Typhoon disaster resident resettlement selection prediction method and device

Abstract

The application discloses a typhoon disaster resident resettlement selection prediction method and device, comprising the following steps: constructing a multi-dimensional population attribute index system and collecting resident resettlement sample data; performing feature screening on the sample data to obtain index features for model input; constructing a plurality of single machine learning models, screening a basic learner based on performance and constructing a soft voting combination prediction model; training and performance evaluating the prediction model; inputting population attribute data of a resident to be predicted into the trained model to output a probability of resettlement selection. The application adopts a feature screening strategy to screen out key features and reduce feature dimensions, can simplify a data collection process, meets a quick decision-making demand in a disaster, and reduces data acquisition difficulty for subsequent practical application.

Description

Technical Field

[0001] This invention relates to the field of urban planning technology, and more specifically, to a method and apparatus for predicting typhoon disaster resident resettlement options. Background Technology

[0002] In typhoon disaster emergency response, timely evacuation and resettlement of affected people are crucial measures to reduce disaster losses and protect lives. However, residents' evacuation choices vary individually, and the decision to choose government-established centralized resettlement sites is influenced by many factors such as age, income, housing conditions, and health status. Accurately predicting the resettlement choices of different residents is of great significance for the rational allocation of evacuation site resources and improving the scientific level of emergency management.

[0003] Existing assessment methods for typhoon evacuation and relocation decisions are mainly divided into two categories: one is mechanism simulation-based assessment methods, which simulate the evacuation and relocation process based on the evolution mechanism of typhoon disasters and through mathematical modeling, statistical analysis, and other means. However, this type of method has high data requirements, focuses on macro-level assessment, and is difficult to combine with micro-level population attributes for individual resettlement decision analysis; the other is data-driven assessment methods, which use mathematical statistics, machine learning, and other techniques to mine variable correlations and extract features from high-dimensional data, and are suitable for micro-level evacuation and resettlement decision assessment.

[0004] However, existing research largely focuses on evacuation and relocation, failing to delve into the precise assessment of resettlement decisions based on micro-level demographic attributes, thus hindering the development of refined emergency resettlement plans. Therefore, a method is urgently needed that can combine demographic characteristics to accurately predict residents' resettlement choices, thereby enhancing the targeting and effectiveness of emergency management. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and device for predicting the resettlement choices of residents during typhoon disasters. By exploring the nonlinear correlation between micro-population attributes and resettlement behavior, it can achieve accurate prediction of individual resettlement choices (centralized resettlement or decentralized resettlement).

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting typhoon disaster resettlement options includes the following steps:

[0008] Step S1: Construct a multi-dimensional population attribute indicator system and collect resident resettlement sample data;

[0009] Step S2: Perform feature filtering on the sample data to obtain indicator features for model input;

[0010] Step S3: Construct multiple single machine learning models, select the base learner based on performance, and construct a soft-voting combined prediction model;

[0011] Step S4: Train and evaluate the performance of the prediction model;

[0012] Step S5: Input the demographic attribute data of the residents to be predicted into the trained model and output the probability of their resettlement choices.

[0013] Furthermore, step S2 includes the following steps:

[0014] Step S21: Preprocess the sample data, taking centralized resettlement and decentralized resettlement as target variables and population attribute indicators as input features;

[0015] Step S22: Using Spearman correlation analysis, calculate the correlation between each input feature and the target variable, and select significant related indicators based on the set significance threshold to form the input indicator set A;

[0016] Step S23: The minimum redundancy maximum correlation algorithm is used to screen features of the sample data. Through a forward stepwise search strategy, features are selected according to the objective function to form the input index set B.

[0017] Step S24: Use input indicator set A and input indicator set B as input features for model construction.

[0018] Furthermore, in step S22, Spearman correlation analysis is performed by calculating the rank difference, specifically by calculating the rank difference between the two variables for each observation and determining the correlation coefficient based on the sample size.

[0019] Furthermore, in step S23, the objective function of the minimum redundancy maximum correlation algorithm is to maximize the correlation between the feature and the target variable, while minimizing the redundancy between features.

[0020] Furthermore, step S3 includes the following steps:

[0021] Step S31: Construct multiple single machine learning models, including logistic regression model, support vector machine model, random forest model, LightGBM model, XGBoost model and CatBoost model;

[0022] Step S32: Based on the performance of each individual machine learning model, select the three best-performing models as the base learners and construct a soft voting combined model.

[0023] Step S33: Incorporate the weight selection process into the model training framework and automatically determine the optimal weight combination through grid search.

[0024] Furthermore, in step S32, the soft voting combination model obtains the final prediction probability by weighted averaging of the class probabilities output by each basic learner, and determines the final prediction class according to the set discrimination threshold.

[0025] Furthermore, step S4 includes the following steps:

[0026] Step S41: The prediction model is trained and evaluated using a hierarchical five-layer nested cross-validation framework. This framework includes an outer loop and an inner loop. The outer loop is used to evaluate the model performance, and the inner loop is used for parameter optimization.

[0027] Step S42: Divide the original dataset into five mutually exclusive subsets of similar size and consistent class proportions using stratified sampling. In the outer cross-validation, select one subset as the test set and the other four subsets as the training set.

[0028] Step S43: Construct an inner five-fold cross-validation based on the training set, perform hyperparameter optimization, and use the evaluation index obtained from the inner cross-validation as the optimization objective to determine the optimal parameter combination.

[0029] Step S44: After obtaining the optimal parameter combination, the model is refitted on the complete training set using the optimal parameter combination in the outer layer to obtain the final model under the current fold.

[0030] Step S45: Perform predictions on the independent test set using the final model to obtain the performance evaluation results of this outer layer cross-validation.

[0031] Step S46: Repeat steps S42 to S45 until all five iterations of the outer cross-validation are completed, and obtain the set of independent performance indicators for each prediction model under five different data partitions.

[0032] Step S47: For the combined model, weight optimization is performed simultaneously in the nested cross-validation framework. For each candidate weight combination, the evaluation index is calculated in the inner validation set, and the weight combination that makes the inner evaluation index optimal is selected as the optimal weight for the current step.

[0033] Step S48: Use multiple evaluation metrics to evaluate the performance of the prediction model.

[0034] Furthermore, in step S48, the evaluation metrics include AUC, recall, accuracy, precision, and F1 score.

[0035] Furthermore, in step S5, outputting the resettlement selection probability includes outputting the probability value of whether the resident chooses centralized resettlement or decentralized resettlement.

[0036] The present invention also provides a device for predicting typhoon disaster resident resettlement options, comprising:

[0037] Memory, used to store computer programs and data;

[0038] The processor is used to execute computer programs to implement the steps of the above-mentioned method for predicting the resettlement of residents affected by typhoon disasters.

[0039] The beneficial effects of this invention are:

[0040] 1. This invention addresses high-dimensional, small-sample scenarios by employing a feature selection strategy to identify key features, reduce feature dimensionality, streamline the data collection process, meet the needs of rapid decision-making during disasters, and reduce the difficulty of data acquisition for subsequent practical applications.

[0041] 2. This invention strictly separates hyperparameter tuning and performance evaluation through layered five-layer nested cross-validation to avoid data leakage. The optimal model soft voting combination model achieves an AUC of 0.930 and a recall of 0.892 on the test set, demonstrating stable performance and reliable generalization ability.

[0042] 3. In this invention, the key features selected are easily obtained through questionnaires, which can simplify the data collection process; it meets the needs of rapid decision-making during disasters and can be transferred to the prediction of resident resettlement for other types of disasters such as earthquakes and floods.

[0043] 4. This invention explores the application of machine learning models in predicting resident refuge and resettlement. By collecting data from field questionnaires, it achieves prediction of resident refuge and resettlement. Attached Figure Description

[0044] Figure 1 This is a flowchart of a typhoon disaster resident resettlement selection prediction method in this embodiment;

[0045] Figure 2 This is a graph showing the correlation coefficients between the feature variables and the choice of refuge and resettlement in this embodiment;

[0046] Figure 3 This is a graph showing the trend of AUC changes under different K values ​​in this embodiment;

[0047] Figure 4 This is a framework diagram of a five-layer nested cross-validation training method in this embodiment;

[0048] Figure 5a This is a graph showing the mean AUC and its trend for each model in this embodiment under feature subset A and feature subset B;

[0049] Figure 5b Box plots of AUC for each model in this embodiment under two feature subsets;

[0050] Figure 6aThis is a diagram showing the predicted refuge and resettlement results for sample a in this embodiment;

[0051] Figure 6b This is a diagram showing the predicted refuge and resettlement results for sample b in this embodiment;

[0052] Figure 6c This is a diagram showing the predicted refuge and resettlement results for sample c in this embodiment;

[0053] Figure 6d This is a diagram showing the predicted refuge and resettlement results for sample d in this embodiment;

[0054] Figure 6e This is a diagram showing the predicted refuge and resettlement results for sample e in this embodiment;

[0055] Figure 6f This is a diagram showing the predicted refuge and resettlement results for sample f in this embodiment;

[0056] Figure 7 This is a structural block diagram of a typhoon disaster resident resettlement selection prediction device in this embodiment.

[0057] Reference numerals: Memory 1, Processor 2. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Example: A method for predicting resident resettlement options during typhoon disasters, such as... Figure 1 As shown, it includes the following steps:

[0060] Step S1: Construct a multi-dimensional population attribute indicator system and collect resident resettlement sample data.

[0061] Specifically, through literature review and field research, a population attribute indicator system for refugee resettlement was constructed, which includes:

[0062] Demographic characteristics are used to characterize differences in basic individual attributes, including gender, age, occupation, education level, and household registration.

[0063] Socioeconomic conditions are used to depict family resources and living conditions, including private car ownership, housing conditions, building age, and social security status.

[0064] Family and social network status, used to reflect support and care pressure, including the status of relatives and friends near the shelter, living alone, local immediate family members, and whether there are other elderly people or children living with the individual besides themselves;

[0065] Disaster preparedness and response capabilities focus on information acquisition and response capabilities, including participation in disaster prevention and evacuation training or drills, and having two or more information acquisition and communication channels.

[0066] The characteristics of vulnerable groups are used to identify differences in mobility and assistance needs, including people with disabilities, those on minimum living allowances, and those with independent mobility.

[0067] Geographical and historical factors are used to reflect the impact of residential environment risk exposure and past resettlement experience, including residential area, geographical environment, whether there are rivers or lakes near the residence, distance from refuge sites, historical resettlement situation, and risk sources.

[0068] The study focused on areas with high typhoon landfall frequency and typical disaster characteristics, encompassing various evacuation and resettlement sites. Multi-dimensional data on residents in centralized shelters were collected through questionnaires, while sample data from dispersed residents were also collected simultaneously, ensuring the sample covered disaster-affected groups with diverse geographical environments and demographics. All indicators were converted to numerical data according to a standardized coding format for subsequent analysis.

[0069] Step S2: Perform feature filtering on the sample data to obtain indicator features for model input.

[0070] Specifically, it includes the following steps:

[0071] Step S21: Preprocess the sample data collected in step S1, taking centralized resettlement and decentralized resettlement as target variables and population attribute indicators as input features.

[0072] Step S22: Using Spearman correlation analysis, calculate the correlation between each input feature and the target variable. The formula is as follows:

[0073]

[0074] In the formula, The Spearman correlation coefficient; Let be the difference in rank between the two variables in the i-th sample; This represents the number of samples.

[0075] Then, based on the set significance threshold ( The characteristic variables that are significantly related to refuge and resettlement decisions are selected to form the input indicator set A.

[0076] Step S23: The minimum redundancy maximum correlation (mRMR) algorithm is used to screen features of the sample data. Through a forward stepwise search strategy, features are selected according to the objective function to form the input index set B. The core idea of ​​this algorithm is to maximize the correlation between features and the target variable (refuge and resettlement selection) while minimizing the redundancy between features.

[0077] The objective function for minimizing redundancy and maximizing correlation is:

[0078]

[0079] In the formula, The set of candidate features; The number of features selected; Representation of features With target variable Mutual information between them; Representation of features With features Mutual information between them; and The feature index represents different feature variables.

[0080] Step S24: Input indicator set A and input indicator set B are used together as input features for model construction to compare and analyze the impact of different feature selection strategies on model performance.

[0081] Step S3: Construct multiple individual machine learning models, select the base learner based on performance, and construct a soft-voting combined prediction model.

[0082] Specifically, it includes the following steps:

[0083] Step S31: Construct multiple single machine learning models, including a logistic regression model, a support vector machine model, a random forest model, a LightGBM model, an XGBoost model, and a CatBoost model.

[0084] Step S32: After completing the training, tuning, and performance evaluation of the above six models, the three best-performing models are selected as the base learners based on the overall performance of the models, and a soft voting combination model is constructed.

[0085] The soft-voting combination model obtains the final prediction probability by weighted averaging of the class probabilities output by each basic learner, and determines the final prediction class based on a set discrimination threshold.

[0086] Let the three basic learners selected be as follows: , and For any sample The predicted probabilities of the output categories (centralized) of each basic learner are as follows: , and The soft-voting combination model calculates the predicted probability by weighting the class probabilities output by each base learner:

[0087]

[0088] In the formula, The probability of concentrated resettlement predicted by the combined model; For the first The weights of each basic learner represent the relative contribution of that basic learner to the overall decision. For the first The settling probability output by each basic learner.

[0089] In the classification decision-making stage, a discrimination threshold is set. The final prediction category of the combined model is:

[0090]

[0091] In the formula, For the final predicted category; To determine the threshold, it is usually set to 0.5.

[0092] Step S33: To avoid bias caused by subjective weight setting, the weight selection process is incorporated into the model training framework, and the optimal weight combination is automatically determined through grid search.

[0093] Considering the limited sample size and computational cost constraints, a discrete weight candidate set is used to achieve a controllable search scale. Set as: Then the weight candidate space of the three-model combination can be expressed as: ,Right now Each component is from the weighted candidate set Take the value from the middle.

[0094] Step S4: Train and evaluate the performance of the prediction model.

[0095] Specifically, such as Figure 4 As shown, it includes the following steps:

[0096] Step S41 employs a hierarchical five-layer nested cross-validation framework to train and evaluate the prediction model. This framework consists of an outer loop and an inner loop; the outer loop evaluates the model, while the inner loop optimizes the parameters. By combining this with grid search, the parameter selection and performance evaluation processes are made independent of each other, preventing data leakage.

[0097] Step S42: Divide the original dataset into five mutually exclusive subsets of similar size and consistent class proportions using stratified sampling. In the outer cross-validation, select one subset as the test set and the other four subsets as the training set.

[0098] Let the original sample dataset be:

[0099]

[0100] in, This represents the feature vector of the i-th sample. This indicates the corresponding refuge / relocation option category. .

[0101] By using stratified sampling, the entire original dataset Divided into 5 mutually exclusive subsets of similar size and consistent category proportions , , , and and satisfies the following properties:

[0102] , ,

[0103] in, Indicates will , , , and The union of these 5 subsets; This represents the intersection of any two distinct subsets (the i-th and j-th subsets).

[0104] In the outer layer cross-validation Second-rate( In the iteration, select the first... One subset is used as the test set, and the remaining four subsets are used as the training set:

[0105] ,

[0106] In the formula, For the first The test set for the next iteration; For the first The training set and test set for the next iteration. This data will be retained as independent validation data and used only for the final evaluation in step S4. It will not be used in any parameter optimization or model training process in subsequent inner loops to prevent information leakage.

[0107] Step S43: Construct inner five-fold cross-validation based on the training set partitioned in step S42, and perform hyperparameter optimization. For the combination of hyperparameters of the model, For the hyperparameter space of the model, a grid search is used to traverse the hyperparameter space in the inner loop. The optimal parameter combination is determined by using the evaluation index obtained from inner-layer cross-validation as the optimization objective. :

[0108]

[0109] In the formula, Represents the predictive model; Indicates the inner verification set; This is the model performance evaluation function (AUC is used as the evaluation metric in the inner cross-validation). This represents the value of the independent variable that maximizes the objective function, i.e., the optimal combination of parameters in the hyperparameter space that optimizes the performance evaluation index.

[0110] Step S44: Obtain the optimal parameter combination Then, exit the inner loop and utilize the optimal parameter combination in the outer loop. In the complete training set The model is refitted to obtain the final model that has been learned most fully at the current fold. .

[0111] Step S45, the trained final model The independent test set retained in step S42 The above is used for prediction, and the first is obtained. Performance evaluation results of the second outermost cross-validation.

[0112] Due to independent test sets The parameter selection in step S42 and the model training process in step S43 are completely invisible. The performance evaluation results obtained in this step can unbiasedly reflect the model's true predictive ability for unknown data.

[0113] Step S46: Repeat steps S42 to S45 until all 5 iterations of the outer cross-validation are completed, thus obtaining a set of independent performance metrics for each prediction model under 5 different data partitions:

[0114]

[0115] Finally, the mean of this set is used as the evaluation result of the overall predictive performance of the model, and the standard deviation is used to measure the stability of the model's predictive performance.

[0116] Step S47: After completing nested cross-validation training of the six single models, the three best-performing models are selected as the base learners for the combined model based on their overall performance. To ensure fairness and rigor in the comparison of the combined models, the combined model is also trained and evaluated within a nested cross-validation framework of outer evaluation and inner optimization. Since different base models differ in predictive ability and error type, the weight configuration directly affects the final performance of the combined model. To avoid subjective bias caused by manually setting weights and to ensure that the weight selection process does not introduce information leakage, the determination of the combined weights is also performed within the nested cross-validation framework. The discrete candidate value set for the weights is set as follows: The candidate weight space for the combination of the three models is:

[0117]

[0118] For each candidate weight combination, the soft-voting weighted probability is calculated using the prediction probability formula in step S32, and the corresponding evaluation index is obtained in the inner layer validation set. Finally, the weight combination that optimizes the inner layer evaluation index is selected as the outer layer's [number missing]. Optimal weights after the break:

[0119]

[0120] In the formula, This represents a soft-voting combination model with given weights. For the inner verification set; The evaluation function is AUC.

[0121] To obtain the optimal combination weights of the basic learners Then, in the outer training set The combined model was refitted with these weights, and finally tested on the independent test set. Prediction and performance evaluation are performed on the above. Repeat. The outer iterations are used to obtain the performance set of the combined model in five outer tests.

[0122] Step S48: The performance of the prediction model is evaluated using multiple evaluation metrics, including AUC and recall. ), accuracy ( ), accuracy ( ) and F1 score ( The calculation formulas for each indicator are as follows:

[0123]

[0124]

[0125]

[0126]

[0127] In the formula, True cases (the number of correctly predicted positive cases); The number of true negative examples (the number of correctly predicted negative examples); The number of false positives (the number of incorrectly predicted positive examples); This is a false negative (the number of cases incorrectly predicted as negative).

[0128] Step S5: Input the demographic attribute data of the resident to be predicted into the trained model, and output the probability of the resident choosing centralized or decentralized resettlement. This probability value can be used as a reference for decision-making by the emergency management department.

[0129] This embodiment also provides a device for predicting typhoon disaster resident resettlement options, such as... Figure 7 As shown, it includes a memory 1 and a processor 2; wherein, the memory 1 is used to store computer programs and data; the processor 2 is used to implement the steps of the above-mentioned typhoon disaster resident resettlement selection prediction method when executing the computer program.

[0130] To verify the effectiveness and superiority of the method provided in this embodiment, experimental verification was conducted based on field survey data from the disaster-stricken area, as follows:

[0131] This study employs a literature review approach to summarize domestic and international research findings on the influencing factors of disaster avoidance, evacuation behavior, and resettlement decisions. Based on this, the influencing factors are summarized into six dimensions: demographic characteristics, socioeconomic status, family and social network status, disaster preparedness and response capacity, vulnerable group attributes, and geographical and historical factors. Table 1 provides an explanation of each dimension and its meaning.

[0132] Table 1 Summary of factors related to population refuge selection

[0133]

[0134] Within the aforementioned six-dimensional framework, to achieve the transformation from "theoretical factors to measurable variables," this method further selects operable and representative indicators to form a set of key indicators suitable for questionnaire surveys and subsequent modeling (see Table 2). Among them, demographic characteristics are used to characterize differences in basic individual attributes; socioeconomic status depicts family resources and living conditions; family and social networks reflect support systems and caregiving pressures; disaster preparedness and response capabilities focus on information acquisition and coping abilities; the vulnerable group dimension is used to identify differences in action capabilities and relief needs; and geographical and historical factors reflect the impact of residential environment risk exposure and past resettlement experiences.

[0135] Table 2. Set of Indicators for Population Refuge and Resettlement Selection

[0136]

[0137] The final indicator system and coding specifications resulting from this method are shown in Tables 3.1 and 3.2. The indicator system includes a target variable and six types of explanatory variables, totaling 24 variables, providing a unified data structure foundation for subsequent statistical testing and predictive model input.

[0138] Table 3.1 Indicator System Coding Table

[0139]

[0140] Table 3.2 Indicator System Coding Table

[0141]

[0142] This method combines Spearman correlation analysis and minimum redundancy maximum correlation (mRMR) algorithm to select features from two levels: statistical significance and prediction guidance. It constructs feature subsets for each level, providing a foundation for subsequent prediction model construction and comparative analysis.

[0143] Spearman correlation analysis results:

[0144] Calculate the correlation coefficients and significance of each feature with the target variable. The correlation analysis ranking results are as follows: Figure 2 As shown, filter out There are 10 features in total, which constitute feature subset A, as shown in Table 4.

[0145] Table 4 Feature subset A based on Spearman correlation analysis

[0146]

[0147] Note: express , express

[0148] Results of the minimum redundancy maximum correlation algorithm screening:

[0149] Based on the ranking results of the minimum redundancy and maximum relevance features, the top [number] features are selected in sequence. A logistic regression model was constructed using 10 features, and 5-fold cross-validation was used to evaluate the model performance. The AUC metric was selected as the evaluation metric. Model performance was assessed for different numbers of features. Figure 3 As shown in Table 5, seven feature variables were selected to construct feature subset B.

[0150] Table 5 Feature subset B based on the minimum redundancy maximum correlation algorithm

[0151]

[0152] Based on a five-fold nested cross-validation training framework, a systematic evaluation was conducted on six single models—logistic regression, support vector machine, random forest, LightGBM, XGBoost, and CatBoost—as well as a combined model based on soft voting under different feature subset conditions. All performance results were derived from the outer test set of the nested cross-validation, ensuring that the model hyperparameter selection, combined weight optimization, and final generalization performance evaluation processes were independent, thus avoiding data leakage. Model evaluation focused on overall discriminative ability (AUC), minority class identification ability (Recall), and overall balance ability (F1-score). Simultaneously, standard deviation analysis was used to analyze the model's stability under different data partitions, aiming to select the optimal model structure and feature subset combination more suitable for predicting resident evacuation and resettlement choices in typhoon disaster scenarios.

[0153] To verify the effectiveness of feature engineering, two types of feature inputs were constructed: a feature subset A selected based on correlation analysis, and a feature subset B selected based on the minimum redundancy maximum correlation algorithm. Under a unified five-fold nested cross-validation framework, six single-model and combined models were trained and evaluated. The outer-layer validation results are summarized in Table 6. Figure 5a Used to compare the mean AUC and its changing trend of each model under feature subset A and feature subset B; Figure 5b Box plots were used to compare the distribution characteristics and dispersion of AUC for each model under two feature subsets to reflect the stability of model performance.

[0154] Table 6. Comparison of prediction performance between different feature subsets and models

[0155]

[0156] Note: Recall is the recall rate for a minority class (centralized placement).

[0157] Considering both model performance and algorithmic mechanism characteristics, Random Forest (RF), LightGBM, and CatBoost were ultimately selected as the base learners for the soft voting ensemble model. Using these three base models to construct the soft voting ensemble model, predictions were made on samples. The overall performance under both feature subsets was superior to a single model, especially under feature subset B, where AUC, Recall, and F1-score all reached optimal levels. Six samples were selected as representative samples, such as... Figures 6a-6f As shown, the waterfall plot represents the predicted value of a single sample as the baseline value. The result is obtained by summing the SHAP (SHapley Additive ex Planations) values ​​of each feature. Red bars indicate that the feature drives the prediction towards "centralized resettlement," while blue bars indicate that it drives it towards "decentralized resettlement." This summation yields the final individual predicted value. .

[0158] Samples a, b, and c are predicted dispersed placement samples, from Figures 6a-6c As can be seen, the predicted values ​​for this type of sample are generally lower than the baseline (0.178), with multiple features contributing negatively. Figure 6c For example, the predicted value for this sample was 0.116, significantly lower than the baseline. The main negative factors contributing to this were age (x2=43), historical resettlement experience (x23=0), and the presence of family and friends (x10=1). Since this resident is in their prime, has not participated in centralized resettlement, and has a support network of family and friends near the resettlement site, all three factors negatively impacted the predicted result. Driven by these multiple factors, the model is more inclined to predict that the resident will choose decentralized resettlement. Figure 6a and Figure 6b Although there are differences in the specific contribution magnitude, the overall structure is similar: sufficient social support, lack of experience in centralized resettlement, and relatively strong action capabilities constitute the common characteristics of the dispersed resettlement prediction.

[0159] Samples d, e, and f are predicted clustered placement samples; the predicted values ​​for these samples are significantly higher than the baseline values. For example... Figure 6e In the preliminary results, the predicted value reached 0.933, with multiple features contributing positively. Historical resettlement experience (x23=1) and age (x2 - advanced age) were particularly significant, each contributing a large positive SHAP value and becoming the main factors driving the predicted result of centralized resettlement. The building's construction year (x8=2) and geographical environment (x20=1) also had some positive impact. In some centralized resettlement samples, the situation of relatives and friends (x10=1) still showed a negative contribution, but its influence was offset by advanced age or previous centralized resettlement experience. Ultimately, the overall trend still pointed to centralized resettlement, indicating that in cases of advanced age and prior refugee resettlement experience, support from relatives and friends did not prompt them to choose dispersed resettlement; instead, they might support them to go to centralized resettlement sites with better facilities for safety reasons. This suggests that the final judgment is formed by the combined effect of multiple factors, rather than being determined by a single variable.

[0160] These partial explanations reveal significant differences in the model's decision-making paths across individuals. Even when certain variables have the same values, such as the presence of family and friends support, their impact on the prediction direction varies depending on factors like age and historical resettlement experience. The model does not classify based on a single rule but rather derives its results from the combined effects of multiple variables, further demonstrating that the predictive model can more meticulously characterize individual differences in residents' resettlement decisions and accurately predict residents' resettlement needs.

[0161] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting resident resettlement options during typhoon disasters, characterized in that, Includes the following steps: Step S1: Construct a multi-dimensional population attribute indicator system and collect resident resettlement sample data; Step S2: Perform feature filtering on the sample data to obtain indicator features for model input; Step S3: Construct multiple single machine learning models, select the base learner based on performance, and construct a soft-voting combined prediction model; Step S4: Train and evaluate the performance of the prediction model; Step S5: Input the demographic attribute data of the residents to be predicted into the trained model and output the probability of their resettlement choices.

2. The method for predicting resident resettlement options during typhoon disasters according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Preprocess the sample data, taking centralized resettlement and decentralized resettlement as target variables and population attribute indicators as input features; Step S22: Using Spearman correlation analysis, calculate the correlation between each input feature and the target variable, and select significant related indicators based on the set significance threshold to form the input indicator set A; Step S23: The minimum redundancy maximum correlation algorithm is used to screen features of the sample data. Through a forward stepwise search strategy, features are selected according to the objective function to form the input index set B. Step S24: Use input indicator set A and input indicator set B as input features for model construction.

3. The method for predicting resident resettlement options during typhoon disasters according to claim 2, characterized in that, In step S22, Spearman correlation analysis is performed by calculating the rank difference, specifically by calculating the rank difference between the two variables for each observation and determining the correlation coefficient based on the sample size.

4. The method for predicting typhoon disaster resident resettlement options according to claim 2, characterized in that, In step S23, the objective function of the minimum redundancy maximum correlation algorithm is to maximize the correlation between the feature and the target variable, while minimizing the redundancy between features.

5. The method for predicting typhoon disaster resident resettlement options according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Construct multiple single machine learning models, including logistic regression model, support vector machine model, random forest model, LightGBM model, XGBoost model and CatBoost model; Step S32: Based on the performance of each individual machine learning model, select the three best-performing models as the base learners and construct a soft voting combined model. Step S33: Incorporate the weight selection process into the model training framework and automatically determine the optimal weight combination through grid search.

6. The method for predicting typhoon disaster resident resettlement options according to claim 5, characterized in that, In step S32, the soft voting combination model obtains the final prediction probability by weighted averaging of the class probabilities output by each basic learner, and determines the final prediction class according to the set discrimination threshold.

7. The method for predicting typhoon disaster resident resettlement options according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: The prediction model is trained and evaluated using a hierarchical five-layer nested cross-validation framework. This framework includes an outer loop and an inner loop. The outer loop is used to evaluate the model performance, and the inner loop is used for parameter optimization. Step S42: Divide the original dataset into five mutually exclusive subsets of similar size and consistent class proportions using stratified sampling. In the outer cross-validation, select one subset as the test set and the other four subsets as the training set. Step S43: Construct an inner five-fold cross-validation based on the training set, perform hyperparameter optimization, and use the evaluation index obtained from the inner cross-validation as the optimization objective to determine the optimal parameter combination. Step S44: After obtaining the optimal parameter combination, the model is refitted on the complete training set using the optimal parameter combination in the outer layer to obtain the final model under the current fold. Step S45: Perform predictions on the independent test set using the final model to obtain the performance evaluation results of this outer layer cross-validation. Step S46: Repeat steps S42 to S45 until all five iterations of the outer cross-validation are completed, and obtain the set of independent performance indicators for each prediction model under five different data partitions. Step S47: For the combined model, weight optimization is performed simultaneously in the nested cross-validation framework. For each candidate weight combination, the evaluation index is calculated in the inner validation set, and the weight combination that makes the inner evaluation index optimal is selected as the optimal weight for the current step. Step S48: Use multiple evaluation metrics to evaluate the performance of the prediction model.

8. The method for predicting typhoon disaster resident resettlement options according to claim 7, characterized in that, In step S48, the evaluation metrics include AUC, recall, accuracy, precision, and F1 score.

9. The method for predicting typhoon disaster resident resettlement options according to claim 1, characterized in that, In step S5, the output of the resettlement selection probability includes the output of the probability value of whether the resident chooses centralized resettlement or decentralized resettlement.

10. A device for predicting resident resettlement options during typhoon disasters, characterized in that, include: Memory (1), used to store computer programs and data; The processor (2) is used to execute a computer program to implement the steps of the typhoon disaster resident resettlement selection prediction method as described in any one of claims 1-9.

Images