SaaS-based multi-platform online car-hailing driver retention prediction method

Abstract

The invention relates to the field of data management, and discloses a SaaS-based multi-platform online car-hailing driver retention prediction method, which is used for solving the problem that a platform cannot timely discover and take corresponding retention measures when a driver with a high user loss risk appears on the platform, and comprises the following steps: collecting the data of the driver on each platform; the method comprises the steps of preprocessing collected data, performing feature extraction on the preprocessed data, selecting a machine learning model to train the model, selecting an optimal model according to a training result, evaluating the reliability of the model along with the change of data volume, predicting user loss risks, and performing early warning on crowds with high user loss risk levels. According to the method and the system, the user loss risk level is high, operation crowd package intelligent circle selection is performed on crowds with high user loss risk levels, the users are classified and analyzed, corresponding operation strategies and measures are formulated, drivers with high loss risks are accurately screened out, and the driver resource stability of the platform is improved.

Description

Technical Field

[0001] The present invention relates to the field of data management, and more specifically to a SaaS-based multi-platform online car-hailing driver retention prediction method. Background Art

[0002] Online ride-hailing driver retention refers to the situation in which drivers who register and start serving on an online ride-hailing platform continue to use the platform over a period of time. In the online ride-hailing industry, driver retention is crucial to the long-term stable operation of the platform. Driver retention usually indicates the driver's loyalty and satisfaction with the platform, as well as their willingness to continue to serve the platform. A higher driver retention rate means that the platform can maintain a stable supply, provide better services and experience, and maintain a dominant position in market competition.

[0003] The existing online car-hailing driver retention prediction mainly relies on the Internet product user prediction prototype. It relies on user viewing, usage and consumption behaviors, and predicts retention based on user habit data. The feature dimension is single, and the model mainly relies on a few indicators in user behavior data for prediction. The model may not cover too many different types of features, but focuses on the most representative or most influential features for prediction.

[0004] In view of the above problems, the present invention proposes a solution. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a SaaS-based multi-platform online car-hailing driver retention prediction method to solve the problems existing in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A SaaS-based multi-platform online car-hailing driver retention prediction method includes the following steps:

[0008] S001: Collecting the driver's data on various platforms, including the driver's registration and login data on various platforms, driver attribute data, vehicle dispatch behavior data, reward data, and violation penalty data;

[0009] S002: Preprocessing the collected data, including cleaning outliers, supplementing missing values, and normalizing the data;

[0010] S003: extracting features from the preprocessed data, wherein the features include driver attribute features, driver behavior features, and driver cross-platform switching behaviors;

[0011] S004: Select a machine learning model, train the model using historical data, and select the final model based on the training results;

[0012] S005: Use the learning curve to evaluate the reliability of the final model as the amount of data changes, and iteratively optimize the model with low reliability. After optimization, continue to evaluate the reliability until the model reliability meets the standard.

[0013] S006: Use a model that has passed the reliability test to predict the risk of user churn, issue early warnings to people with high risk of user churn, and intelligently circle the people with high risk of user churn for operational populations, classify and analyze these users, identify user groups who will continue to use the platform, and formulate corresponding operational strategies and measures.

[0014] Preferably, the step of preprocessing the collected data is:

[0015] Delete duplicate records in the data set to ensure data uniqueness and perform outlier detection and processing;

[0016] Find missing values ​​in the data set, delete missing values, use mean and median to fill missing values ​​for numerical data, and use mode to fill missing values ​​for categorical data;

[0017] Use Min-Max normalization to scale the data to a specified range. The Min-Max normalization formula is: Transform the data into a distribution with mean 0 and standard deviation 1, where X norm Expressed as standardized values, X is the original data, X min and X max Expressed as the minimum and maximum values ​​of the data.

[0018] Preferably, the steps of selecting a machine learning model, training the model using historical data, and selecting a final use model according to the training results are:

[0019] Define the evaluation basis of the model;

[0020] Collect historical data and divide it into training set and test set;

[0021] Try and screen classification models;

[0022] Use GridSearchCV to perform model cross-validation tuning.

[0023] Preferably, the step of evaluating the reliability of the final use model as the amount of data changes through the learning curve is:

[0024] The cross-validation method is used to divide the data set into training set and validation set;

[0025] Gradually increase the amount of data in the training set, starting with a small subset and then gradually increasing the amount of data;

[0026] Perform model training on training sets of each data size, and verify on verification sets of the same size, and record the performance indicators of training and verification, including loss function, accuracy, and recall;

[0027] The model reliability index is obtained by comprehensive evaluation of the loss function, accuracy and recall rate, and its calculation formula is: Where RB represents the model reliability index, LS represents the loss function, PC represents the accuracy, RC represents the recall rate, a1, a2, and a3 represent the weight coefficients of the loss function, accuracy, and recall rate, and the sum of a1, a2, and a3 is 1.

[0028] Preferably, the loss function calculation step is:

[0029] Choose a loss function that is suitable for the task. Common ones include mean square error, cross entropy loss function, and logarithmic loss function.

[0030] Use the feature data of the training set to input the mean square error model to obtain the prediction results of the model;

[0031] Compare the model's predictions with the true labels of the training set and calculate the loss value;

[0032] Sum the loss values ​​of all samples and find their average to get the average loss of the entire training set, which is the loss function value. The calculation formula is: in represents the loss value of sample i, n represents the number of samples, y i is the true label of sample i, It is represented as the predicted value of the model for sample i.

[0033] Preferably, the step of predicting the user churn risk using the model with passed reliability is:

[0034] The service activity change coefficient is calculated based on the change in service activity;

[0035] Collect the driver's total income on this platform and on the switched platform. The total income is the sum of the average daily income, average order price, rewards and subsidies. The income change coefficient is calculated based on the driver's total income on this platform and on the switched platform.

[0036] Collect the number of violations that the driver committed on this platform and on the switching platform in the same time period, and calculate the violation change coefficient;

[0037] Collect the number of positive service reviews that drivers received on this platform and the number of positive service reviews that they received on the switched platform during the same period of time, and calculate the coefficient of change in positive reviews;

[0038] The driver's cross-platform behavior change coefficient is calculated based on the service activity change coefficient, income change coefficient, violation change coefficient, and positive review change coefficient;

[0039] Collect the time interval between when the driver switches from other platforms to this platform and when he switches from this platform, count the number of times the driver switches from this platform during the time period, count the total amount of orders completed by the driver on this platform during the time period, and comprehensively evaluate to obtain the driver's cross-platform switching coefficient;

[0040] The user churn risk index is calculated based on the driver’s cross-platform behavior change coefficient and the driver’s cross-platform switching coefficient.

[0041] Preferably, the steps of identifying the user groups who will continue to use the platform and formulating corresponding operation strategies and measures are:

[0042] Classify and analyze user groups with high churn risk, segment them according to their behavioral characteristics and preferences, and conduct in-depth analysis of user groups in different categories to understand their characteristics and behavioral patterns;

[0043] Identify the user groups that will continue to use the platform among those with a high risk of churn, and find the common characteristics and behavioral patterns of these users through data analysis and mining;

[0044] Formulate targeted operational strategies and measures based on the analysis results of user groups, including personalized recommendations, preferential activities and exclusive services;

[0045] Implement the formulated operational strategies and measures, and continuously monitor and evaluate their effectiveness. Based on the monitoring results, timely adjust and optimize the operational strategies.

[0046] Technical effects and advantages of the present invention:

[0047] Collect data from drivers on various platforms, preprocess the collected data, extract features from the preprocessed data, select a machine learning model to train the model, and select the best model based on the training results. Evaluate the reliability of the model as the amount of data changes, predict user churn risk, issue early warnings to people with a high level of user churn risk, and intelligently circle people with a high level of user churn risk for operational population packages. Classify and analyze these users, and formulate corresponding operational strategies and measures to accurately screen out drivers with a high risk of churn, thereby increasing the number of orders completed by drivers on the platform, the efficiency of selecting retained drivers, and increasing the time drivers stay on the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the overall flow chart of the present invention.

[0049] Figure 2 This is a flow chart of model training and model application of the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures recorded in the following implementation modes are merely illustrative. The SaaS-based multi-platform online car-hailing driver retention prediction method involved in the present invention is not limited to the various structures recorded in the following implementation modes. All other implementation modes obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0051] The present invention provides a SaaS-based multi-platform online car-hailing driver retention prediction method, comprising the following steps:

[0052] S001: Collecting the driver's data on various platforms, including the driver's registration and login data on various platforms, driver attribute data, vehicle dispatch behavior data, reward data, and violation penalty data;

[0053] S002: Preprocessing the collected data, including cleaning outliers, supplementing missing values, and normalizing the data;

[0054] S003: extracting features from the preprocessed data, wherein the features include driver attribute features, driver behavior features, and driver cross-platform switching behaviors;

[0055] S004: Select an appropriate machine learning model, train the model using historical data, and select the final model based on the training results;

[0056] S005: Use the learning curve to evaluate the reliability of the final model as the amount of data changes, and iteratively optimize the model with low reliability. After optimization, continue to evaluate the reliability until the model reliability meets the standard. The learning curve is a visualization tool for evaluating the performance of machine learning models. The learning curve usually uses the number of training samples as the horizontal axis and the model performance index as the vertical axis to show the performance change trend of the model under training data of different scales;

[0057] S006: Use the model with passed reliability to predict the risk of user churn, issue early warnings to people with high risk of user churn, and intelligently select people with high risk of user churn for operation groups, classify and analyze these users, identify user groups that may continue to use the platform, and formulate corresponding operation strategies and measures to improve user retention rate;

[0058] The overall model training and model application process is divided into two parts: offline and online. Figure 2 As shown, offline is used for model training iteration and model evaluation, and online is used for model application and real-time computing.

[0059] In this embodiment, it should be specifically explained that the steps of preprocessing the collected data are:

[0060] Delete duplicate records in the data set to ensure the uniqueness of the data, and detect and process outliers through methods such as box plots and scatter plots;

[0061] Find missing values ​​in the data set and delete them. If the number of missing values ​​is small, you can consider directly deleting the records containing missing values. For numerical data, you can use the mean, median or other statistics to fill the missing values. For categorical data, you can use the mode to fill the missing values. You can also use machine learning algorithms for prediction and filling.

[0062] Use Min-Max standardization or Z-score standardization to scale the data to a specified range. The Min-Max standardization formula is: The Z-score normalization formula is: Transform the data into a distribution with mean 0 and standard deviation 1, where X is the original data and X min and X max It is expressed as the minimum and maximum values ​​of the data, μ is the mean of the data, and σ is the standard deviation of the data.

[0063] In this embodiment, it should be specifically explained that the driver attribute characteristics include basic information such as the driver's gender, age, city, and platform registration time; the driver behavior characteristics include the driver's first registration time, the driver's daily login status, the driver's daily driving status, the driver's daily order completion status, the driver's daily rewards, and all driver penalties and disposals; the driver's cross-platform switching behavior includes the driver's cross-platform behavior comparison characteristics before and after the cross-platform switch and the driver's cross-platform switching RFM characteristics.

[0064] The driver's daily login status includes login frequency, login duration, login time distribution and continuous login status. The higher the login frequency, the higher the driver's attention and participation in the platform. A long login time may mean that the driver spends more time on the platform and may be more likely to participate in order services. Recording the time distribution of the driver's login to the platform can help the platform understand the driver's activity habits and preferences, and help optimize scheduling and operation strategies. Continuous logins may indicate the driver's continuous attention and participation in the platform, which is a good performance of activity.

[0065] The driver's daily driving situation includes the number of orders, service time, service area distribution and order type distribution. The number of orders can reflect the driver's workload and service capabilities. A large number of orders may indicate a high level of driver activity. The length of service time can reflect the driver's working hours and service efficiency. The distribution of the driver's daily service area is recorded. This can help the platform understand the driver's service scope and activity area. The different order type distributions can reflect the driver's service preferences and business characteristics.

[0066] The driver's daily rewards include the reward type, reward amount and reward issuance time, and the driver's total punishment and disposal include the violation type, punishment type, number of punishments and punishment execution status.

[0067] The comparison characteristics of the driver's cross-platform behavior before and after include:

[0068] Changes in service activity: Compare the service activity of drivers before and after switching platforms. This includes changes in the number of orders received, service time, service frequency, etc. Whether the driver shows obvious changes in service activity before and after switching platforms can help assess their adaptation to the new platform and their work enthusiasm;

[0069] Changes in income levels: Analyze changes in drivers’ income levels before and after switching platforms. You can compare indicators such as drivers’ average daily income, average order price, rewards and subsidies on different platforms to understand whether drivers’ income is affected by platform switching;

[0070] Violations: Compare the violations of drivers before and after switching platforms. This includes changes in violations such as canceling orders, refusing to pick up passengers, and violating traffic rules. Whether there are significant changes in the violations of drivers before and after switching platforms can help assess their compliance with the rules and regulations of the new platform;

[0071] Service quality evaluation: Evaluate the driver's service quality evaluation before and after switching platforms. You can compare indicators such as passenger evaluation, complaint rate, and service complaint handling to understand whether the driver's service quality performance on the new platform has changed.

[0072] The driver cross-platform switching RFM features include:

[0073] Last switch time: the time between the last time a driver switched to a certain platform and the time between the last time they switched out of that platform. Generally speaking, the longer the time between the last switch, the more satisfied or fonder the customer is of that platform or platform service;

[0074] Platform switching frequency: the number of times a customer switches platforms within a certain period of time. Drivers with lower switching frequency are more satisfied customers, have higher loyalty, and are more valuable as drivers.

[0075] Service amount: refers to the total amount of orders completed by the driver within a certain period of time. This feature can be used to evaluate the change in the driver's income level before and after switching platforms, that is, whether the driver has higher order income after switching platforms;

[0076] The RFM feature is a commonly used model for customer analysis. In the driver's cross-platform switching behavior, the RFM model can be applied to the driver's behavioral data to evaluate the changes in the driver's behavioral characteristics before and after the platform switching. By analyzing the RFM features of the driver's cross-platform switching, the platform can understand the changes in the driver's behavioral characteristics before and after the platform switching, so as to formulate corresponding management strategies and optimization measures to promote the driver's work enthusiasm and service quality.

[0077] In this embodiment, it should be specifically explained that the steps of selecting a machine learning model, training the model using historical data, and selecting a final use model according to the training results are as follows:

[0078] Define the evaluation basis of the model, and its implementation code is:

[0079] fromsklearn.metricsimportroc_auc_scoredefperformance_metric(y_true,y_pred):score=roc_auc_score(y_true,y_pred)returnscore;

[0080] Collect historical data and divide it into training set and test set. The implementation code is as follows:

[0081] fromsklearn.cross_validationimporttrain_test_splity=df1['retained']

[0082] X = df1.drop(['retained'], axis = 1) X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.3, random_state = 1)

[0083] X_test.head()

[0084] Attempt to select and filter classification models, and the implementation code is as follows:

[0085] from sklearn.linear_model import LogisticRegression

[0086] from sklearn.svm import SVC

[0087] from sklearn.neighbors import KNeighborsClassifier

[0088] from sklearn.ensemble import RandomForestClassifier clf_lr = LogisticRegression()

[0089] clf_svc = SVC()

[0090] clf_knn = KNeighborsClassifier()

[0091] clf_rf = RandomForestClassifier()

[0092] for clf in (clf_lr, clf_svc, clf_knn, clf_rf): clf.fit(X_train, y_train) y_train_pred = clf.predict(X_train) y_test_pred = clf.predict(X_test) print(clf.__class__.__name__) print('Score on training dataset:', performance_metric(y_train, y_train_pred)) print('Score on test dataset:', performance_metric(y_test, y_test_pred));

[0093] The training results of each model are shown in Table 1:

[0094]

[0095] Table 1

[0096] From the comprehensive evaluation in the above table, it can be seen that the random forest model performs better than other models, so the random forest model is selected for further refinement and parameter adjustment;

[0097] GridSearchCV is used to perform model cross-validation tuning. The GridSearchCV is a tool for parameter tuning and model selection provided in the scikit-learn library. It evaluates the performance of the model under different parameter combinations through cross-validation and selects the parameter combination with the best performance. Using GridSearchCV for model cross-validation tuning is a commonly used machine learning parameter tuning method. It can automatically search for the optimal parameter combination within a given parameter range and evaluate the performance of the model through cross-validation.

[0098] In this embodiment, it should be specifically explained that the steps of using GridSearchCV to perform model cross validation tuning are:

[0099] Determine the parameter space of the model and use the GridSearchCV class in scikit-learn to create a grid search object. When creating the object, you need to specify the model to be optimized, the parameter space, the number of cross-validation folds, and other parameters;

[0100] Call the fit method of the GridSearchCV object, pass in the training data and target label, and start searching for the optimal parameter combination in the parameter space. GridSearchCV will use cross-validation to evaluate the performance of each parameter combination;

[0101] After the search is complete, you can get the best parameter combination through the best_params_ attribute. In addition, you can also get the best model under the optimal parameters through the best_estimator_ attribute;

[0102] The model trained with the optimal parameters is evaluated on the test set to calculate the performance indicators of the model.

[0103] In this embodiment, it should be specifically explained that the step of evaluating the reliability of the final use model as the amount of data changes through the learning curve is:

[0104] The cross-validation method is used to divide the data set into training set and validation set;

[0105] Gradually increase the amount of data in the training set, starting with a smaller subset and then gradually increasing the amount of data;

[0106] Perform model training on training sets of each data size, and verify on verification sets of the same size, and record the performance indicators of training and verification, including loss function, accuracy, and recall;

[0107] The model reliability index is obtained by comprehensive evaluation of the loss function, accuracy and recall rate, and its calculation formula is: Wherein RB represents the model reliability index, LS represents the loss function, and a lower loss function value indicates that the model has better performance, because the loss function is an indicator to measure the difference between the model prediction result and the true label. Therefore, a model with a higher reliability index usually has a lower loss function value. PC represents the accuracy rate, which represents the proportion of samples correctly predicted by the model to the total number of samples. The higher the accuracy rate, the better the consistency between the model prediction result and the true label, and the better the performance of the model. Therefore, the larger the accuracy rate, the larger the model reliability index. RC represents the recall rate, which is used to measure the ability of the model to correctly identify positive examples, that is, how many samples can be correctly predicted by the model among all samples that are actually positive examples. The higher the recall rate, the more capable the model is of identifying true positive examples, and the better the performance of the model. Therefore, the higher the recall rate, the larger the model reliability index. a1, a2, and a3 are represented as weight coefficients of the loss function, the accuracy rate, and the recall rate. The sum of a1, a2, and a3 is 1, and this embodiment does not perform specific calculations on the specific values ​​of a1, a2, and a3.

[0108] In this embodiment, it should be specifically explained that the loss function calculation steps are:

[0109] Choose a loss function that is suitable for the task. Common ones include mean square error (MSE), cross entropy loss function (CrossEntropy Loss) and logarithmic loss function (Log Loss).

[0110] Use the feature data of the training set to input the mean square error model to obtain the prediction results of the model;

[0111] Compare the model's predictions with the true labels of the training set and calculate the loss value;

[0112] Sum the loss values ​​of all samples and find their average to get the average loss of the entire training set, which is the loss function value. The calculation formula is: in represents the loss value of sample i, n represents the number of samples, y i is the true label of sample i, It is represented as the predicted value of the model for sample i.

[0113] In this embodiment, it should be specifically explained that the accuracy calculation steps are:

[0114] The model predicts each sample in the test set and obtains the prediction result;

[0115] Compare the model's predictions with the true labels of the test set, and count the number of samples correctly predicted by the model;

[0116] The accuracy is obtained by dividing the number of correctly predicted samples by the total number of samples. The calculation formula is: Where PC represents the accuracy, n represents the number of samples, NUM R It is expressed as the number of samples predicted correctly.

[0117] In this embodiment, it should be specifically explained that the recall rate calculation steps are:

[0118] The model predicts each sample in the test set and obtains the prediction result;

[0119] Compare the model's predictions with the true labels of the test set, and count the number of samples that the model correctly predicts as positive examples and the number of samples that are incorrectly predicted as negative examples but are actually positive examples;

[0120] The recall rate is obtained by dividing the number of true positive examples by the total number of samples that are true positive examples. The total number of samples that are true positive examples includes the number of true positive examples and the number of false negative examples. The calculation formula is: Where RC represents the recall rate, NUM 真 Expressed as the number of true positives, NUM 假 It is expressed as the number of false negative examples, the number of true positive examples is expressed as the number of positive examples correctly identified by the model, and the number of false negative examples is expressed as the number of positive examples incorrectly predicted as negative examples by the model.

[0121] In this embodiment, it should be specifically explained that the steps of predicting the user churn risk using the model with passed reliability are:

[0122] The service activity change coefficient is calculated based on the change in service activity, and its calculation formula is: Among them, CV represents the service activity variation coefficient, BZ1 represents the standard deviation of the number of orders received, service time, and service frequency indicators of the driver on this platform, BZ2 represents the standard deviation of the number of orders received, service time, and service frequency indicators of the driver on the switching platform, JZ1 represents the mean of the number of orders received, service time, and service frequency indicators of the driver on this platform, and JZ2 represents the mean of the number of orders received, service time, and service frequency indicators of the driver on the switching platform;

[0123] Collect the total income of the driver on this platform and on the switching platform. The total income is the sum of the average daily income, average order price, rewards and subsidies. The income change coefficient is calculated based on the total income of the driver on this platform and on the switching platform. The calculation formula is: Where IC represents the income change coefficient, IC1 represents the total income of the current platform, and IC2 represents the total income of the switching platform;

[0124] Collect the number of violations of the driver on this platform and the number of violations when switching platforms in the same time period, and calculate the violation change coefficient, which is calculated as follows: Where VL represents the violation variation coefficient, VL1 represents the number of violations committed by the driver on the current platform, and VL2 represents the number of violations committed by the driver when switching platforms;

[0125] Collect the number of positive service reviews that drivers received on this platform and the number of positive service reviews that they received on the switching platform during the same period of time, and calculate the positive review change coefficient, which is calculated as follows: EL represents the coefficient of change of favorable comments, EL1 represents the number of favorable comments the driver received on the current platform, and EL2 represents the number of favorable comments the driver received after switching platforms;

[0126] The driver's cross-platform behavior change coefficient is calculated based on the service activity change coefficient, income change coefficient, violation change coefficient, and positive review change coefficient. The calculation formula is: BC represents the coefficient of change of driver's cross-platform behavior, CV represents the coefficient of change of service activity, IC represents the coefficient of change of income, VL represents the coefficient of change of violation, and EL represents the coefficient of change of positive reviews;

[0127] Collect the time interval between the driver switching from other platforms to this platform and switching from this platform, count the number of times the driver switches from this platform during the time period, count the total amount of orders completed by the driver on this platform during the time period, and comprehensively evaluate to obtain the driver's cross-platform switching coefficient, which is calculated as follows: Where ST is the driver's cross-platform switching coefficient, TI is the time interval between the driver switching from other platforms to this platform and switching from this platform, NUM S It is represented by the number of times the driver cuts out of the platform during the time period, and TO is represented by the total amount of orders completed by the driver on the platform during the time period;

[0128] The user churn risk index is calculated based on the driver's cross-platform behavior change coefficient and the driver's cross-platform switching coefficient, and its calculation formula is CR=b1×BC+b2×ST, where CR represents the user churn risk index, BC represents the driver's cross-platform behavior change coefficient, the greater the driver's behavior changes on different platforms, the higher the risk of churn, and the greater the difference in the driver's behavior on different platforms, the greater the possibility of churn. This situation may indicate that the changes in the environment, conditions, policies, etc. faced by drivers on different platforms have led to the instability of their behavior, thereby increasing the risk of churn. ST represents the driver's cross-platform switching coefficient. Frequent switching between different platforms by drivers may increase their risk of churn, and frequent switching between different platforms by drivers may increase their possibility of churn. b1 and b2 represent the weight coefficients of the driver's cross-platform behavior change coefficient and the driver's cross-platform switching coefficient, and this embodiment does not specifically calculate the specific values ​​of b1 and b2.

[0129] The step of issuing an early warning to people with a high user churn risk level is to compare the user churn risk index with a preset threshold. If the user churn risk index is less than the preset threshold, it is determined that the driver churn risk is small and no churn risk warning is issued. If the user churn risk index is greater than the preset threshold, it is determined that the driver churn risk is large and a churn risk warning is issued for the driver.

[0130] In this embodiment, it should be specifically explained that the steps of identifying the user group that may continue to use the platform and formulating corresponding operation strategies and measures are as follows:

[0131] Classify and analyze user groups with high churn risk, segment them according to their behavioral characteristics, preferences, etc., and conduct in-depth analysis of user groups in different categories to understand their characteristics and behavioral patterns;

[0132] Identify the user groups that are likely to continue using the platform among those with a high risk of churn, and find the common characteristics and behavioral patterns of these users through data analysis and mining;

[0133] Based on the analysis results of user groups, formulate targeted operational strategies and measures to improve the retention rate and activity of these users, including personalized recommendations, preferential activities and exclusive services;

[0134] Implement the formulated operational strategies and measures, and continuously monitor and evaluate their effectiveness. Based on the monitoring results, timely adjust and optimize the operational strategies to improve user retention and platform activity.

[0135] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0136] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A SaaS-based multi-platform online car-hailing driver retention prediction method, characterized in that: The following steps are involved: S001: Collecting the driver's data on various platforms, including the driver's registration and login data on various platforms, driver attribute data, vehicle dispatch behavior data, reward data, and violation penalty data; S002: Preprocessing the collected data, including cleaning outliers, supplementing missing values, and normalizing the data; S003: extracting features from the preprocessed data, wherein the features include driver attribute features, driver behavior features, and driver cross-platform switching behaviors; S004: Select a machine learning model, train the model using historical data, and select the final model based on the training results; S005: Use the learning curve to evaluate the reliability of the final model as the amount of data changes, and iteratively optimize the model with low reliability. After optimization, continue to evaluate the reliability until the model reliability meets the standard. S006: Use a model that has passed the reliability test to predict the risk of user churn, issue early warnings to people with high risk of user churn, and intelligently circle the people with high risk of user churn for operational populations, classify and analyze these users, identify user groups who will continue to use the platform, and formulate corresponding operational strategies and measures.

2. According to claim 1, a SaaS-based multi-platform online car-hailing driver retention prediction method is characterized by: The steps of preprocessing the collected data are as follows: Delete duplicate records in the data set to ensure data uniqueness and perform outlier detection and processing; Find missing values ​​in the data set, delete missing values, use mean and median to fill missing values ​​for numerical data, and use mode to fill missing values ​​for categorical data; Use Min-Max normalization to scale the data to a specified range. The Min-Max normalization formula is: Transform the data into a distribution with mean 0 and standard deviation 1, where X norm Expressed as standardized values, X is the original data, X min and X max Expressed as the minimum and maximum values ​​of the data.

3. According to claim 1, a SaaS-based multi-platform online car-hailing driver retention prediction method is characterized by: The steps of selecting a machine learning model, training the model using historical data, and selecting a final model for use based on the training results are as follows: Define the evaluation basis of the model; Collect historical data and divide it into training set and test set; Try and screen classification models; Use GridSearchCV to perform model cross-validation tuning.

4. The SaaS-based multi-platform online car-hailing driver retention prediction method according to claim 1 is characterized by: The steps of evaluating the reliability of the final use model as the amount of data changes through the learning curve are: The cross-validation method is used to divide the data set into training set and validation set; Gradually increase the amount of data in the training set, starting with a small subset and then gradually increasing the amount of data; Perform model training on training sets of each data size, and verify on verification sets of the same size, and record the performance indicators of training and verification, including loss function, accuracy, and recall; The model reliability index is obtained by comprehensive evaluation of the loss function, accuracy and recall rate, and its calculation formula is: Where RB represents the model reliability index, LS represents the loss function, PC represents the accuracy, RC represents the recall rate, a1, a2, and a3 represent the weight coefficients of the loss function, accuracy, and recall rate, and the sum of a1, a2, and a3 is 1.

5. A SaaS-based multi-platform online car-hailing driver retention prediction method according to claim 4, characterized in that: The loss function calculation steps are: Choose a loss function that is suitable for the task. Common ones include mean square error, cross entropy loss function, and logarithmic loss function. Use the feature data of the training set to input the mean square error model to obtain the prediction results of the model; Compare the model's predictions with the true labels of the training set and calculate the loss value; Sum the loss values ​​of all samples and find their average to get the average loss of the entire training set, which is the loss function value. The calculation formula is: in represents the loss value of sample i, n represents the number of samples, y i is the true label of sample i, It is represented as the predicted value of the model for sample i.

6. The SaaS-based multi-platform online car-hailing driver retention prediction method according to claim 1 is characterized by: The steps of predicting the user churn risk using the model with passed reliability are as follows: The service activity change coefficient is calculated based on the change in service activity; Collect the driver's total income on this platform and on the switched platform. The total income is the sum of the average daily income, average order price, rewards and subsidies. The income change coefficient is calculated based on the driver's total income on this platform and on the switched platform. Collect the number of violations that the driver committed on this platform and on the switching platform in the same time period, and calculate the violation change coefficient; Collect the number of positive service reviews that drivers received on this platform and the number of positive service reviews that they received on the switched platform during the same period of time, and calculate the coefficient of change in positive reviews; The driver's cross-platform behavior change coefficient is calculated based on the service activity change coefficient, income change coefficient, violation change coefficient, and positive review change coefficient; Collect the time interval between when the driver switches from other platforms to this platform and when he switches from this platform, count the number of times the driver switches from this platform during the time period, count the total amount of orders completed by the driver on this platform during the time period, and comprehensively evaluate to obtain the driver's cross-platform switching coefficient; The user churn risk index is calculated based on the driver’s cross-platform behavior change coefficient and the driver’s cross-platform switching coefficient.

7. The SaaS-based multi-platform online car-hailing driver retention prediction method according to claim 2 is characterized by: The steps to identify the user groups who will continue to use the platform and formulate corresponding operation strategies and measures are as follows: Classify and analyze user groups with high churn risk, segment them according to their behavioral characteristics and preferences, and conduct in-depth analysis of user groups in different categories to understand their characteristics and behavioral patterns; Identify the user groups that will continue to use the platform among those with a high risk of churn, and find the common characteristics and behavioral patterns of these users through data analysis and mining; Formulate targeted operational strategies and measures based on the analysis results of user groups, including personalized recommendations, preferential activities and exclusive services; Implement the formulated operational strategies and measures, and continuously monitor and evaluate their effectiveness. Based on the monitoring results, timely adjust and optimize the operational strategies.

Images