Method and system for predicting resident travel selection in dynamic evolution of rail transit network

Abstract

This invention belongs to the field of resident travel choice prediction technology, and particularly relates to a method and system for predicting resident travel choices in the dynamic evolution of rail transit networks. It includes preprocessing multi-source resident travel data, dividing it into training and test sets; constructing a gradient boosting decision tree model, training the model using the training set, and optimizing the model parameters using a GBDT hyperparameter automatic tuning method based on multi-objective Bayesian optimization; obtaining prediction results on the test set using the trained model, calculating the SHAP value and partial dependency graph of each influencing factor, and obtaining the nonlinear mechanism and dynamic evolution law of each influencing factor. This invention uses multi-source resident travel data from different periods of the dynamic evolution of rail transit networks for analysis and research, and innovatively introduces a multi-objective Bayesian optimization framework for automatic hyperparameter tuning, dynamically capturing the nonlinear laws of resident travel choices during the evolution of rail transit networks.

Description

Technical Field

[0001] This invention belongs to the field of resident travel choice prediction technology, and in particular relates to a method and system for predicting resident travel choices based on the dynamic evolution of rail transit networks. Background Technology

[0002] Accelerated urbanization has led to a surge in motor vehicle ownership, highlighting traffic congestion and environmental pollution. Rail transit has become a core strategy for cities to alleviate traffic pressure and achieve sustainable development. The rapid expansion of rail transit networks is reshaping urban spatial structures and residents' travel patterns. Accurately understanding residents' travel choices under these dynamic changes is crucial for optimizing route planning, improving operational efficiency, and formulating relevant policies.

[0003] Existing research on residents' travel choices under the dynamic evolution of rail transit networks has three limitations: First, the analysis mainly uses static cross-sectional data, which fails to reveal the dynamic changes in the factors and mechanisms influencing travel decisions during network evolution; Second, traditional linear models are difficult to capture complex nonlinear relationships and threshold effects. Although machine learning models are introduced, existing hyperparameter tuning methods are inefficient and have a single objective, which can easily lead to model overfitting and make it difficult to balance prediction accuracy and generalization ability. Third, the "black box" nature of high-precision machine learning models hinders the transformation of research results into practical strategies. Summary of the Invention

[0004] To overcome the shortcomings of the existing technologies, this invention provides a method and system for predicting residents' travel choices in the dynamic evolution of rail transit networks. It designs a research framework integrating machine learning and interpretability analysis, analyzes multi-source data on residents' travel at different stages of the dynamic evolution of rail transit networks, and innovatively introduces a multi-objective Bayesian optimization framework for automatic hyperparameter tuning. Ultimately, it can dynamically capture the nonlinear patterns of residents' travel choices during the evolution of rail transit networks and reveal the underlying driving mechanisms. This dynamically and precisely characterizes residents' rail transit travel choices, filling the technological gap between macro-level passenger flow prediction and micro-level individual decision-making understanding, and providing a more refined and forward-looking planning basis for the design of subway rail transit.

[0005] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for predicting residents' travel choices based on the dynamic evolution of rail transit networks.

[0006] A method for predicting residents' travel choices based on the dynamic evolution of rail transit networks includes the following steps: Multiple influencing factors were identified to obtain multi-source data on residents' travel during different periods of the dynamic evolution of the rail transit network; Preprocess the multi-source data on residents' travel to divide it into training and test sets; A gradient boosting decision tree model is constructed, and the model is trained using the training set. The model parameters are optimized using the GBDT hyperparameter automatic tuning method based on multi-objective Bayesian optimization. The trained model is used to obtain prediction results on the test set. The SHAP value and partial dependency graph of each influencing factor are calculated to obtain the nonlinear action mechanism and dynamic evolution law of each influencing factor.

[0007] The second aspect of this invention provides a resident travel choice prediction system for the dynamic evolution of rail transit networks.

[0008] A predictive system for residents' travel choices based on the dynamic evolution of rail transit networks includes: The data acquisition module is configured to: divide multiple influencing factors and acquire multi-source data on residents' travel at different stages of the dynamic evolution of the rail transit network; The preprocessing module is configured to preprocess multi-source data on residents' travel and divide it into training and test sets. The parameter tuning module is configured to: construct a gradient boosting decision tree model, train the model using the training set, and optimize the model parameters using the GBDT hyperparameter automatic tuning method based on multi-objective Bayesian optimization. The data analysis module is configured to: use the trained model to obtain prediction results on the test set, calculate the SHAP value and partial dependency graph of each influencing factor, and obtain the nonlinear action mechanism and dynamic evolution law of each influencing factor.

[0009] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the method for predicting resident travel choices in the dynamic evolution of rail transit networks as described in the first aspect of the present invention.

[0010] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the method for predicting resident travel choices in the dynamic evolution of rail transit networks as described in the first aspect of the present invention.

[0011] The above one or more technical solutions have the following beneficial effects: (1) Methodological Integration and Innovation: This invention systematically integrates the strong predictive power of Gradient Boosting Decision Tree (GBDT) with the strong interpretive power of SHAP value and Partial Dependency Graph (PDP), constructing an integrated "prediction-interpretation" analytical framework. This framework can not only simulate residents' travel choices with high accuracy, but also quantitatively and visually reveal the nonlinear mechanism and dynamic evolution of various influencing factors, effectively breaking through the "black box" application bottleneck of machine learning models in the field of public policy research.

[0012] (2) Innovative Data Preprocessing: A dual data optimization process was innovatively introduced before machine learning modeling. First, propensity score matching (PSM) was used to address sample selection bias in cross-period comparative studies, ensuring the scientific rigor of dynamic analysis. Then, SMOTEENN mixed sampling technology was used to address class imbalance, improving the model's accuracy in identifying the critical minority class (choosing rail transit travel). This preprocessing process significantly improved data quality and the reliability of subsequent model analysis.

[0013] (3) Innovative Parameter Optimization Strategy: Addressing the technical bottlenecks of low computational efficiency and difficulty in balancing multiple objectives in traditional grid search methods, this invention innovatively introduces a multi-objective Bayesian optimization framework for automatic hyperparameter tuning. This method constructs a surrogate model and acquisition function to achieve intelligent directional search in the parameter space, simultaneously optimizing both model prediction accuracy and structural complexity, and automatically generating a Pareto optimal solution set. Compared to traditional methods, this strategy significantly improves parameter optimization efficiency and significantly reduces model complexity while maintaining prediction accuracy, providing a stable and reliable parameter benchmark for long-term dynamic comparative studies and ensuring the optimal balance between model performance and generalization ability.

[0014] (4) Innovation in research perspective and application value: This invention breaks through the traditional research focus on the macro perspective of "station passenger flow" and refines the analysis granularity to the level of "individual travel decision-making". Through dynamic comparison, it clearly depicts how the macro urban process of rail transit network expansion microscopically affects individual behavior, thus providing urban decision-makers with more refined and forward-looking planning basis, such as paying more attention to connection services and multi-center layout during the network maturity period, rather than simply pursuing line extension.

[0015] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a flowchart illustrating the method steps of Example 1.

[0018] Figure 2 This is a flowchart of the method in Example 1.

[0019] Figure 3 This is a schematic diagram illustrating the principle of the GBDT algorithm in Example 1.

[0020] Figure 4(a) shows the geographical distribution of the 2018 sample of Example 1.

[0021] Figure 4(b) shows the geographical distribution of the sample in 2023 for Example 1.

[0022] Figure 5(a) is a visualization of the relative importance of features in 2018 for Example 1.

[0023] Figure 5(b) is a visualization of the relative importance of features in 2023 for Example 1.

[0024] Figure 6 This is a partial age dependency graph for 2018 and 2023 in Example 1.

[0025] Figure 7 This is a partial dependency graph of household car ownership in 2018 and 2023 for Example 1.

[0026] Figure 8 This is a partial dependency graph of electric bicycle ownership in 2018 and 2023 for Example 1.

[0027] Figure 9 This is a partial dependency graph of commuting distance in 2018 and 2023 for Example 1.

[0028] Figure 10 This is a nonlinear correlation diagram between the distance from the residence to the nearest subway station and the probability of taking rail transit, as shown in Example 1.

[0029] Figure 11 This is a nonlinear correlation graph showing the distance from the workplace to the nearest subway station and the probability of taking rail transit, as shown in Example 1.

[0030] Figure 12 This is a nonlinear correlation graph showing the distance from the residence to the city center and the probability of taking rail transit, as shown in Example 1.

[0031] Figure 13 This is a nonlinear correlation graph showing the distance from the workplace to the city center and the probability of taking rail transit, as shown in Example 1.

[0032] Figure 14 This is a nonlinear correlation diagram between the distance from the residence to the secondary city center and the probability of taking rail transit, as shown in Example 1.

[0033] Figure 15 This is a nonlinear correlation diagram between the distance from the workplace to the secondary city center and the probability of taking rail transit, as shown in Example 1. Detailed Implementation

[0034] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0035] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0036] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0037] Example 1 This embodiment discloses a method for predicting residents' travel choices in the dynamic evolution of rail transit networks. It proposes a research framework that integrates machine learning and interpretability analysis, aiming to dynamically capture the nonlinear patterns of residents' travel choices during the evolution of rail transit networks and reveal the driving mechanisms behind them. It dynamically and precisely characterizes residents' rail transit travel choice behavior, deeply explains its internal driving mechanisms, and fills the technical gap between macro-level passenger flow prediction and micro-level individual decision-making understanding.

[0038] like Figure 1 As shown, the method for predicting residents' travel choices based on the dynamic evolution of rail transit networks includes the following steps: Multiple influencing factors were identified to obtain multi-source data on residents' travel during different periods of the dynamic evolution of the rail transit network; Preprocess the multi-source data on residents' travel to divide it into training and test sets; A gradient boosting decision tree model is constructed, and the model is trained using the training set. The model parameters are optimized using the GBDT hyperparameter automatic tuning method based on multi-objective Bayesian optimization. The trained model is used to obtain prediction results on the test set. The SHAP value and partial dependency graph of each influencing factor are calculated to obtain the nonlinear action mechanism and dynamic evolution law of each influencing factor.

[0039] like Figure 2 As shown, this embodiment provides a method for analyzing residents' rail transit travel choices based on machine learning, interpretability, and dynamic comparison. Its core innovation lies in constructing a four-stage collaborative optimization technical framework, specifically including: (1) Cross-period sample alignment stage: Propensity score matching (PSM) was used to eliminate systematic bias between the 2018 and 2023 survey samples to ensure the scientific nature of the dynamic comparison; (2) Class balance enhancement stage: The SMOTEENN hybrid sampling technology is introduced to simultaneously solve the problems of minority class oversampling and majority class noise removal, which significantly improves the model's sensitivity to the recognition of rail transit travel behavior; (3) Multi-objective parameter optimization stage: Introduce a multi-objective Bayesian optimization framework to simultaneously optimize the dual objectives of prediction accuracy and structural complexity of the GBDT model. The optimal hyperparameter combination is automatically obtained through Pareto front analysis to improve the model's generalization ability while ensuring prediction performance. (4) Nonlinear prediction modeling stage: Based on gradient boosting decision tree (GBDT), a high-precision prediction model is constructed to automatically capture the high-order interaction and threshold effect between personal attributes, traffic characteristics and built environment attributes; (5) Mechanism Interpretation Mining Stage: Integrating SHAP values ​​and Partial Dependency Graph (PDP) to achieve multi-granularity interpretation from “global feature importance” to “local nonlinear response”, and directly outputting decision-making basis for planning practice.

[0040] This research framework achieves the transition from "data bias correction" to "behavioral mechanism explanation," addressing the fundamental shortcomings of existing methods in terms of dynamism, robustness, and interpretability.

[0041] (1) Gradient Boosting Decision Tree (GBDT) Model Model Introduction In this embodiment, we employ the Gradient Boosting Decision Tree (GBDT) model to analyze the nonlinear relationship between personal attribute characteristics, traffic attribute characteristics, and built environment characteristics and the rail transit travel behavior of residents in a certain city. GBDT is a powerful ensemble learning method that progressively improves model performance by integrating multiple weak learners (usually decision trees), and is widely used in tasks such as regression, classification, and ranking. GBDT has significant advantages, especially in nonlinear data modeling.

[0042] GBDT stands for Gradient Boosting, a decision tree model trained using the Gradient Boosting strategy. Gradient Boosting is an ensemble learning method. Ensemble learning improves prediction performance by combining multiple models. Its basic idea is to combine multiple weak learners (usually those slightly better than random guessing learners) into a strong learner. Ensemble learning accomplishes the learning task by constructing and combining multiple learners. Gradient Boosting is an implementation of Boosting, primarily using gradient descent to gradually reduce model bias. In each iteration, Gradient Boosting trains a new weak learner based on the prediction error (residual) of the previous model, and the final prediction result is the weighted sum of all model predictions.

[0043] For example, let's assume there are M weak learners. The final strong learner It is a weighted combination of these weak learners: (1) in, It is the weight of the m-th weak learner.

[0044] Meanwhile, the decision tree used in the GBDT model is the CART regression tree. The regression tree uses mean squared error as the loss function, and during tree generation, it recursively partitions the space according to the optimal feature and the optimal value under the optimal feature until a stopping condition is met. The steps for constructing the regression tree are as follows: Input: Training dataset

[0045] Output: Regression Tree T (1) Solve for the selection of the segmentation feature j and the segmentation feature value s. J divides the training set D into two parts. and After being segmented according to (j,s), it is as follows: (2) (3) in, , These are the means of the left and right sub-regions, respectively. The optimal split point can be found by iterating through every possible value of each variable to divide the dataset.

[0046] (2) Iterate through all possible solutions (j,s) to find the optimal one. The optimal solution minimizes the corresponding loss, according to the optimal feature. You can use ) to divide it.

[0047] (3) Recursively call (1)~(2) until the stopping condition is met.

[0048] (4) Return to decision tree T Note: The stopping condition can be set manually. For example, the sample size of a certain node can be set to be less than a given threshold c, or the loss reduction after splitting can be less than a given threshold or less than a given ε, then the splitting can be stopped and leaf nodes can be generated.

[0049] Algorithm Steps The core idea of ​​the GBDT model is to gradually improve the model's predictive ability by iteratively training multiple weak learners (usually decision trees). For example... Figure 3 As shown, each new tree fits the "residual" of the current model to correct the errors of the previous model, thereby improving the overall prediction accuracy. The specific steps are as follows: 1. Initialize the model: Initialize a constant value as the first step in the model's prediction.

[0050] (4) in, It is a loss function. This is the initial predicted value.

[0051] 2. Iterative training tree: For each iteration Perform the following steps: Calculate the residual for each sample based on the difference between the current model's predicted values ​​and the actual values: (5) in, Indicates the first The sample at the th The wheel's residual.

[0052] Train a new decision tree using the residuals as the objective. To predict residuals: (6) 3. Update the model: Update the current model using a linear weighted approach. (7) in, It is the learning rate, which controls the step size of each iteration update.

[0053] 4. Output the final model: The final model is a weighted sum of all decision trees: (8) GBDT offers significant advantages over traditional linear models (such as linear regression and logistic regression), especially when handling non-linear relationships. This machine learning model effectively captures complex non-linear relationships between features. It doesn't require assuming linear relationships between features; instead, it automatically captures complex relationships through decision tree splitting. Secondly, the GBDT model is robust to outliers and noise in the data. It can find reasonable split points within different data intervals and is less susceptible to extreme values. During each decision tree split, it can handle missing values ​​and learn from the remaining valid data. However, in traditional linear regression, handling missing values ​​requires preprocessing steps (such as imputing missing values); otherwise, the model may not be well-suited to this missing data.

[0054] (2) SHAP value analysis Machine learning possesses remarkable learning capabilities, but as a black-box model, it inevitably suffers from certain problems. These include the difficulty in interpreting the model's output and its over-reliance on data; if the data is perturbed, the model's output can change significantly, even leading to completely opposite results. Therefore, it is necessary to infer the causal relationship between the results of machine learning algorithms and their input features to enhance the model's interpretability. Currently, commonly used methods include attribution analysis-based explanations and counterfactual explanations. Common methods based on attribution analysis include Shapley Values ​​and LIME; counterfactual explanations primarily use a sample feature processing method to change the model's results, thereby proving the causal relationship between features and results. This paper selects SHAP values ​​based on attribution analysis to explain the proposed model.

[0055] SHAP (Shapley Additive exPlanations) is a game theory-based machine learning interpretation method that measures the contribution of each feature to the model's predictions. It originates from Shapley values ​​in game theory and was proposed by economist Lloyd Shapley in the 1950s. This method obtains the contribution value of each feature, representing the magnitude and direction of that feature's contribution to the model's predictions. The mathematical formula is as follows: (9) in, It is a feature The SHAP value; It is a subset of features (excluding features) ); It is the total set of features; It is a feature set The model output; It is a feature set Add features The model output after that.

[0056] SHAP values ​​have three important properties: local interpretability: SHAP values ​​interpret individual predictions, not the behavior of the entire model; global consistency: the sum of the SHAP values ​​of all features equals the model's predicted value minus the baseline value (such as the baseline prediction value); and fairness: SHAP values ​​ensure that the contribution of each feature is fairly distributed, conforming to the axioms of Shapley values.

[0057] The technical solution of this embodiment will now be explained in detail with reference to the accompanying drawings.

[0058] (1) Multi-source data acquisition and structured processing.

[0059] This embodiment primarily utilizes three types of data: personal attribute data, traffic attribute data, and built-up area environmental data. All three types of data originate from resident travel surveys conducted in a certain city in 2018 and 2023. The resident travel survey is a household sampling survey. Researchers used residential communities within the urban area of ​​the city (six administrative districts: Gusu District, Industrial Park, High-tech Zone, Wuzhong District, Xiangcheng District, and Wujiang District) as sampling units, employing a structured survey through face-to-face interviews. The sample distribution is shown in Figures 4(a) and 4(b). The questionnaire required respondents to fill in personal attribute information of their families and family members, daily travel information, commuting information, and housing characteristics. A total of 12,598 survey data points were collected in 2018, and 9,745 survey data points were collected in 2023. After screening and cleaning the survey data, removing missing and duplicate data, we ultimately obtained 10,510 valid travel data points from 2018 and 6,732 valid travel data points from 2023.

[0060] Collect "Resident Travel Survey" data for the target city (City X) at different stages of development (early and mature stages of rail transit network formation). This data should structure and include three core characteristic variables: 1) Individual and family attributes: including age, number of cars owned by the family, and number of electric bicycles owned by the family.

[0061] 2) Transportation characteristics: including commuting distance, distance from residence to the nearest subway station, and distance from workplace to the nearest subway station.

[0062] 3) Architectural environment characteristics: Based on the “5D” dimensions, the key points include: distance from residence to city center, distance from workplace to city center, distance from residence to the nearest urban sub-center, and distance from workplace to the nearest urban sub-center.

[0063] In this embodiment, the dependent variable is a binary indicator representing whether residents choose rail transit (1 = choose rail transit, 0 = do not choose rail transit). This variable is derived from the survey of residents' main modes of transportation in the resident travel survey questionnaire.

[0064] Categorical variables such as age are coded using ordinal scales, while numerical variables such as vehicle ownership and distance are treated as continuous variables. Table 1 lists the descriptive statistics for all independent variables.

[0065] Table 1. Variable Description and Statistical Analysis

[0066] (2) Data preprocessing and sample optimization Sample Range Matching (PSM) To eliminate systematic biases caused by natural evolution (such as urban expansion and changes in population structure) in samples from different years, a propensity score matching method was adopted. Using the "survey year" as the treatment variable and all characteristic variables as covariates, a logistic regression model was constructed to calculate the propensity score for each sample. The specific formula is as follows: (10) in, It is an intervention variable; It indicates 2023. Indicates 2018; It is a vector of covariates; Indicates the first One sample; It is the regression coefficient.

[0067] Then, nearest neighbor matching is used, employing Euclidean distance to calculate the distance between samples.

[0068] (11) in, and These are samples and samples In the Propensity scores on each covariate; It is a sample and samples The Euclidean distance between them represents their similarity in the propensity score space.

[0069] Considering that the initial sample size in 2018 (10510) is greater than that in 2023 (6732), we set that each sample from 2023 can be matched with a maximum of three samples from 2018, and that no sample from 2018 is matched repeatedly. Finally, we obtained 5875 data points from 2018 and 5869 data points from 2023.

[0070] Meanwhile, this embodiment uses the standardized mean difference (SMD) of each covariate between 2018 and 2023 to test the matching effect. As shown in Table 2, the SMD of each covariate is less than 0.1, indicating a good matching effect (the closer the SMD is to 0, the more similar the distributions of the two sets of data are after matching).

[0071] Table 2. Effects of propensity score matching variable Standardized mean difference (SMD) age 0.0777 Car ownership 0.0332 Electric vehicles -0.0638 Commuting distance -0.0214 Distance from residence to the nearest subway station 0.0255 Distance from workplace to the nearest subway station 0.0447 Distance from residence to city center 0.0281 Distance from workplace to city center 0.0401 Distance from residence to secondary center 0.0299 Distance from workplace to secondary center 0.0459 Sample class balance (SMOTEENN) To address the class imbalance problem in rail transit travel samples, which are typically classified as a "minority class," a SMOTEENN hybrid sampling technique was employed. First, the SMOTE method was used to oversample the minority class samples, generating synthetic samples through K-nearest neighbor interpolation. Then, the ENN method was used to undersample the majority class samples, removing noisy or boundary samples whose classes are inconsistent with their neighbors. This process significantly improved the model's ability to identify the minority class and prevented the model from biased towards predicting the majority class. Ultimately, 10,327 data points were obtained in 2018 and 7,815 data points in 2023, with the specific sample distribution shown in Table 3.

[0072] Table 3 SMOTEENN processing results

[0073] (3) Construction and training of nonlinear machine learning models Parameter tuning based on multi-objective Bayesian optimization The GBDT algorithm gradually reduces prediction error through multiple iterations; however, this approach can easily lead to overfitting on the training data, especially when the tree depth is large or the number of training iterations is excessive. To address this issue, this embodiment utilizes the Scikit-learn library in Python 3.10 to develop the model and optimizes the model parameters through the following steps: (1) Dataset partitioning: For the data from 2018 and 2023, this embodiment randomly partitions the dataset into a training set and a test set, with the training set containing 80% of the samples and the test set containing 20% ​​of the samples. This ensures that the model can access enough data during training, while also reserving some data for model performance verification.

[0074] (2) Parameter Optimization: To find the optimal combination of model parameters, this embodiment proposes an automatic hyperparameter tuning method for GBDT based on multi-objective Bayesian optimization. This method combines the Bayesian optimization framework with the theory of multi-objective optimization, enabling simultaneous optimization of multiple potentially conflicting objective functions. In the hyperparameter tuning of the Gradient Boosting Decision Tree (GBDT) model, this method not only pursues the optimization of model prediction performance but also takes into account the control of model complexity, thereby obtaining a parameter configuration that achieves the best balance among multiple objectives.

[0075] Compared to traditional grid search methods, the multi-objective Bayesian optimization automatic tuning proposed in this embodiment significantly reduces the number of model training and evaluation iterations required through intelligent sampling strategies and surrogate models, thereby improving the efficiency of hyperparameter search. Furthermore, it can simultaneously optimize multiple objectives and find the optimal balance point among them, avoiding the problems of model overcomplexity or underfitting that may result from single-objective optimization. The specific principles and implementation steps are as follows: a. Multi-objective function design Traditional single-objective optimization aims to find the solution that optimizes a single objective function, while multi-objective optimization requires optimizing multiple objective functions simultaneously. For the GBDT hyperparameter optimization problem, we formalize it as follows: (12) in, Represents the hyperparameter space. For hyperparameter combination, For the first One objective function.

[0076] In this embodiment, we set two main optimization objectives: minimizing prediction error and minimizing model complexity.

[0077] Mean squared error (MSE) is used as a metric for model predictive performance: (13) Taking into account multiple structural parameters of the GBDT model, a complexity evaluation function is constructed: (14) Among them, the weighting coefficient , , , This reflects the relative impact of different parameters on the complexity of the model.

[0078] b. Dopareto optimality In multi-objective optimization, due to the frequent conflicts between objectives, there is usually no single solution that simultaneously optimizes all objectives. Pareto optimality defines the concept of an optimal solution in the multi-objective case: a solution that is not dominated by any other solution is called a Pareto optimal solution. The surface formed in the objective space by all Pareto optimal solutions is called the Pareto front.

[0079] After obtaining the Pareto front, we need to select a final solution from it for model construction. This embodiment uses a weighted normalization method: Objective value normalization: Normalize each objective value of each solution on the Pareto front.

[0080] Trade-off score calculation: Calculate the trade-off score for each solution: (15) Among them, take , .

[0081] Optimal solution selection: Select the solution with the highest trade-off score as the final hyperparameter configuration.

[0082] c. Implementation steps The basic optimization process is as follows: Initialization: Randomly select a small number of hyperparameter combinations for preliminary evaluation; Constructing a proxy model: Based on existing evaluation results, establish a probabilistic model of the relationship between the objective function and hyperparameters; Select the next evaluation point: Determine the most promising combination of hyperparameters based on the acquisition function; Evaluation and Update: Evaluate the performance of the selected hyperparameter combination and update the surrogate model; Iteration: Repeat steps 2-4 until the preset number of evaluations is reached.

[0083] Through multi-objective Bayesian optimization and automatic parameter tuning, this embodiment ultimately determines the optimal parameter combinations for the 2018 and 2023 models as follows: 2018: n_estimators=158, learning_rate=0.10717364601244753, max_depth=8, min_samples_split=4, min_samples_leaf=5, subsample=0.9881098752741272.

[0084] 2023: n_estimators=56, learning_rate=0.12026008341996716, max_depth=8, min_samples_split=5, min_samples_leaf=5, subsample=0.6375689228139025.

[0085] (3) Model performance evaluation Five-fold cross-validation can estimate a model's performance and generalization ability. It involves dividing the original dataset into five equal subsets: four subsets are used as the training dataset, and the remaining subset is used to test the model.

[0086] Combining this method enables the model to have good predictive ability on both training and testing data. Therefore, in this section, to effectively evaluate the model's training and prediction performance, this embodiment selects four metrics—mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²)—to evaluate model performance. These metrics comprehensively reflect the model's prediction accuracy and fitting ability from different perspectives. MSE and RMSE are mainly used to measure the degree of deviation between predicted and true values, MAE focuses more on the absolute value of the error, and R² is used to measure the model's ability to interpret data. As shown in Tables 4 and 5, this paper compares the models from 2018 and 2023, respectively, using the grid search method and multi-objective Bayesian optimization for automatic tuning of the four metrics.

[0087] Table 4 Comparison of the effects of different parameter optimization methods in the model in 2018

[0088] Table 5. Comparison of the effects of different parameter optimization methods in the 2023 model.

[0089] As shown in the table, the multi-objective Bayesian optimization method significantly improves model performance compared to the traditional grid search method. In Table 4, for the 2018 data, parameter tuning using multi-objective Bayesian optimization reduced the MSE of the GBDT model from 0.03750 to 0.01622, a decrease of 56.7%; the MAE improved by 48.2%; and the R² increased from 0.84977 to 0.93500, enhancing the model's interpretability by 10.0%. A similar improvement trend is evident in the 2023 data, with a 17.2% decrease in MSE, a 12.7% improvement in MAE, and an 8.0% increase in R². This verifies the effectiveness of multi-objective Bayesian optimization in hyperparameter tuning.

[0090] (4) Model interpretability analysis and mechanism mining Feature Importance Quantification (SHAP) This embodiment employs a game theory-based SHAP value interpretation framework. For each prediction result of the model on the test set, the SHAP value of each feature variable is calculated. This value quantifies the magnitude and direction (positive or negative) of the marginal contribution of the feature to the prediction result. By summing the SHAP values ​​of all samples, a global feature importance ranking can be obtained, accurately identifying the key driving factors influencing residents' rail transit travel choices. Table 6 shows the relative importance of all feature variables to residents' rail transit travel.

[0091] Table 6. Ranking of Feature Importance

[0092] With the continuous expansion of the rail transit network, the influence of traffic characteristics on rail transit travel has increased significantly, reaching a total relative importance of 59.45%, while the influence of built environment characteristics has decreased to 23.28%. This change reflects that residents' choice of rail transit travel is no longer largely limited by the distribution of the built environment, but is more influenced by the convenience of transportation. Commuting time and distance from the subway station are currently the main factors that residents consider when choosing rail transit travel.

[0093] Overall, from 2018 to 2023, the gradual expansion of the rail transit network significantly altered the main factors influencing residents' rail transit travel. Initially, architectural environment characteristics dominated, while later, the influence of traffic characteristics increased significantly. This reflects the profound impact of urban transportation systems on residents' travel behavior.

[0094] Meanwhile, this embodiment uses a summary plot of SHAP values ​​for visualization, which can show the marginal contribution of feature variables to the model prediction results and their impact mechanism. Figure 5 shows the ranking and trend of the importance of various feature variables to residents' rail transit travel in 2018 and 2023.

[0095] As shown in Figure 5(a), in 2018, the rail transit network was still in its development stage, with limited coverage and insufficient station density. At this stage, "distance from workplace to the nearest urban sub-center" and "distance from residence to the nearest urban sub-center" were the two most important factors influencing residents' rail transit travel, both with high relative importance. Furthermore, these two characteristics had a significant negative impact on rail transit travel; that is, the farther a resident's workplace or residence was from the urban sub-center, the less likely they were to choose rail transit.

[0096] By 2023, as shown in Figure 5(b), with the further expansion of the rail transit network, commuting distance became the most critical factor influencing residents' rail transit travel, and its relative importance increased significantly. This change also illustrates the increased attractiveness of the expanded rail transit network for long-distance commutes, especially the extension of rail transit lines and the increase in station density, making rail transit the preferred mode of transportation for long-distance commutes. In 2023, the second significant variable was "distance from workplace to the nearest subway station," which had a significant negative impact on rail transit travel. This phenomenon indicates that although the coverage of the rail transit network has expanded, station accessibility remains an important factor influencing residents' travel choices.

[0097] Visualization of Nonlinear Relationships (PDP) This embodiment uses a partial dependency plot (PDP) to depict the average marginal effect of a single feature on the model's predicted output (i.e., the probability of choosing rail transit) under different values ​​by marginalizing the influence of all other features. This plot visually illustrates the nonlinear patterns, trends, and potential threshold effects of the influence. This embodiment first plots the PDP for 10 variables—personal attribute features, transportation attribute features, and built environment features—in 2018 and 2023. Then, it compares and analyzes the differences in the nonlinear impact of the same variable on the probability of residents choosing rail transit at different stages of rail transit network expansion.

[0098] (1) Age like Figure 6 As shown, in 2023, age was negatively correlated with residents' overall rail transit usage, with the 6-30 age group having the highest probability of rail transit use. Although the 61-80 age group had a slightly higher probability of rail transit use than the 46-60 age group, the overall trend was that older people used rail transit less. In 2018, the 21-40 age group had the highest probability of rail transit use, followed by students aged 6-20. Comparatively, whether in 2018 when rail transit was less developed or in 2023 when the rail transit network was more complete, younger people were more willing to choose rail transit, while older people were less likely to choose it.

[0099] (2) Number of family cars and electric bicycles Figure 7The study shows the relationship between household car ownership and the probability of residents choosing rail transit. In both 2018 and 2023, there was a trend that the more cars a household owned, the lower the probability of rail transit use, indicating a competitive relationship between private cars and rail transit. However, by 2023, the gap in the probability of rail transit use between households with one car and those without narrowed significantly, indicating that the improved rail transit network attracted more car owners to switch to rail transit, and residents' travel choices became more comprehensive.

[0100] Figure 8 This illustrates the relationship between electric vehicles and rail transit. In 2023, residents without electric vehicles were most likely to use rail transit, decreasing as the number of electric vehicles increased. In 2018, however, having one electric vehicle resulted in the highest probability of rail transit use. This may be because in 2018, the rail transit network was less developed, and electric vehicles served as a supplementary connecting mode of transport. By 2023, the rail network had wider coverage and denser stations, reducing reliance on electric vehicles, and the two became competitors for short-distance travel.

[0101] (3) Commuting distance Figure 9 The study shows a non-linear relationship between commuting distance and the probability of residents using rail transit. In 2023, the two were positively correlated, indicating that as the rail network improved, its ability to serve long-distance commutes increased, attracting more long-distance commuters. In contrast, the situation in 2018 was different: within 10 kilometers, residents chose walking or cycling more often, resulting in lower rail transit usage; the 10-20 kilometer range was the highest point of travel probability due to the obvious advantages of rail transit in speed, punctuality, and cost; however, when the distance exceeded 20 kilometers, the probability of rail travel decreased due to insufficient network coverage and limited service capacity at that time.

[0102] (4) Distance from residence (workplace) to the nearest subway station like Figure 10 and Figure 11 As shown, the farther the residence or workplace is from the subway station, the lower the probability of residents choosing rail transit, showing an overall negative correlation—a phenomenon consistent with conventional understanding. Comparing 2023 and 2018 reveals that with the further expansion of the rail transit network, the distance from residence and workplace to the subway station has a more significant impact on travel choices. This conclusion is consistent with the changes in the importance ranking of SHAP features: the rankings of both distance variables increased in 2023.

[0103] (5) Distance from residence (workplace) to city center like Figure 12 and Figure 13As shown, the distance from residence to the city center and the probability of using rail transit were negatively correlated in both 2018 and 2023, meaning the greater the distance, the lower the probability of travel. However, the impact of the distance from workplace to the city center changed significantly: while it remained negatively correlated overall in 2018, it peaked at multiple distance points in 2023, indicating that people were willing to choose rail transit even if their workplace was far from the city center. In the SHAP analysis, the importance of this variable dropped from third to eighth, further confirming its weakening influence. This shift reflects that with the improvement of the rail network and the flattening of urban spatial structures, distance to the city center is no longer the key factor determining travel mode, and residents' travel behavior patterns have undergone profound changes.

[0104] (6) Distance from residence (workplace) to secondary center like Figure 14 and Figure 15 As shown, the distance from the residence and workplace to the secondary city center has different effects on the probability of taking rail transit. Figure 14 The data shows that the distance from "residence to the secondary city center" had no significant impact in 2023, while in 2018, it showed a significant negative impact when the distance exceeded 10 kilometers. This is consistent with the result that this variable dropped significantly from 2nd to 10th place in the SHAP importance ranking. Figure 15 This indicates that the distance from the workplace to the secondary city center remained a significant influencing factor throughout the two years, showing a negative correlation. This negative correlation became particularly pronounced in 2023 when the distance exceeded 2 kilometers. Overall, the improved rail transit network has boosted the development of the city's sub-center, shifting the focus of residents' travel decisions from primarily being influenced by the distance to the city center to now being more significantly affected by the distance to the secondary city center.

[0105] This embodiment focuses on understanding how residents adjust their rail transit travel choices to cope with the expansion of the rail transit network. Taking a certain city as an example, based on resident travel survey data from 2018 and 2023, the study uses the Gradient Boosting Decision Tree (GBDT) model to deeply analyze the nonlinear influence and dynamic changes of personal attributes, traffic characteristics, and built environment factors on the probability of residents' rail transit travel choices under different stages of rail transit network expansion. The model results are interpreted from multiple dimensions using SHAP values ​​and Partial Dependency Graphs (PDPs). The following core achievements are made in terms of theoretical and methodological innovation and practical application: (1) This embodiment adopts the gradient boosting decision tree (GBDT) model, combined with multi-objective Bayesian parameter tuning and five-fold cross-validation, to explore in depth the nonlinear correlation between personal attribute characteristics, traffic attribute characteristics, and architectural environment characteristics and the rail transit travel behavior of residents in a certain city. It explores the nonlinear relationship and threshold effect of architectural environment factors on the probability of rail transit travel. It analyzes the dynamic changes in residents' rail transit travel behavior brought about by the expansion of the rail transit network.

[0106] (2) The study, through the ranking analysis of the relative importance of various characteristic variables obtained using SHAP, found that from 2018 to 2023, the gradual expansion of the rail transit network significantly changed the main factors influencing residents' rail transit travel. In the initial stage, building environment characteristics dominated (accounting for 59.80%), while in the later stage, the influence of traffic characteristics increased significantly (accounting for 59.45%). Commuting distance became the primary factor influencing residents' rail transit travel choices. At the same time, residents' demand for efficient connections for the "first mile" and "last mile" is also increasing.

[0107] (3) This embodiment uses a partial dependency graph (PDP) to analyze the interpretability of the model results. Through analysis of the partial dependency graphs of the two variables, "distance to the city center" and "distance to secondary centers," the study found that at the current stage, even if the workplace is far from the city center, residents are more inclined to choose rail transit. The expansion of the rail transit network has significantly improved rail transit accessibility, not only changing residents' travel mode choices but also influencing the city's transportation pattern and development model. This has led to a more dispersed urban structure, a flatter urban layout, and has also promoted the development of sub-centers.

[0108] The research methods and findings can provide a scientific basis for urban master planning and the development of rail transit networks, helping urban planners to better grasp the trends in residents' rail transit travel behavior and formulate more reasonable transportation and planning schemes.

[0109] Example 2 This embodiment discloses a resident travel choice prediction system based on the dynamic evolution of rail transit networks.

[0110] A predictive system for residents' travel choices based on the dynamic evolution of rail transit networks includes: The data acquisition module is configured to: divide multiple influencing factors and acquire multi-source data on residents' travel at different stages of the dynamic evolution of the rail transit network; The preprocessing module is configured to preprocess multi-source data on residents' travel and divide it into training and test sets. The parameter tuning module is configured to: construct a gradient boosting decision tree model, train the model using the training set, and optimize the model parameters using the GBDT hyperparameter automatic tuning method based on multi-objective Bayesian optimization. The data analysis module is configured to: use the trained model to obtain prediction results on the test set, calculate the SHAP value and partial dependency graph of each influencing factor, and obtain the nonlinear action mechanism and dynamic evolution law of each influencing factor.

[0111] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0112] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in the method for predicting resident travel choices in the dynamic evolution of rail transit networks as described in Embodiment 1 of this disclosure.

[0113] Example 4 The purpose of this embodiment is to provide an electronic device.

[0114] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the method for predicting resident travel choices in the dynamic evolution of rail transit networks as described in Embodiment 1 of this disclosure.

[0115] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0116] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0117] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting resident trip choices in dynamic evolution of rail transit network, characterized in that, Includes the following steps: Multiple influencing factors were identified to obtain multi-source data on residents' travel during different periods of the dynamic evolution of the rail transit network; Preprocess the multi-source data on residents' travel to divide it into training and test sets; A gradient boosting decision tree model is constructed, and the model is trained using the training set. The model parameters are optimized using the GBDT hyperparameter automatic tuning method based on multi-objective Bayesian optimization. The trained model is used to obtain prediction results on the test set. The SHAP value and partial dependency graph of each influencing factor are calculated to obtain the nonlinear action mechanism and dynamic evolution law of each influencing factor.

2. The method of predicting resident travel choices for dynamic evolution of rail transit networks of claim 1, wherein, Multiple influencing factors were identified, and multi-source data on resident travel at different stages of the dynamic evolution of the rail transit network were obtained, specifically including: Identify the target city and the corresponding survey years for different periods of the dynamic evolution of the target city's rail transit network; Determine multi-source data on residents' travel in different survey years, including personal attribute data, traffic attribute data, and built-up area environmental data; Identify multiple influencing factors and determine characteristic variables, specifically including: Personal attribute data: including age, number of cars owned by the family, and number of electric bicycles owned by the family; Transportation attribute data: including commuting distance, distance from residence to the nearest subway station, and distance from workplace to the nearest subway station; Built-up area environmental data: including the distance from residence to city center, distance from workplace to city center, distance from residence to the nearest sub-city center, and distance from workplace to the nearest sub-city center.

3. The method of predicting resident travel choices for dynamic evolution of rail transit networks of claim 2, wherein, Preprocessing of multi-source data on residents' travel includes: Propensity score matching is used to address sample selection bias in cross-period comparative studies. Use SMOTEENN hybrid sampling technique to solve the class imbalance problem; Among these measures, propensity score matching is used to address sample selection bias in cross-period comparative studies, specifically including: Using the year of the survey as the treatment variable and all characteristic variables as covariates, a logistic regression model was constructed to calculate the propensity score for each sample. Based on the propensity score of each sample, nearest neighbor matching is used, and the distance between samples is calculated using the Euclidean distance metric to achieve matching of samples from different survey years.

4. The method of predicting resident travel choices for dynamic evolution of rail transit networks of claim 1, wherein, The GBDT hyperparameter automatic tuning method based on multi-objective Bayesian optimization is used to optimize model parameters, specifically including: The optimization objectives are determined to be minimizing prediction error and minimizing model complexity; Mean squared error is used as a metric for model prediction performance. A complexity evaluation function is constructed by comprehensively considering multiple structural parameters of the GBDT model; Construct a proxy model and acquisition function to achieve intelligent directional search in the parameter space, while optimizing the dual objectives of model prediction accuracy and model complexity, and automatically generating Pareto optimal solution set; Normalize the objective values ​​of each solution in the Pareto optimal solution set and calculate the tradeoff score for each solution; Choose the solution with the highest trade-off score as the final hyperparameter configuration.

5. The method of predicting resident travel choices for dynamic evolution of rail transit networks of claim 4, wherein, The complexity evaluation function is specifically as follows: ； wherein, , , and are weight coefficients, reflecting the relative influence degree of different parameters on the complexity of the model; represents the maximum number of iterations of the weak learner; represents the maximum depth of the decision tree; represents the minimum number of samples required for internal node re-partitioning; represents the minimum number of samples required for leaf node.

6. The method of predicting resident travel choices for dynamic evolution of rail transit networks of claim 1, wherein, The SHAP values ​​and partial dependency plots of each influencing factor are calculated to obtain the nonlinear action mechanism and dynamic evolution law of each influencing factor, specifically including: The SHAP value of each influencing factor quantifies the magnitude and direction of the marginal contribution of the current feature to the current prediction result; By summing the SHAP values ​​of all samples, a global feature importance ranking is obtained, which accurately identifies the key driving factors influencing residents' choices of rail transit. Plot partial dependency graphs of feature variables and compare and analyze the differences in the nonlinear impact of the same variable on the probability of residents choosing rail transit travel at different stages of rail transit network expansion.

7. The method of predicting resident travel choices for dynamic evolution of rail transit networks of claim 4, wherein, During the automatic tuning of GBDT hyperparameters: Initialization: Randomly select a small number of hyperparameter combinations for preliminary evaluation; Constructing a proxy model: Based on existing evaluation results, establish a probabilistic model of the relationship between the objective function and hyperparameters; Select the next evaluation point: Determine the most promising combination of hyperparameters based on the acquisition function; Evaluation and Update: Evaluate the performance of the selected hyperparameter combination and update the surrogate model; Iteration: Repeat the above process until the preset number of evaluations is reached.

8. A resident trip choice prediction system for dynamic evolution of rail transit network, characterized in that, include: The data acquisition module is configured to: divide multiple influencing factors and acquire multi-source data on residents' travel at different stages of the dynamic evolution of the rail transit network; The preprocessing module is configured to preprocess multi-source data on residents' travel and divide it into training and test sets. The parameter tuning module is configured to: construct a gradient boosting decision tree model, train the model using the training set, and optimize the model parameters using the GBDT hyperparameter automatic tuning method based on multi-objective Bayesian optimization. The data analysis module is configured to: use the trained model to obtain prediction results on the test set, calculate the SHAP value and partial dependency graph of each influencing factor, and obtain the nonlinear action mechanism and dynamic evolution law of each influencing factor.

9. A computer-readable storage medium having stored thereon a program, characterized in that, When executed by the processor, the program implements the steps in the method for predicting resident travel choices in the dynamic evolution of rail transit networks as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for predicting resident travel choices in the dynamic evolution of rail transit networks as described in any one of claims 1-7.

Images