Two-channel loss risk grading and intelligent recall method fusing XGBoost prediction and user portrait construction

Abstract

The invention relates to a two-channel loss risk grading and intelligent callback method fusing XGBoost prediction and user portrait construction. The method comprises the following steps: collecting user data based on an online state grid platform, and constructing a power user feature set; constructing a Boruta model based on active dimension features in the power user feature set, and screening out key features used for judging user loss from high-dimensional features; constructing a potential lost user prediction model based on an XGBoost algorithm; calculating a correlation coefficient between the preference dimension features and the screened user loss key features, and identifying features highly related to user loss in the preference dimension; constructing differentiated lost user portraits; and establishing a two-dimensional mapping table of risk levels * group preferences. According to the method, fine grading of user loss risks is realized, the scientificity of parameter configuration is improved while the feature effectiveness is ensured, and automatic optimization of model hyper-parameters is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital data processing, and particularly relates to a double-channel churn risk grading and intelligent recall method fusing XGBoost prediction and user portrait construction. BACKGROUND

[0002] With the rapid development of digital economy, various mobile applications providing electricity payment and life services have emerged in the market. Online State Grid is facing the competitive pressure of user diversion through multiple channels, and it is urgent to accurately identify potential churn users from a large number of user groups and develop effective user recall strategies. The current online State Grid churn user identification method mainly relies on user activity data, determines the threshold by statistical analysis of the last active time, determines the churn user when the user is inactive for a certain period of time, and uses standardized in-station messages or short messages to reach out for recall. This method is based on the last active time of the user for analysis, focuses on post-recall of user churn, lacks advanced churn user prediction methods and tools, and fails to identify potential churn users in advance, making it impossible to intervene and guide before the user churns and maximize user retention. On the other hand, the existing recall strategy lacks matching degree with the characteristics of churn users. When implementing the recall of churn users, the behavior characteristics and use preferences of users before they churn are not analyzed, and there is a lack of personalized support. Therefore, it is necessary to carry out more refined analysis and modeling based on the portrait of churn users, so as to realize the accurate matching and effective implementation of the recall strategy.

[0003] The traditional user churn risk prediction model is to identify the user groups that have churned by statistical analysis of the last active related data of the customers in the context of electric power big data, and to establish an evaluation system accordingly. With the rapid development of the artificial intelligence industry, research on the establishment of electric power customer churn risk prediction models using machine learning algorithms has emerged. These algorithms can mine useful information from user data and analyze user churn risk levels. First, the random forest algorithm is used to perform a feature selection on the key factors affecting user churn. Then, the back propagation neural network is used to preliminarily predict the user churn probability and output the credit score of user activity and loyalty. Finally, the logistic regression model is used for secondary discrimination and verification to effectively grade and warn the user churn risk.

[0004] While the aforementioned methods can improve prediction accuracy to some extent, they still have shortcomings: First, relying solely on random forests for feature selection can easily miss key features with non-linear relationships or strong correlations, resulting in a suboptimal subset of selected features. Second, the inherent complexity and black-box nature of neural network models lead to weak interpretability, failing to provide a clear basis for subsequent recall strategies. Secondary discrimination using logistic regression struggles to fully utilize the interaction information of complex features, potentially resulting in insufficiently precise identification of medium- to high-risk users. Furthermore, this method primarily focuses on risk assessment, neglecting to incorporate user functional features or behavioral preferences for feature mining, and lacks the construction of churned user profiles, leading to a disconnect between prediction results and personalized recall measures.

[0005] Explanation of relevant terms:

[0006] Dual-channel ensemble learning: Dual-channel ensemble learning is an ensemble learning method that enhances the robustness and adaptability of a model by constructing two different information processing channels to learn different features of the data in parallel or interactively. Each channel can employ a different model structure, feature set, or learning strategy, and the outputs of the two channels are ultimately fused to improve overall prediction performance.

[0007] User profiling: This involves characterizing the key features and behavioral patterns of potential churned users. Through integrated analysis of churned users' basic attributes, usage habits, interaction preferences, and historical behaviors, a differentiated user profile model is created. This profile provides a basis for the layered design and personalized intervention of recall strategies, thereby improving the effectiveness of recall measures.

[0008] Accuracy: This directly reflects the model's prediction accuracy across the entire sample set. It represents the proportion of samples that can be correctly predicted out of all samples.

[0009] AUC: Used to measure the model's ability to distinguish between positive and negative samples. The ROC curve is a graphical representation method. The true positive rate is the proportion of samples that are actually positive that are correctly predicted as positive by the model. The false positive rate is the proportion of samples that are actually negative that are incorrectly predicted as positive by the model. The true positive rate is plotted on the vertical axis and the false positive rate is plotted on the horizontal axis, which shows the model's ability to predict positive samples. The AUC value is the area under the ROC curve, and its value ranges from [0,1]. The closer the value is to 1, the stronger the model's ability to distinguish between positive and negative samples.

[0010] XGBoost is an ensemble learning algorithm based on the gradient boosting framework. Its core principle is to minimize the loss function by iteratively building a decision tree model and gradually correcting the prediction residuals of preceding models. It optimizes the objective function through second-order Taylor expansion, effectively handling high-dimensional sparse data. Its objective function consists of a loss function and a regularization term.

[0011] ,

[0012] Where n is the number of samples; This represents the true label of sample i (0 for normal users, 1 for users in arrears). Let i be the predicted value after the t-th iteration. For the t-th tree, Let w be the number of leaf nodes in the k-th tree, and w be the leaf weight. , are hyperparameters, representing the leaf node complexity penalty coefficient and the L2 regularization parameter, respectively; The loss function is calculated as follows:

[0013] ,

[0014] in, Let be the predicted probability of sample i. XGBoost approximates the loss function through a second-order Taylor expansion and uses a regularization term to prevent overfitting while improving the model's generalization ability. A greedy algorithm is then used to calculate the optimal split gain.

[0015] ,

[0016] , Summing the first-order gradients of the left and right child nodes. , This is the sum of the second-order gradients of the left and right child nodes. The denominator is the regularized second-order gradient statistic, used to balance loss optimization and model complexity. Summary of the Invention

[0017] This invention addresses the problem of user churn risk prediction for the State Grid's online app by proposing a dual-channel churn risk classification and intelligent recall method that integrates XGBoost prediction and user profiling. This method achieves accurate prediction of user churn risk through a classification model based on XGBoost and optimized by Bayesian methods. Simultaneously, it utilizes feature correlation analysis and K-means clustering to construct churn user profiles, accurately depicting the behavioral patterns and usage preferences of different user groups, thus achieving dual-dimensional empowerment for churn risk classification and churn user recall. The specific technical solution is as follows:

[0018] A dual-channel churn risk classification and intelligent recall method integrating XGBoost prediction and user profiling includes the following process:

[0019] Step 1: Based on the State Grid online platform, collect user data and construct a power user feature set;

[0020] The constructed feature set is divided into two dimensions:

[0021] The activity dimension specifically includes: basic information characteristics and active behavior characteristics, which are used to predict the risk level of potential churned users;

[0022] Preference dimensions specifically include: user preference behavior characteristics, used to build refined churn user profiles;

[0023] Step 2: Based on the active dimension features in the above power user feature set, construct the Boruta model and combine it with feature importance ranking to initially screen out key features for judging user churn from the high-dimensional features;

[0024] Step 3: Construct a potential churn user prediction model based on the XGBoost algorithm, and train and learn the key features selected above; the potential churn user prediction model outputs a churn probability value between 0 and 1 for each user to measure the possibility of user churn, and divides users into three levels of low risk, medium risk and high risk according to the set grading threshold range, so as to realize the hierarchical identification of potential churn users.

[0025] Step 4: Introduce the feature correlation analysis method to calculate the correlation coefficient between the preference dimension features and the selected key user churn features, and identify the features in the preference dimension that are highly correlated with user churn;

[0026] Step 5: Based on the selected preference features that are highly correlated with user churn, use clustering methods to group users and build differentiated churn user profiles to characterize the behavioral patterns and needs of potential churn users;

[0027] Step 6: Combine the user churn risk level output in Step 3 with the churn user group profile constructed in Step 5 to establish a two-dimensional mapping table of "risk level × group preference". In each cross unit, a corresponding recall strategy is preset. The system locates the user in the corresponding risk-profile category according to the combination of the user's risk level and group preference, and automatically matches the corresponding recall plan from the preset recall strategy library, thereby realizing differentiated and personalized recall for users with different risk levels and different group characteristics.

[0028] Furthermore, the basic information features of the active dimension in step 1 include the following fields: gender, age, address, registration time, binding status, number of bound account numbers, electricity usage category, whether my subscription is activated, and whether I am a member;

[0029] The active behavior features in step 1 include the following fields:

[0030] Last active time, last active click count, average monthly active users, average monthly logins, average monthly activity participations, average click count, average monthly power consumption, average monthly active users in the six months prior to the last active time, average monthly logins in the six months prior to the last active time, average monthly activity participations in the six months prior to the last active time, average click count in the six months prior to the last active time, average monthly active users in the three months prior to the last active time, average monthly logins in the three months prior to the last active time, average monthly activity participations in the three months prior to the last active time, average click count in the three months prior to the last active time, active users in the month prior to the last active time, logins in the month prior to the last active time, activity participations in the month prior to the last active time, average click count in the month prior to the last active time;

[0031] The preference dimension mentioned in step 1 includes the following fields:

[0032] Number of times participating in promotional activities, number of times discounts were claimed, number of discounts used, number of expired discounts, first preferred function, second preferred function, third preferred function, last used function, number of customer service inquiries, average monthly clicks on featured payment products, average monthly clicks on featured services, and the type of product most recently used. The first preferred function refers to the function the user uses most frequently, the second most frequently used function refers to the function the user uses second most frequently, and the third most frequently used function refers to the function the user uses third most frequently.

[0033] Furthermore, step 2 constructs a Boruta model based on the active dimension features of the aforementioned electricity user feature set, and filters out key features for judging user churn from the high-dimensional features, specifically including the following process:

[0034] Create corresponding random "shadow" features for the 28 real features, and calculate the importance scores of the real features and shadow features in the random forest model according to the following formula.

[0035] ,

[0036] in, Represents the user feature vector. The number of trees in the random forest. The proportion of samples reaching node n, i.e., the node weight, is determined when node n uses features. After partitioning, the improvement in Gini purity is: ;

[0037] By comparing feature importance scores, the model filters out true features whose scores are higher than the highest importance score among the shadow features. The model then outputs a ranking of the importance of each feature, such as... Figure 2 As shown, the 17 features ranked 1 are the important features selected by the model, while the remaining ranked features are redundant features that have been removed.

[0038] Furthermore, step 2 includes the following specific processes:

[0039] Randomly shuffle the original features Based on the feature values, create corresponding shadow features. ;

[0040] Using datasets Training a random forest model;

[0041] Calculate the importance scores of the original features and shadow features. , ;

[0042] Find the maximum value in the shadow feature importance score. ;

[0043] If the importance score of the original feature Greater than Then the feature Mark as important; otherwise, leave as pending.

[0044] Delete items with an importance score less than Features;

[0045] Repeat the above process until the preset maximum number of iterations is reached or all features are correctly classified.

[0046] Furthermore, step 2 also includes the following process:

[0047] Based on the 17 important features initially selected by the Boruta algorithm, a second screening was conducted according to their feature importance scores to obtain 10 key features for predicting churned users.

[0048] Furthermore, step 3 includes the following specific process:

[0049] Using XGBoost as the core algorithm, a user churn risk prediction model is constructed. By sequentially building multiple decision trees, subsequent trees continuously learn and correct the prediction residuals of the previous tree. With its powerful nonlinear fitting capability, it accurately captures the complex mapping relationship between user behavior patterns and churn risk.

[0050] After receiving the key features selected in step 2, the model predicts the churn probability for each user and classifies users into risk levels based on a preset threshold range. When a user is identified as low-risk, a value of 0 is assigned; when When a user is identified as a medium-risk user, a value of 1 is assigned; when At that time, the user was identified as a high-risk user and assigned a value of 2.

[0051] Furthermore, a Bayesian optimization method is introduced to automatically search and tune the model's hyperparameters. This method constructs a probabilistic surrogate model of the objective function and employs a trial-and-error feedback iterative mechanism to guide the hyperparameter combination to converge rapidly toward the optimal performance, efficiently and accurately finding the globally optimal or near-optimal parameter configuration. The parameter combinations are as follows:

[0052] The maximum depth of a tree is max_depth: 6;

[0053] The minimum sum of the weights of all samples in the child node, min_child_weight: 3;

[0054] Minimum loss reduction gamma required for node splitting: 0.44775015309733157;

[0055] The weight of the L1 regularization term, reg_alpha, is 0.021982987387329393.

[0056] The weight of the L2 regularization term, reg_lambda, is 7.257213105405618.

[0057] Learning rate: 0.0700723844724174;

[0058] The proportion of training data randomly sampled when training each tree: 0.9650435093433541;

[0059] The proportion of features randomly sampled when constructing each tree, colsample_bytree: 0.6.

[0060] Furthermore, step 4 includes the following specific process:

[0061] The key features selected in step 2 Behavioral characteristics in the user preference dimension Feature correlation analysis was conducted, and the relationship coefficients between features were calculated using three indicators: Pearson correlation coefficient, Spearman rank correlation coefficient, and ANOVA.

[0062] A dynamic damping adjustment mechanism is introduced to adaptively learn the weight coefficients of three indicators: Pearson, Spearman, and ANOVA. The calculation results of the three indicators are then weighted and combined to construct a feature correlation matrix, thereby uncovering behavioral features that are significantly correlated with key features.

[0063] Furthermore, step 5 includes the following specific process:

[0064] Based on the features highly correlated with user churn identified in step 4, the K-means algorithm is used to segment users. Specifically, the three feature data are input into K-means, and through continuous iterative optimization, the goal is to minimize the sum of squared distances from samples within a cluster to its cluster center, thereby completing user segmentation and constructing a churn user profile with group behavior characteristics.

[0065] Step 6 includes the following specific process:

[0066] Personalized recall strategies are matched based on the user churn risk level output in step 3 and the user group profile output in step 5; different risk levels correspond to different push channels and different push frequencies, and different group profiles correspond to different push content.

[0067] High-risk users: daily SMS messages, 3-5 phone calls per week, and customized push content based on user profiles;

[0068] Medium-risk users: Daily app pop-ups, 1-2 SMS messages per week, and customized push content based on user profile;

[0069] Low-risk users: 3-5 app pop-ups per week, with content customized based on user profile.

[0070] Furthermore, it also includes the following processes:

[0071] Step 7: Introduce a proportional-integral-derivative (PID) control algorithm. By monitoring three parts of information—the current error, cumulative error, and error change trend—between the actual recall rate and the target recall rate of churned users, a control signal is generated to dynamically adjust the churn risk threshold and achieve adaptive optimization of risk classification.

[0072] The advantages of this invention over the prior art are as follows:

[0073] To address the multidimensional characteristics of electricity users, a churn risk prediction model based on Boruta-XGBoost and driven by Bayesian optimization was constructed. Through the synergistic effect of key feature selection and churn probability prediction, a refined classification of user churn risk was achieved. The model's hyperparameters were fine-tuned using Bayesian optimization methods, improving the scientific nature of parameter configuration while ensuring feature effectiveness, thus achieving automatic optimization of model hyperparameters. In this embodiment, a PID dynamic control mechanism was introduced to adjust the risk classification threshold in real time based on recall deviation, ensuring an adaptive balance between coverage and recall efficiency, and realizing dynamic classification and intelligent management of user churn risk.

[0074] This invention employs three types of indicators—Pearson correlation coefficient, Spearman rank correlation coefficient, and analysis of variance—to comprehensively measure the correlation between user preference dimensions and key churn characteristics from multiple perspectives, including linear and nonlinear, continuous and categorical analyses. By introducing a dynamic damping adjustment mechanism, the weights of each indicator are adaptively adjusted according to the data distribution characteristics of different features, constructing a more reasonable feature correlation matrix. Based on highly correlated features, differentiated churn user profiles are built, enabling refined modeling of the behavioral patterns and needs of potential churn users, providing reliable support for subsequent personalized recall strategies.

[0075] This invention draws on a dual-channel ensemble learning mechanism to construct a dual-channel joint framework for churn risk prediction and user profile construction. This overcomes the limitation of traditional methods where churn risk grading and churn user profile construction cannot be trained collaboratively. Through cross-model information fusion, it achieves information complementarity and collaborative optimization between models. A two-dimensional matrix mapping is introduced to cross-integrate risk levels and group profiles, forming a differentiated recall strategy table, enabling precise matching between recall strategies and user characteristics. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the process of a dual-channel churn risk classification and intelligent recall method that integrates XGBoost prediction and user profile construction according to the present invention.

[0077] Figure 2 This is a dual-channel churn risk classification and intelligent recall method that integrates XGBoost prediction and user profile construction in an embodiment of the present invention. In step 2, by comparing the feature importance scores, the true features with scores higher than the maximum importance score among the shadow features are selected, and the model outputs a schematic diagram of the importance ranking of the features.

[0078] Figure 3 This is a schematic diagram of a dual-channel churn risk classification and intelligent recall method that integrates XGBoost prediction and user profile construction in an embodiment of the present invention. In step 2, secondary mining is performed by ranking the importance of features, and finally the top 10 key features in importance are retained.

[0079] Figure 4 This invention presents a dual-channel churn risk grading and intelligent recall method that integrates XGBoost prediction and user profile construction. Step 4 involves behavioral features that are significantly correlated with 10 key features. As shown in the figure, the darker the color in the matrix, the stronger the correlation between the corresponding features. Detailed Implementation

[0080] A dual-channel churn risk classification and intelligent recall method integrating XGBoost prediction and user profiling is described below. Figure 1 As shown, the process includes the following:

[0081] Step 1: After cleaning and fusion processing the high-dimensional heterogeneous user data collected from the State Grid online platform, a complete set of power user features is constructed to characterize the complex interactive behaviors of users in the State Grid online APP.

[0082] Based on business needs and industry experience, the constructed feature set is divided into two main dimensions:

[0083] The activity dimension specifically includes: basic information features and active behavior features, which are used to predict the level of user churn risk.

[0084] Preference dimension is used to characterize users' usage preferences and behavioral characteristics in various functional modules, and to support the construction of refined churn user profiles;

[0085] Table 1 lists the user feature fields required for this invention.

[0086] Table 1 User Feature Fields

[0087]

[0088] Step 2: Based on the active dimension features in the aforementioned high-dimensional electricity user feature set, construct a Boruta model and combine it with feature importance ranking to initially screen out key features for judging user churn from the high-dimensional features, i.e., key factors that have a significant discriminative effect on user churn. This method effectively removes redundant and noisy information from the high-dimensional feature set, and mines active behavioral features highly correlated with user churn, providing reliable support for subsequent churn risk classification prediction. Specifically, the process includes the following steps:

[0089] Create corresponding random "shadow" features for the 28 real features, and calculate the importance scores of the real features and shadow features in the random forest model according to the following formula.

[0090] ,

[0091] in, Represents the user feature vector. The number of trees in the random forest. The proportion of samples reaching node n, i.e., the node weight, is determined when node n uses features. After partitioning, the improvement in Gini purity is: ;

[0092] By comparing feature importance scores, the model filters out true features whose scores are higher than the highest importance score among the shadow features. The model then outputs a ranking of the importance of each feature, such as... Figure 2As shown, the 17 features ranked 1st are the important features selected by the model, while the remaining ranked features are redundant features that were discarded. The specific steps of Boruta's feature selection are as follows:

[0093] Randomly shuffle the original features Based on the feature values, create corresponding shadow features. ;

[0094] Using datasets Training a random forest model;

[0095] Calculate the importance scores of the original features and shadow features. , ;

[0096] Find the maximum value in the shadow feature importance score. ;

[0097] If the importance score of the original feature Greater than Then the feature Mark as important; otherwise, leave as pending.

[0098] Delete items with an importance score less than Features;

[0099] Repeat the above process until the preset maximum number of iterations is reached or all features are correctly classified.

[0100] The Boruta algorithm selected 17 important features from 28 input features, achieving initial dimensionality reduction of high-dimensional features. However, redundant features with low importance still existed in the results. To further compress the feature dimensionality and extract more representative key features, this invention performs secondary mining based on feature importance ranking on the 17 feature subsets. Finally, the top 10 most important features for churn prediction are retained, such as... Figure 3 As shown, this ensures the representativeness and effectiveness of the selected features in the prediction task.

[0101] Step 3: Based on the 10 key features mined, XGBoost is used as the core algorithm to construct a user churn risk prediction model. This involves sequentially building multiple decision trees, allowing subsequent trees to continuously learn and correct the prediction residuals of the previous trees. This powerful non-linear fitting capability accurately captures the complex mapping relationship between user behavior patterns and churn risk. After receiving the 10 key features, the model predicts the churn probability for each user and classifies users into risk levels based on a preset threshold range. When a user is identified as low-risk, a value of 0 is assigned; when When a user is identified as a medium-risk user, a value of 1 is assigned; when At that time, the user was identified as a high-risk user and assigned a value of 2.

[0102] To fully leverage the performance potential of the XGBoost model and overcome the blindness and inefficiency of manual hyperparameter tuning, this invention innovatively introduces a Bayesian optimization method for automated search and tuning of the model's hyperparameters. This method constructs a probabilistic surrogate model of the objective function and employs a trial-and-error feedback iterative mechanism to guide the hyperparameter combination towards rapid convergence towards performance optimization, thereby efficiently and accurately finding the globally optimal or near-optimal parameter configuration. This invention yields the following optimal parameter combination:

[0103] The maximum depth of a tree is max_depth: 6;

[0104] The minimum sum of the weights of all samples in the child node, min_child_weight: 3;

[0105] Minimum loss reduction gamma required for node splitting: 0.44775015309733157;

[0106] The weight of the L1 regularization term, reg_alpha, is 0.021982987387329393.

[0107] The weight of the L2 regularization term, reg_lambda, is 7.257213105405618.

[0108] Learning rate: 0.0700723844724174;

[0109] The proportion of training data randomly sampled when training each tree: 0.9650435093433541;

[0110] The proportion of features randomly sampled when constructing each tree, colsample_bytree: 0.6.

[0111] To verify the superiority of the proposed solution, mainstream classification models such as logistic regression, support vector machine (SVM), random forest, and LightGBM were selected and compared with the XGBoost+Bayesian optimization method used in this invention. In terms of both prediction accuracy and area under the ROC curve (AUC), the proposed method significantly outperformed the comparative models, demonstrating the best classification performance.

[0112] Step 4: Introduce the feature correlation analysis method to calculate the correlation coefficient between the preference dimension features and the selected key user churn features, and identify the features in the preference dimension that are highly correlated with user churn;

[0113] To further explore the behavioral characteristics of users before churn and formulate targeted recall strategies, this invention mines 10 key features based on the Boruta+ feature importance ranking. 12 behavioral characteristics in the user preference dimension Feature correlation analysis was conducted, and the relationship coefficients between features were calculated using three indicators: Pearson correlation coefficient, Spearman rank correlation coefficient, and ANOVA.

[0114] Pearson's law applies to continuous variables and can be used to measure the degree of linear correlation between features. It is calculated using the following formula:

[0115] ,

[0116] in, Features The value of the i-th sample, Features The mean, The closer The stronger the correlation, the better.

[0117] Spearman's law is suitable for ordinal variables or continuous data that do not follow a normal distribution. It can be used to measure non-normal distributions or non-linear monotonic relationships, and is calculated using the following formula:

[0118] ,

[0119] in, This represents the difference in ranking of sample i based on two features, where n is the number of samples. The larger the value, the stronger the monotonic correlation.

[0120] ANOVA is suitable for measuring the correlation between categorical and continuous variables. By comparing the differences in the means of the continuous variable in different categorical groups, it determines whether the categorical variable has a significant impact on the continuous variable. The core statistic is:

[0121] ,

[0122] Where K is the number of groups, i.e., the number of categories in the categorical variable, and n is the total number of samples. Let k be the number of samples in the k-th group. For the i-th sample in the k-th group, This represents the mean of the k-th group. F represents the population mean. When F is significantly large, it indicates that there is a correlation between the variables.

[0123] This invention introduces a dynamic damping adjustment mechanism to adaptively learn the weight coefficients of three indicators: Pearson, Spearman, and ANOVA. The calculation results of these three indicators are then weighted and combined to construct a feature correlation matrix, revealing behavioral features significantly correlated with 10 key features, such as... Figure 4 As shown in the matrix, the darker the color, the stronger the correlation between the corresponding features. The key behavioral features identified include the number of times a user participated in a promotional activity, the number of expired promotions, and the type of product most recently used, revealing the underlying behavioral patterns and potential needs of potential churned users.

[0124] Step 5: Based on the three behavioral features strongly correlated with the key features identified through feature correlation analysis, this invention employs the K-means algorithm to segment users. Specifically, the number of times a user participated in promotional activities, the number of expired promotions, and the type of product most recently used are input into the K-means algorithm. Through continuous iterative optimization, the goal is to minimize the sum of squared distances from samples within a cluster to its cluster center, thus completing the user segmentation. This invention divides the samples into five user groups, each corresponding to different preference features, constructing a churn user profile with group behavioral characteristics. The specific features are as follows:

[0125] Group 0 information-seeking users: those who have not participated in promotional activities, have no record of expired promotions, and prefer query services;

[0126] Group 1: Discount-sensitive and information-seeking users: They have not participated in promotional activities, have expired discount records, prefer inquiry services, and have potential demand for service processing and payment services.

[0127] Group 2: Interactive and information-seeking users: They actively participate in promotional activities, have no record of expired promotions, prefer inquiry and payment services, and have low demand for processing services.

[0128] Group 3: Business-oriented users: They did not participate in promotional activities, have no record of expired promotions, and prefer to handle business-related services;

[0129] Group 4 consists of users who pay frequently: they did not participate in promotional activities, have no record of expired promotions, and prefer payment services.

[0130] Step 6: This invention combines the user churn risk level output from the risk prediction channel with the churn user group profile constructed by the user profile construction channel to establish a two-dimensional mapping table of "risk level × group preference". A corresponding recall strategy is preset in each cross-unit, as shown in Table 2. Based on the user's risk level and group preference combination, the system locates the user in the corresponding risk-profile category and automatically matches the corresponding recall plan from the preset recall strategy library, thereby achieving differentiated and personalized recall for users with different risk levels and different group characteristics.

[0131] Table 2 Risk Level × Group Preference Mapping Table

[0132]

[0133] Step 7: To optimize the setting of the user churn risk level threshold, this invention introduces a proportional-integral-derivative (PID) control algorithm. By monitoring three parts of information—the current error, cumulative error, and error trend—between the actual recall rate and the target recall rate of churned users, a control signal is generated to dynamically adjust the churn risk threshold, achieving adaptive optimization of risk classification. Its core idea is to balance risk coverage and churn recall rate through continuous feedback and adjustment, ensuring that operational resources are prioritized for potential churned users with higher churn risk.

[0134] Recall rate refers to the proportion of successfully recalled users out of the total number of users who should have been recalled; current error refers to the proportional term, cumulative error refers to the integral term, and error trend refers to the differential term.

[0135] The output of the PID controller is as follows:

[0136] ,

[0137] in, It is the difference between the actual recall rate and the target recall rate. For the proportional control item P, based on the current... The size is linearly adjusted, and when the actual recall rate is much lower than the target recall rate, the error is quickly responded to, the threshold is lowered to expand the scope of high-risk users, and immediate compensation control is provided.

[0138] For the integral control term I, long-term correction is performed based on the historical cumulative recall rate error. When the actual recall rate is consistently lower than the target value over a period of time, the system will automatically optimize the model parameters to eliminate steady-state error and improve recall performance.

[0139] The differential control term D is used to predict future trends based on the rate of change of recall error, and to take measures in advance to prevent over-adjustment and system oscillation.

[0140] By organically combining proportional, integral, and derivative components, the PID controller can achieve both rapid response to recall deviations and long-term optimization and trend prediction. Through this dynamic control mechanism, the system can adaptively optimize the churn risk grading threshold, ensuring more accurate churn user identification and recall strategies, thereby significantly improving resource allocation efficiency and user retention rates.

Claims

1. A dual-channel churn risk grading and intelligent recall method fusing XGBoost prediction and user portrait construction, characterized in that, The method comprises the following steps: Step 1: Based on the online state grid platform, user data is collected, and a power user feature set is constructed; The constructed feature set is divided into two dimensions: Active dimension, specifically including: basic information features and active behavior features, used to predict the risk level of potential churn users; Preference dimension, specifically including: user preference behavior features, used to construct a refined churn user portrait; Step 2: Based on the active dimension features in the above power user feature set, a Boruta model is constructed and combined with feature importance ranking to preliminarily screen out key features for judging user churn from high-dimensional features; Step 3: Based on the XGBoost algorithm, a potential churn user prediction model is constructed, and the above screened key features are trained and learned; the potential churn user prediction model outputs a churn probability value between 0 and 1 for each user, which is used to measure the possibility of user churn, and according to the set classification threshold range, the users are divided into three levels of low risk, medium risk and high risk, realizing the hierarchical identification of potential churn users; Step 4: Introducing a feature correlation analysis method, the correlation coefficient between the preference dimension features and the screened user churn key features is calculated, and the features in the preference dimension highly related to user churn are identified; Step 5: For the screened preference features highly related to user churn, clustering method is used to group users, and a differentiated churn user portrait is constructed, realizing the description of the behavior mode and demand preference of potential churn users; Step 6: The user churn risk level output by step 3 and the churn user group portrait constructed in step 5 are combined to establish a two-dimensional mapping table of "risk level x group preference", and a corresponding recall strategy is preset in each intersection unit. The system positions the user to the corresponding risk-image category according to the combination of the user's risk level and group preference, and automatically matches the corresponding recall scheme from the preset recall strategy library, thereby realizing the differentiation and individualization of recall for users with different risk levels and different group characteristics.

2. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling construction according to claim 1, wherein, The basic information features of the active dimension in step 1 include the following fields: gender, age, address, registration time, binding state, number of bound account numbers, electricity category, whether to open my subscription, whether to be a member; The active behavior features of the active dimension in step 1 include the following fields: Last active time, last active click count, monthly active count, monthly login count, monthly activity participation count, average click count, monthly electricity consumption, last active half-year monthly active count, last active half-year monthly login count, last active half-year monthly activity participation count, last active half-year average click count, last active three-month monthly active count, last active three-month monthly login count, last active three-month monthly activity participation count, last active three-month average click count, last active one-month active count, last active one-month login count, last active one-month activity participation count, last active one-month average click count; The preference dimension in step 1 includes the following fields: The number of participation times of the preferential activity, the number of preferential acquisition times, the number of preferential use quantities, the number of preferential expiration quantities, the first function preference, the second function preference, the third function preference, the last use function, the customer service consultation times, the monthly average click times of the special payment product, the monthly average click times of the special service, and the product type used last time.

3. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling construction according to claim 2, characterized in that, Step 2: Based on the active dimension features in the above power user feature set, a Boruta model is constructed, and the feature importance ranking is combined to preliminarily screen out key features for judging user churn from high-dimensional features, including the following processes: Create corresponding random "shadow" features for the 28 real features, and calculate the importance scores of the real features and the shadow features in the random forest model according to the following formula: ， wherein, represents a user feature vector, is the number of trees in the random forest, is the proportion of samples reaching node n, i.e. the node weight, when node n uses feature After partitioning, the Gini purity improvement is ; By comparing the feature importance scores, screen out real features with scores higher than the maximum importance score in the shadow features. The model outputs the importance ranking of each feature, and the top 17 features are the important features selected by the model, and the remaining features are redundant features that are excluded.

4. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling construction according to claim 3, characterized in that, Step 2 includes the following specific processes: randomly shuffle the original feature values to create corresponding shadow features ;​ Using the dataset Training a random forest model; computing importance scores for the original features and shadow features , ; finding the maximum value in the shadow feature importance scores ; If the importance score of the original feature Greater than Then the feature Mark as important; otherwise, leave as pending. deleting features with importance scores less than 0.5; Repeat the above process until the maximum number of iterations is reached or all features are correctly classified.

5. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling construction according to claim 4, characterized in that, The step 2 also includes the following processes: Based on the 17 important features preliminarily screened out by the Boruta algorithm, according to the feature importance score ranking, perform secondary screening to obtain 10 key features for predicting churn users.

6. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling construction according to claim 5, wherein, The step 3 includes the following specific processes: XGBoost is used as the core algorithm to build a user churn risk prediction model. By sequentially building multiple decision trees, the subsequent trees continuously learn and correct the prediction residuals of the previous tree, so as to accurately capture the complex mapping relationship between user behavior patterns and churn risk with powerful nonlinear fitting capability; After receiving the key features selected in step 2, the model predicts the churn probability for each user and classifies users into risk levels based on a preset threshold range. When a user is identified as low-risk, a value of 0 is assigned; when When a user is identified as a medium-risk user, a value of 1 is assigned; when At that time, the user was identified as a high-risk user and assigned a value of 2.

7. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling construction according to claim 6, characterized in that, The Bayesian optimization method is introduced to automatically search and optimize the hyperparameters of the model. This method builds a probability proxy model of the objective function, uses the "trial and feedback" iterative mechanism to guide the hyperparameter combination to quickly converge to the performance optimal direction, and efficiently and accurately finds the global optimal or near-optimal parameter configuration; the parameter combination is as follows: The maximum depth of a tree max_depth: 6; The minimum value of the sum of all sample weights in the child node min_child_weight: 3; The minimum loss reduction required for node splitting gamma: 0.44775015309733157; The weight of the L1 regularization term reg_alpha: 0.021982987387329393; The weight of the L2 regularization term reg_lambda: 7.257213105405618； Learning rate learning_rate: 0.0700723844724174; The proportion of randomly sampled training data when training each tree subsample: 0.9650435093433541; The proportion of randomly sampled features when building each tree colsample_bytree: 0.

6.

8. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling construction according to claim 1, wherein, The step 4 includes the following specific processes: Key features screened out in step 2 Behavioral features in the user preference dimension Correlation analysis of features, Pearson correlation coefficient (Pearson), Spearman rank correlation coefficient (Spearman) and analysis of variance (ANOVA) three indicators are used to calculate the relationship coefficient between features; The dynamic damping adjustment mechanism is introduced to adaptively learn the weight coefficients of the Pearson, Spearman and ANOVA three indexes, the calculation results of the three indexes are weighted and combined, a feature correlation matrix is constructed, and behavior features with significant correlation with key features are mined.

9. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling construction according to claim 8, wherein, The step 5 includes the following specific process: Based on the step 4, the features highly related to user churn in the preference dimension are identified, and the K-means algorithm is used to divide the users; specifically, the three feature data are input into the K-means, and the distance square sum of the cluster center to the sample in the cluster is minimized through continuous iteration optimization, the user division is completed, and the churn user portrait with group behavior characteristics is constructed; The step 6 includes the following specific process: According to the user churn risk level output by the step 3 and the user group portrait output by the step 5, the personalized recall strategy is matched; wherein different risk levels correspond to different push channels and different push frequencies, and different group portraits correspond to different push contents; High-risk users: daily mobile phone messages, 3-5 times of telephone per week, and push content customized according to the user portrait; Medium-risk users: daily APP pop-up window, 1-2 times of mobile phone message per week, and push content customized according to the user portrait; Low-risk users: 3-5 times of APP pop-up window per week, and push content customized according to the user portrait.

10. The dual-channel churn risk grading and intelligent recall method of predicting and user profiling according to any one of claims 1-9, characterized in that, Further including the following process: Step 7: introducing the proportional-integral-derivative PID control algorithm, generating a control signal by monitoring the current error, cumulative error and error trend between the actual recall rate and the target recall rate of the churn users, dynamically adjusting the churn risk threshold, and realizing adaptive optimization of risk classification.

Images