High-value user operation preservation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of communication operation, and provides a high-value user operation preservation method and system based on artificial intelligence, and the method comprises the steps: dividing a plurality of high-value user groups from stock users; for each high-value user group, constructing and training an exclusive loss risk prediction model; and inputting the multi-dimensional feature data of the target user into the loss risk prediction model of the affiliated customer group to obtain a loss probability of the target user and a key risk feature influencing user loss, the loss probability and the key risk feature of the target user being used for executing an operation retention strategy of the target user. According to the method and the device, an operator can perform targeted retention according to key risk characteristics influencing user loss, the retention success rate of high-value users is improved, and the marketing resource waste rate is reduced.

Description

Technical Field

[0001] This application relates to the field of communication operation technology, and in particular to a method and system for operating and retaining high-value users based on artificial intelligence. Background Technology

[0002] In the telecommunications industry, the conventional technical path for user churn prediction includes data collection, data preprocessing, model training, and strategy push. Specifically, the data layer focuses on basic user attributes and consumption behaviors; the prediction models primarily use logistic regression and single decision trees, with long training cycles and fixed parameters; and the strategy push application stage often employs a unified marketing model. For example, patent publication number CN114219668A describes a machine learning-based method for predicting telecommunications user churn. This method constructs a feature set using basic user information, call and data usage data, trains the model using a random forest algorithm, retrains it every six months, and in the application phase, classifies risk levels based on churn probability and pushes general package discounts. Patent publication number CN115082876B describes a method for retaining high-value users on internet platforms, using consumption amount and login frequency as core features, training a model based on the XGBoost algorithm, and pushing benefit reminders via SMS.

[0003] However, the mobile communications industry currently faces challenges such as intensified competition in the existing market and increasing pressure to churn high-value users. For example, other network operators are launching regionalized competition strategies targeting their core customer groups. While the aforementioned existing technologies provide a basic framework for user management for telecommunications operators, the accuracy of predicting churn risk still needs improvement. Summary of the Invention

[0004] To address the aforementioned problems in the prior art, this application provides a method and system for managing and retaining high-value users based on artificial intelligence, thereby improving the accuracy of predicting the risk of high-value user churn.

[0005] To achieve the above objectives, the technical solution adopted in this application is as follows: Firstly, this application provides a method for high-value user operation and retention based on artificial intelligence, including: Step S1: Identify multiple high-value user groups from existing users; Step S2: For each high-value user group, build and train a dedicated churn risk prediction model; Step S3: Input the multi-dimensional feature data of the target user into the churn risk prediction model of the customer group to obtain the churn probability of the target user and the key risk features affecting user churn. The churn probability and key risk features of the target user are used to implement the operation and retention strategy of the target user.

[0006] The beneficial effects of this application are as follows: This application identifies multiple high-value user groups and trains them to obtain exclusive churn risk prediction models, thereby significantly improving the prediction accuracy and coverage of high-value user churn risk; at the same time, it outputs key risk characteristics, enabling operators to carry out targeted retention based on the key risk characteristics affecting user churn, thereby improving the retention success rate of high-value users and reducing the waste rate of marketing resources.

[0007] Optionally, the churn risk prediction model adopts the LightGBM algorithm framework, and the key risk characteristics affecting user churn obtained in step S3 include: By analyzing the trained decision tree structure in the churn risk prediction model, the gain of each feature at the split node is calculated to identify the key risk features affecting user churn. The formula for calculating the gain is as follows: ; In the formula, k is the index of the dimensional feature data, T is the total number of decision trees in the model, and Loss is... before,t Loss after,t denoted as , representing the losses before and after the split of the t-th tree.

[0008] Optionally, step S3 further includes: The target user's overall value score is calculated using a weighted dynamic scoring function. This overall value score is also used to implement the target user's operational retention strategy. The weighted dynamic scoring function is as follows: Score= ; In the formula, x i w represents the feature data of the i-th dimension. i represents the weight coefficient of the feature data in the i-th dimension.

[0009] Optionally, after step S3, the method further includes: Based on the newly generated multi-dimensional feature data and operational feedback results, the churn risk prediction model is updated using a gradient descent update function, which is: ; In the formula, θ (t) To update the model parameters, η new The incremental learning rate is n, where n is the historical sample size. new To increase the sample size, L old , L new These are the loss gradients for historical samples and newly added samples, respectively.

[0010] As described above, daily model updates are achieved through incremental learning formulas, which further improves the prediction accuracy of high-value user churn risk, reduces the speed of early warning response, and adapts to the offline service coverage needs of operators.

[0011] Optionally, after step S3, the following steps are included: According to a preset period, the churn probability of the target user is fused with the prediction results of the auxiliary model according to the corresponding model weight coefficients to obtain the fused churn probability of the target user.

[0012] Optionally, the training process of the churn risk prediction model includes: Calculate the absolute value of the gradient of the current training samples and sort them. High-gradient samples are retained, and random samples are taken from the remaining low-gradient samples. The gradients of the sampled low-gradient samples are then weighted and compensated to obtain the final training samples. The high-gradient samples are the current training samples before a preset sorting ratio. The gradient of the final training samples is... for: ; In the formula, g i denoted as the gradient of the original sample, a is the preset sorting ratio of high gradient samples, and b is the random sampling ratio.

[0013] Optionally, the churn risk prediction model includes: The values ​​of continuous features in multi-dimensional feature data are mapped to corresponding discrete intervals to obtain discrete features; Numerical features are obtained by encoding the numerical values ​​that belong to categorical features in multidimensional feature data.

[0014] Optionally, the training process of the churn risk prediction model includes: For the t-th round of model training, the objective function is defined as the sum of the historical loss and the regularization term. The objective function is: ; ; In the formula, n is the sample size of the high-value user group, and y i For users' actual churn, These are the model predictions from the first t-1 rounds. Let l(·) be the output of the t-th decision tree, and l(·) be the logarithmic loss function. Let γ be the regularization term, T be the number of leaf nodes in the decision tree of the high-value user group, γ be the regularization coefficient, λ be the weight decay coefficient, and w be the weight decay coefficient. j Let be the weight of the j-th leaf node.

[0015] Optionally, the high-value user groups segmented in step S1 include at least the package upgrade customer group, the high-income stable customer group, and the border customer group; The multi-dimensional feature data includes at least the following dimensions: basic user profile, core integrated package, upgrade behavior characteristics, consumption and usage habits, internal and external marketing correlation, and risk retention correlation. Step S3 includes: The risk level is determined based on the churn probability of the target users, and an early warning is pushed out based on the risk level to remind the operations staff to implement the operation and retention strategy for the target users.

[0016] Secondly, this application provides a high-value user operation and retention system based on artificial intelligence, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the high-value user operation and retention method based on artificial intelligence provided in the first aspect.

[0017] The technical effects of the AI-based high-value user operation and retention system provided in the second aspect are described in the relevant description of the AI-based high-value user operation and retention method provided in the first aspect. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the main process of the high-value user operation and retention method based on artificial intelligence according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a high-value user operation and retention system based on artificial intelligence, as described in an embodiment of this application.

[0019] Explanation of reference numerals in the attached figures: 1. A high-value user operation and retention system based on artificial intelligence; 2. Processor; 3. Memory. Detailed Implementation

[0020] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0021] This application targets application scenarios for managing existing users of mobile operators, such as customer profiling, recommendation systems, and business forecasting. Existing technologies employ a unified model trained with uniform parameters and a standardized marketing approach, resulting in low accuracy in predicting user churn risk. This application improves the accuracy of predicting churn risk for high-value users by training churn risk prediction models specific to different customer groups, and simultaneously outputs key risk characteristics to increase the success rate of retaining high-value users.

[0022] The present application will now be described in conjunction with the accompanying drawings and specific embodiments.

[0023] Please refer to Figure 1 High-value user retention and operation methods based on artificial intelligence include: Step S1: Identify multiple high-value user groups from existing users.

[0024] This involves segmenting existing customers of telecommunications operators into several high-value customer groups based on service characteristics. In one example, these high-value customer groups include at least: (1) Customers who upgrade their packages: those who have purchased non-integrated packages, upgraded their packages this month, and whose package fees have increased by 30% compared to last month; (2) High-income and stable customer group: Customers whose consumption in the last three months of the previous year was greater than or equal to 150 yuan and who were online in the current month; (3) Border customers: Customers whose account opening documents are border residents' certificates or who live in the border area for a long time.

[0025] In other examples, high-value customer groups may also include government and enterprise customers, family customers, and so on.

[0026] Step S2: For each high-value user group, build and train a dedicated churn risk prediction model.

[0027] For model training and subsequent model prediction, based on the existing IT architecture, the following implementation environments are deployed in layers: (1) Provincial node: The municipal data center deploys HDFS storage system and Spark preprocessing module to support data processing throughout the province; (2) City / County Nodes: Lightweight training nodes are deployed in key cities to adapt to local customer group analysis; (3) Offline terminals: Dedicated terminals are configured at village service points and business halls and connected to the early warning system; (4) Visualization platform: Based on the existing operation dashboard, develop modules for different customer groups and regions, which can be viewed by three levels of operation personnel.

[0028] The churn risk prediction model employs the LightGBM algorithm framework. LightGBM (Light Gradient Boosting Machine) is a highly efficient gradient boosting tree ensemble learning algorithm. Its core advantages are fast training speed, low memory consumption, and high accuracy. It is widely used in machine learning tasks such as classification, regression, and ranking, and performs particularly well in big data scenarios. LightGBM is essentially an optimized implementation of gradient boosting trees, following the idea of ​​integrating multiple weak classifiers to form a strong classifier. Specifically, it includes: (1) Iterative generation of decision trees: Each time a new decision tree is trained, the goal is to fit the prediction error of the previous model and optimize the loss function through the gradient descent direction; (2) Integrating the results of all trees: The final prediction result is a weighted sum or classification of multiple trees, and the error is continuously reduced by improving the strategy.

[0029] During the training process, for each high-value user group, multi-dimensional feature data is obtained from mobile CRM (Mobile Customer Relationship Management), BOSS (Business & Operation Support System), business support, and offline service systems as samples. In one example, the multi-dimensional feature data includes at least: (1) User basic profile dimension Personal attributes: user age, gender, city / region, duration of network membership, customer star rating / membership level, and attribute city / rural / government / enterprise / campus.

[0030] Account attributes: whether it is registered with real name, whether the payment method is prepaid or postpaid, whether the account status is normal / out of balance / dormant, and the number of associated supplementary cards.

[0031] (2) Core dimensions of the integrated package Basic package information: current bundled package type, package activation date, last month's package monthly fee, monthly fee after this month's upgrade, and the change in package fee.

[0032] Service composition: broadband bandwidth level, mobile phone package minutes / data allowance / number of SMS messages, whether it includes value-added services such as membership and security.

[0033] (3) Upgrade behavior characteristics dimension Upgrade process information: Upgrade application channels include offline business halls / online APP / customer service hotline; application time is at the beginning / middle / end of the month; and whether it is an active application / marketing recommendation.

[0034] Historical behavior: Package change records in the past 6 months / 1 year, activation time of historical integrated services, and whether there were any records of single services being converted to integrated packages.

[0035] (4) Consumption and usage habits dimension Consumption characteristics: average monthly consumption amount in the past 3 months, bill composition, history of arrears, frequency and amount of recharge.

[0036] Usage behavior: Business activity levels such as actual mobile data usage, voice call duration usage rate, and border data usage ratio.

[0037] (5) Internal and external marketing linkage dimension Internal Marketing: Did you receive any marketing pushes or participate in any promotional activities before the upgrade?

[0038] External environment: broadband coverage in the area, price comparison of competitor packages, number of family members, and consumption level of surrounding business districts / neighborhoods.

[0039] (6) Risk retention related dimensions Risk of customer churn: Customer service inquiry records in the past 3 months, and historical intentions to disconnect / port numbers.

[0040] Retention potential: Changes in service usage rate in the first month after the upgrade, whether there are any extra charges outside the package, and whether family members are all operators' users.

[0041] After obtaining multi-dimensional feature data, data preprocessing is also required: (1) Cleaning: Remove abnormal network fluctuation data and mark rural customer groups recorded offline; (2) Standardization: Min-Max standardization is used to process numerical features, and One-Hot encoding is used to process categorical features, such as package type and customer group affiliation; (3) Feature derivation: Generate operator-specific combined features and output standardized datasets for different customer groups.

[0042] In this embodiment, historical data of high-value user groups are used as the training set. The model is trained according to the customer groups, and the parameters are optimized by 5-fold cross-validation to finally obtain the trained churn risk prediction model.

[0043] The trained churn risk prediction model includes: The values ​​of continuous features in multi-dimensional feature data are mapped to corresponding discrete intervals to obtain discrete features; Numerical features are obtained by encoding the numerical values ​​that belong to categorical features in multidimensional feature data.

[0044] Therefore, by discretizing continuous features, the periodicity of operator data can be adapted, significantly reducing memory usage.

[0045] Step S3: Input the multi-dimensional feature data of the target user into the churn risk prediction model of the customer group to obtain the churn probability of the target user and the key risk features affecting user churn. The churn probability and key risk features of the target user are used to implement the operation and retention strategy of the target user.

[0046] For each target user, the process begins by determining which high-value user group they belong to. Then, a churn risk prediction model for that high-value user group is used for exclusive identification. This not only yields the churn probability of the target user but also identifies key risk characteristics based on the impact of different data features on the final churn probability, providing a basis for subsequent operational retention strategies.

[0047] After obtaining the churn probability and key risk characteristics of each target user in each high-value user group, the following key operational groups and corresponding operational retention strategies can be integrated: (1) Potential package upgrade customer group: refers to the user group whose current package cannot match their actual usage needs, who have the willingness and ability to upgrade their consumption, and who can be guided to upgrade to a higher value package through precise operation.

[0048] (2) High-risk customer churn group among high-income stable users: core value user group with high consumption capacity, stable historical consumption and high ARPU value, but has shown signs of behavioral abnormalities and demand mismatch, and has a high possibility of churn.

[0049] (3) High traffic demand customer group among border users: This group of users has a continuous and high demand for traffic. The traffic usage scenarios of this group are complex and significantly affected by policies, geographical environment and international communication rules. It is necessary to formulate precise operation strategies in combination with cross-border communication characteristics and local needs.

[0050] Meanwhile, in this embodiment, step S3 includes: The risk level is determined based on the churn probability of the target users, and early warnings are pushed out based on the risk level to remind operations personnel to implement the target user retention strategy.

[0051] Different churn probabilities correspond to different risk levels. For example, among high-income, stable users, a churn probability ≥50% is considered high risk, 20-50% is medium risk, and the rest are low risk. Therefore, early warning information is pushed through three channels: the operator's BOSS system, the city branch's operation platform, and offline service point terminals. This ensures that early warning delays in rural areas are controlled within 2 hours, and in urban areas, within 15 minutes.

[0052] In one embodiment, the training process of the churn risk prediction model includes: Step S21: Calculate the absolute value of the gradient of the current training sample and sort it.

[0053] Step S22: Retain high-gradient samples, randomly sample from the remaining low-gradient samples, and perform weighted compensation on the gradients of the sampled low-gradient samples to obtain the final training samples. The high-gradient samples are the current training samples before the preset sorting ratio. The gradient of the final training samples is... for: ; In the formula, g i denoted as the gradient of the original sample, a is the preset sorting ratio of high gradient samples, and b is the random sampling ratio.

[0054] The preset sorting ratio of high-gradient samples and the subsequent random sampling ratio can be set according to requirements. For example, if a is 0.2 and b is 0.1, then the weight of low-gradient samples is 8, ensuring that the sample distribution after sampling fits the original distribution.

[0055] Therefore, this embodiment uses a gradient weighted calculation formula to screen high-value samples, which not only improves processing efficiency by reducing the sample size, but also reduces the impact of sample imbalance in border customer groups, ultimately improving the accuracy of model prediction.

[0056] Step S23: For the t-th round of model training, the objective function is defined as the sum of the historical loss and the regularization term. The objective function is: ; ; In the formula, n is the sample size of the high-value user group, and y i For users' actual churn, These are the model predictions from the first t-1 rounds. Let l(·) be the output of the t-th decision tree, and l(·) be the logarithmic loss function. Let γ be the regularization term, T be the number of leaf nodes in the decision tree of the high-value user group, γ be the regularization coefficient, λ be the weight decay coefficient, and w be the weight decay coefficient. j Let be the weight of the j-th leaf node.

[0057] In this context, step S23 serves as the objective function for model training, which is iteratively optimized using a gradient boosting tree. In other embodiments, the tree depth, number of leaves, and L1 / L2 regularization can also be controlled, such as a learning rate of 0.05, a tree depth of 8-10, and 200-250 leaf nodes, to prevent overfitting and adapt to scenarios where the sample distribution is relatively concentrated after the customer groups are segmented in this application.

[0058] In one embodiment, the key risk characteristics affecting user churn obtained in step S3 include: By analyzing the pre-trained decision tree structure in the churn risk prediction model, the gain of each feature at the split node is calculated to identify the key risk features affecting user churn. The formula for calculating the gain is as follows: ; In the formula, k is the index of the dimensional feature data, T is the total number of decision trees in the model, and Loss is... before,t Loss after,t denoted as , representing the losses before and after the split of the t-th tree.

[0059] Among them, the model calculation method uses gain index quantification features to identify key factors of customer churn for operators and outputs key risk features such as Top 3 or Top 5, providing a basis for subsequent intervention strategies. For example, border customers may be involved in cross-border roaming and cross-network inquiries, while customers who upgrade their plans may be involved in upgrade intentions, consumption feedback, etc.

[0060] In one embodiment, step S3 further includes: A weighted dynamic scoring function is used to calculate the overall value score of the target user. This overall value score is also used to implement the target user's operational retention strategy. The weighted dynamic scoring function is as follows: Score= ; In the formula, x i w represents the feature data of the i-th dimension. i represents the weight coefficient of the feature data in the i-th dimension.

[0061] Among them, dimensional feature data such as average monthly consumption, online time, and package replacement cycle, etc., can be trained together when training the churn risk prediction model. They can then be dynamically adjusted according to changes in user behavior. The final calculated user value score is used to accurately formulate operation and retention strategies.

[0062] In one embodiment, after step S3 of the above embodiments, the method further includes: Step S4: Based on the newly generated multi-dimensional feature data and operational feedback results, update the churn risk prediction model using the gradient descent update function. The gradient descent update function is as follows: ; In the formula, θ (t) To update the model parameters, η new The incremental learning rate is n, where n is the historical sample size. new To increase the sample size, L old , L new These are the loss gradients for historical samples and newly added samples, respectively.

[0063] That is, compared with the traditional full update, this embodiment adopts gradient descent update, which reduces the model update time from 8 hours to 1.5 hours for a new data volume of 50GB per day.

[0064] Meanwhile, after step S3, the sea includes: According to the preset period, the churn probability of the target user is fused with the prediction results of the auxiliary model according to the corresponding model weight coefficients to obtain the fused churn probability of the target user.

[0065] The auxiliary model can be an existing model on the market, such as the XGBoost model. The model weight coefficients vary depending on the customer group and are merged and updated quarterly. For example, the model weight coefficient of the churn risk prediction model in the potential upgrade customer group is 0.7, and the model weight coefficient of the churn risk prediction model in the border customer group is 0.8. Different models are merged and calculated every quarter, and the churn probability after merging is compared with the original churn probability to update the churn risk prediction model, thereby further improving the accuracy of the model output.

[0066] This embodiment also includes updating different dimensional feature data monthly. Therefore, this embodiment forms a complete dynamic self-training optimization scheme for the model to match the operator's business rhythm, including daily model updates, monthly feature iterations, and quarterly model fusion. The final output is an optimized churn risk prediction model that differentiates between different customer groups, further improving the prediction accuracy of high-value user churn risk, reducing warning response speed, and adapting to the operator's offline service coverage needs.

[0067] In addition, different customer group differentiation adaptation branches can be set in different churn risk prediction models to improve the model's specific identification capabilities. For example: adding a package upgrade query dimension for potential upgrade customers to ensure the accuracy of the model's output customer groups; adding a business inactivity day dimension for high-risk churn customers among high-income stable users and outputting it in a timely manner to ensure that the output customer groups can be maintained in a timely manner; adding roaming data verification for border customers to ensure that the model fits the business characteristics of each customer group.

[0068] In summary, the various embodiments of this application have the following advantages: (1) Accurate risk identification: The high-value customer group segmentation modeling method of telecommunications operators divides the customer group according to industry characteristics and conducts targeted training. Combined with formula calculation to optimize the model, the model prediction accuracy and coverage are greatly improved, and the regional loss signal such as the tendency of border customers to lose due to cross-border traffic discounts is accurately captured.

[0069] (2) Operational efficiency optimization: By combining online early warning and offline collaboration with the mobile channel layout, and combining the key factors output by feature importance calculation, we can improve the operational capabilities in rural and border areas, increase the retention rate of high-value users, and reduce the waste rate of marketing resources.

[0070] (3) Improved response efficiency: Daily model updates are achieved through incremental learning formulas, shortening the early warning response speed in rural / border areas to within 2 hours and in urban areas to within 15 minutes, adapting to the offline service coverage needs of operators.

[0071] (4) Stable and effective model: The dynamic self-training mechanism that matches the operator's business rhythm ensures the long-term effectiveness of the model.

[0072] Please refer to Figure 2 The embodiments of this application also provide a high-value user operation and retention system 1 based on artificial intelligence, including a memory 3, a processor 2, and a computer program stored on the memory 3 and executable on the processor 2. When the processor 2 executes the computer program, it implements the steps in the above embodiments.

[0073] Since the systems / devices described in the above embodiments of this application are systems / devices used to implement the methods of the above embodiments of this application, those skilled in the art can understand the specific structure and modifications of the system / devices based on the methods described in the above embodiments of this application, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of this application fall within the scope of protection of this application.

[0074] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0075] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0076] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This application can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0077] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0078] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0079] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, then this application should also include these modifications and variations.

Claims

1. A high-value user operation and retention method based on artificial intelligence, characterized in that, include: Step S1: Identify multiple high-value user groups from existing users; Step S2: For each high-value user group, build and train a dedicated churn risk prediction model; Step S3: Input the multi-dimensional feature data of the target user into the churn risk prediction model of the customer group to obtain the churn probability of the target user and the key risk features affecting user churn. The churn probability and key risk features of the target user are used to implement the operation and retention strategy of the target user.

2. The method for operating and retaining high-value users based on artificial intelligence according to claim 1, characterized in that, The churn risk prediction model adopts the LightGBM algorithm framework. Step S3 obtains the key risk characteristics affecting user churn, including: By analyzing the trained decision tree structure in the churn risk prediction model, the gain of each feature at the split node is calculated to identify the key risk features affecting user churn. The formula for calculating the gain is as follows: ; In the formula, k is the index of the dimensional feature data, T is the total number of decision trees in the model, and Loss is... before,t Loss after,t denoted as , representing the losses before and after the split of the t-th tree.

3. The method for operating and retaining high-value users based on artificial intelligence according to claim 1, characterized in that, Step S3 further includes: The target user's overall value score is calculated using a weighted dynamic scoring function. This overall value score is also used to implement the target user's operational retention strategy. The weighted dynamic scoring function is as follows: Score= ; In the formula, x i w represents the feature data of the i-th dimension. i represents the weight coefficient of the feature data in the i-th dimension.

4. The method for operating and retaining high-value users based on artificial intelligence according to claim 1, characterized in that, The step S3 is followed by: Based on the newly generated multi-dimensional feature data and operational feedback results, the churn risk prediction model is updated using a gradient descent update function, which is: ； In the formula, θ (t) To update the model parameters, η new The incremental learning rate is n, where n is the historical sample size. new To increase the sample size, L old , L new These are the loss gradients for historical samples and newly added samples, respectively.

5. The method for operating and retaining high-value users based on artificial intelligence according to claim 1, characterized in that, After step S3, the following is included: According to a preset period, the churn probability of the target user is fused with the prediction results of the auxiliary model according to the corresponding model weight coefficients to obtain the fused churn probability of the target user.

6. The method for operating and retaining high-value users based on artificial intelligence according to any one of claims 1 to 5, characterized in that, The training process of the churn risk prediction model includes: Calculate the absolute value of the gradient of the current training samples and sort them. High-gradient samples are retained, and random samples are taken from the remaining low-gradient samples. The gradients of the sampled low-gradient samples are then weighted and compensated to obtain the final training samples. The high-gradient samples are the current training samples before a preset sorting ratio. The gradient of the final training samples is... for: ； In the formula, g i denoted as the gradient of the original sample, a is the preset sorting ratio of high gradient samples, and b is the random sampling ratio.

7. The method for operating and retaining high-value users based on artificial intelligence according to any one of claims 1 to 5, characterized in that, The churn risk prediction model includes: The values ​​of continuous features in multi-dimensional feature data are mapped to corresponding discrete intervals to obtain discrete features; Numerical features are obtained by encoding the numerical values ​​that belong to categorical features in multidimensional feature data.

8. The method for operating and retaining high-value users based on artificial intelligence according to any one of claims 1 to 5, characterized in that, The training process of the churn risk prediction model includes: For the t-th round of model training, the objective function is defined as the sum of the historical loss and the regularization term. The objective function is: ； ； In the formula, n is the sample size of the high-value user group, and y i For users' true churn tags, These are the model predictions from the first t-1 rounds. Let l(·) be the output of the t-th decision tree, and l(·) be the logarithmic loss function. Let γ be the regularization term, T be the number of leaf nodes in the decision tree of the high-value user group, γ be the regularization coefficient, λ be the weight decay coefficient, and w be the weight decay coefficient. j Let be the weight of the j-th leaf node.

9. The method for operating and retaining high-value users based on artificial intelligence according to any one of claims 1 to 5, characterized in that, The high-value user groups identified in step S1 include at least the package upgrade customers, high-income stable customers, and border customers. The multi-dimensional feature data includes at least the following dimensions: basic user profile, core integrated package, upgrade behavior characteristics, consumption and usage habits, internal and external marketing correlation, and risk retention correlation. Step S3 includes: The risk level is determined based on the churn probability of the target users, and an early warning is pushed out based on the risk level to remind the operations staff to implement the operation and retention strategy for the target users.

10. A high-value user operation and retention system based on artificial intelligence, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the high-value user operation and retention method based on artificial intelligence as described in any one of claims 1 to 9.

Images