistically analyzed through spatial overlay analysis to obtain a comprehensive feature vector for each node; S5. Perform feature matching based on the individual travel records, the public transportation node space, and the comprehensive feature vector of each node to construct a feature matching dataset; S6. Based on the feature matching dataset, the key features and their nonlinear effects that affect public transportation travel decisions are identified through hierarchical LightGBM model training and SHAP interpretability analysis.

2. The public transportation node efficiency analysis method based on multivariate data and interpretable machine learning according to claim 1 is characterized in that: In S1, the resident travel survey data includes: household number, personal number, personal trip number, starting point longitude and latitude, destination longitude and latitude, departure time, arrival time, transportation mode for the entire travel chain, age, gender, education level, personal annual income, family annual income, work industry, and number of family vehicles; The public transportation station and line data include: bus station name, bus station longitude and latitude, bus line numbers of all bus routes that stop at the bus station, subway station name, subway station longitude and latitude, and subway line number where the subway station is located; The geographic information data includes: projection coordinate system information, urban road network information and urban boundary information; Residential and job population data include: urban population and job density with a grid accuracy of 250m*250m or urban population and job density with a grid accuracy of 80m*80m; The POI data includes: POI name, major category, minor category and latitude and longitude information.

3. The public transportation node efficiency analysis method based on multivariate data and interpretable machine learning according to claim 2 is characterized in that: In S2, the travel time is calculated by: T journey =T arrival -T departure Get, where T journey is the total travel time, T arrival is the arrival time of the trip record, T departure The departure time recorded for the trip; The travel distance is given by: Get, where d is the travel distance, x end is the horizontal coordinate of the plane projection of the travel destination, x start is the horizontal coordinate of the plane projection of the starting point of the trip, y end is the plane projection ordinate of the travel destination, y start It is the plane projection vertical coordinate of the starting point of the trip.

4. The public transportation node efficiency analysis method based on multivariate data and interpretable machine learning according to claim 3 is characterized in that: In S2, the individual travel record includes: personal economic and social characteristics, travel information and travel mode.

5. The public transportation node efficiency analysis method based on multivariate data and interpretable machine learning according to claim 4 is characterized in that: S3 specifically: S31. Generate a buffer zone with a radius of 150 meters with each bus stop as the center. The buffer zone contains all bus route information of the bus stop. Overlay buffers and calculate the number of bus routes passing through each overlay area. The number of bus routes is calculated by: Get, where count spatial_join is the number of bus routes in the superimposed buffer zone, count i is the number of lines with a single station participating in the superposition; S32, if count spatial_join If ≥10, all the bus stops associated with it constitute a transportation node, and the combined area of the 150-meter buffer zone formed by all the stops is regarded as the transportation node space.

6. The public transportation node efficiency analysis method based on multivariate data and interpretable machine learning according to claim 5, characterized in that: S4 is specifically: S41: spatially superimpose the public transportation node space with the POI data, extract POI information within the public transportation node space, and screen POIs in four categories: catering, shopping, life services, and leisure and entertainment. Specifically, the following steps are performed: Implementation, where category is catering or shopping or life services or leisure and entertainment, count category is the number of POIs of a certain category, node is all the stations in the traffic node space, POI k is the kth POI point, I is the indicator function; S42. Calculate the proportion of each category in the node space by: Implementation, where ratio category is the proportion of a certain type of POI to all POIs in the node, count total is the total number of all four categories of POIs in the node; S43. Extract indicators of the area where the location data overlaps with the node space, and construct a comprehensive feature vector for each node.

7. The public transportation node efficiency analysis method based on multivariate data and interpretable machine learning according to claim 6 is characterized in that: S5 is specifically: S51. Determine whether the individual travel record is within the walking influence range of any public transportation node, specifically by: Implementation, where origin is the latitude and longitude coordinates of the starting point of the individual travel record, node i is the center coordinate index of the i-th traffic node space, d is the plane Euclidean distance calculation formula, Indicates the minimum distance from the starting point to all public transportation nodes. origin-pos The value is 1 or 0; S52: First, check each individual travel record to see whether its main travel mode is bus or subway. If not, no subsequent matching process is performed; S52: Filter out candidate nodes from all public transportation nodes that have an intersection with the route number involved in the individual travel record, specifically by: Implementation, where record.lines is the set of transportation line numbers used in individual travel records, node.lines is the set of line numbers contained in the public transportation node space, and CandidateNodes origin A candidate node that intersects with the route number involved in the individual travel record; S54. For the candidate nodes that intersect with the route numbers involved in the individual travel record, calculate the distance between their starting points and the centers of the nodes, specifically by: d origin (node)=distance(record.origin,node.center) Implementation, where record.origin is the latitude and longitude of the starting point in the same geographical location match, node.center is the latitude and longitude coordinates of the center point of the candidate node, d origin (node) is the distance between its starting point and the center of each node; S55. Select the node with the smallest distance by: Implementation, where is the node with the smallest distance; S56, distance judgment, if If the distance is less than 1000 meters, the travel record is considered to match the public transportation node space and the record is: indicator origin =1 Otherwise, both are set to 0, and origin_node_id is the unique identifier of the public transportation node matched to the travel origin: S57. Integrate the individual travel record and the comprehensive feature vector of each node using the unique identifier of the public transportation node matched to the travel starting point to form a multi-source fused feature matching dataset.

8. The public transportation node efficiency analysis method based on multivariate data and interpretable machine learning according to claim 1 is characterized in that: S6 specifically: S61, use the three-layer LightGBM model; At the first level, a binary classification model is used to distinguish between motorized travel and non-motorized travel; At the second level, in motorized travel, the binary classification model further distinguishes between public transportation and non-public transportation; in non-motorized travel, the binary classification model further distinguishes between walking and non-motorized travel; The third layer uses a binary classification model to distinguish between buses and subways for public transportation and taxis and self-driving cars for non-public transportation. S62, directly use the classification accuracy as the training objective function; Accuracy by: get; S63. Using the built-in evaluation indicators of LightGBM, set the metric parameter to 'accuracy' to calculate the classification accuracy of the current model after each iteration; Cross-validation is used to select the hyperparameter combination that maximizes Accuracy, including learning rate, tree depth, and minimum number of samples; The model training of each layer uses accuracy as the main evaluation indicator, and the model finally selected can achieve the highest classification accuracy in its respective classification task; S64. Calculate the SHAP value of each input feature using the SHAP method, where the input features include individual travel characteristics, node spatial characteristics, and user profile information; S65. Calculate the SHAP interaction value to reveal the positive and negative impact of each functional layout feature under different user profiles and different travel purposes; S66. Based on the results of the global SHAP analysis, rationally adjust the scale and spatial layout of public transportation node spaces to adapt them to the needs of the area and achieve intensive and balanced allocation of various functions within the public transportation node area.

Images