High-risk user loss early warning and retention method and system fusing CNN and Informer

Abstract

The invention discloses a CNN and Informer fused high-risk user loss early warning and retention method and system, and belongs to the technical field of big data, and the method comprises the steps: obtaining the multi-source data of each user, and constructing the static features, daily dynamic features, and a time sequence feature matrix of a set time period of the user through the multi-source data; constructing a CNN and Informer fused deep learning model, and predicting the loss probability of the user by using the trained deep learning model based on the static features and daily dynamic features of the user and the time sequence feature matrix of the time period set by the user; constructing an RFL model, and obtaining a user value score by using the RFL model and the multi-source data; and in combination with the value scores and loss probabilities of all the users, grouping all the users by using a Kmenas model, outputting user grouping results and grouping portraits, and performing early warning and retention. According to the method, potential high-risk lost users can be accurately identified, and the retention cost is reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of big data, and particularly relates to a high-risk user loss early warning and retention method and system fusing CNN and Informer. BACKGROUND

[0002] In the era of information development, telecommunication operators are facing fierce market competition, the user market of traditional services is close to saturation, and the problem of user loss is becoming increasingly serious, so new and higher requirements are put forward for stock user management. At present, the user retention of the value-added products of telecommunication operators is mainly based on data analysis and mining of business experience, and after the users are circled, a unified retention strategy is implemented. This mode cannot mine the implicit information in the time series behavior data of users, and has problems such as high calculation cost, slow user group update, inaccurate positioning, short business intervention retention time window, etc. In addition, the unified retention measures without distinguishing the value and characteristics of users will lead to high retention cost and low retention success rate. Therefore, how to accurately identify potential high-risk loss users in advance, reduce retention cost and improve user retention rate has become an important problem to be solved. SUMMARY

[0003] The application provides a high-risk user loss early warning and retention method and system fusing CNN and Informer, which can accurately identify potential high-risk loss users and reduce retention cost.

[0004] The application provides the following technical solutions:

[0005] In a first aspect, a high-risk user loss early warning and retention method fusing CNN and Informer is provided, comprising:

[0006] Obtaining multi-source data of each user, and constructing static features, daily dynamic features and time series feature matrices of a set time period of the user by using the multi-source data;

[0007] Constructing a deep learning model fusing CNN and Informer, and predicting the loss probability of the user by using the trained deep learning model based on the static features, daily dynamic features and time series feature matrices of the set time period of the user;

[0008] Constructing an RFL model, and obtaining a user value score by using the RFL model and the multi-source data;

[0009] Combining all user value scores and loss probabilities, grouping all users by using a Kmenas model, outputting user grouping results and grouping portraits, and performing early warning and retention.

[0010] Optionally, the multi-source data comprises: user basic attribute data, user real-time browsing data in the APP, click data, user subscription, use and unsubscribing behavior data of value-added services and associated services in the APP, and user complaint work orders, and traffic usage data;

[0011] The daily dynamic features comprise: user online behavior, user browsing, click logs, user subscription behavior, whether the user newly subscribes value-added services, whether the user uses value-added services, whether the user uses value-added services, the type of value-added services used by the user, and APP active information.

[0012] The static features comprise: user basic information, user traffic usage data, complaint service types in user complaint work orders, and user preference click and browsing data in the APP.

[0013] Optionally, a time sequence feature matrix of the user in the set time period is constructed, specifically: the daily dynamic features extracted from the user are sorted in time sequence, a sliding window mechanism is used to construct the time sequence feature matrix of the user in the set time period with a set number of days as a window length.

[0014] Optionally, the deep learning model fused by the CNN and the Informer is in a series structure, and the deep learning model comprises: a CNN model layer, an Informer model layer, and a classifier.

[0015] Optionally, the static features of the user, the daily dynamic features, and the time sequence feature matrix of the user in the set time period are used to predict the churn probability of the user by using the trained deep learning model, specifically:

[0016] The static features and the daily dynamic features of the user are input into the CNN model layer to extract effective features.

[0017] After the effective features extracted by the CNN model layer and the time sequence feature matrix of the user in the set time period are aligned and fused, they are input into the Informer model layer, and the classifier is used to predict whether the user chums and the probability of the user churning.

[0018] Optionally, the RFL model is constructed, and the value score of the user is obtained by using the RFL model and the multi-source data, specifically: based on the RFM model, the frequency of the user using value-added services, the number of renewals, and the subscription of value-added services are taken as key indicators, and the RFL model is established.

[0019] The key indicators are weighted and summed by using a weighted average method to obtain the value score of the user in the value-added service field.

[0020] Optionally, the combination of all user value scores and churn probability, using Kmenas model for grouping all users, output user grouping results and grouping portrait, and carry out early warning and retention, specifically:

[0021] According to the value score of the user, the user is divided into high-value, medium-value or low-value user; According to the user's churn probability, the user is divided into high churn, medium churn or low churn user; According to the user's use frequency of value-added services, the user is divided into high frequency, medium frequency and low frequency user;

[0022] Based on value, loss and use frequency, all user types are obtained, and one typical user of each type is selected as the initial center point of the Kmenas model;

[0023] Based on the value score of the user, the use frequency of the user and the churn probability of the user, the distance between other users and all center points is obtained, and each user is assigned to the nearest center point;

[0024] The center point of the user group is updated by using the mean value of the subscription times, use frequency and churn probability of all users in each user group;

[0025] The center point is continuously updated to obtain the final user grouping result, and all user groups and user group portraits after grouping are output;

[0026] According to the value, loss and use frequency of the user group, a preset level of early warning is issued, and a retention scheme meeting the current user group portrait is output according to the preset retention strategy.

[0027] In the second aspect, a high-risk user churn early warning and retention system combining CNN and Informer is provided, comprising:

[0028] The data acquisition module acquires multi-source data of each user, and constructs static features, daily dynamic features and time sequence feature matrix of the user in a set time period by using the multi-source data;

[0029] The loss prediction module constructs a deep learning model combining CNN and Informer, and predicts the user's churn probability by using the trained deep learning model based on the user's static features, daily dynamic features and time sequence feature matrix of the user in a set time period;

[0030] The value evaluation module constructs an RFL model, and acquires the user value score by using the RFL model and multi-source data;

[0031] The grouping early warning module combines all user value scores and churn probability, uses Kmenas model to group all users, outputs user grouping results and grouping portrait, and carries out early warning and retention.

[0032] In a third aspect, a computer device is provided, comprising a processor and a memory; wherein the processor implements the steps of the high-risk user loss early warning and retention method of fusing CNN and Informer according to any one of the first aspect when executing the computer program stored in the memory.

[0033] In a fourth aspect, a computer readable storage medium is provided for storing a computer program; the computer program is executed by a processor to implement the steps of the high-risk user loss early warning and retention method of fusing CNN and Informer according to any one of the first aspect.

[0034] Compared with the prior art, the present application has the following beneficial effects:

[0035] (1) The present application predicts high-risk loss users of value-added products by establishing a CNN and Informer fusion deep learning model, effectively captures the short-term dependence of dynamic features and the meaning of static features by using the local feature processing capability of CNN, and then processes long time series data by using the Informer model, which is complementary in features, obtains the meaning and association hidden by key feature data on time series of different lengths, efficiently learns the change of user behavior trajectory, more accurately identifies the user loss tendency, and improves the user retention effect.

[0036] (2) The present application obtains user daily and monthly features according to the characteristics of value-added services, and constructs a user set time period time series feature matrix by using a sliding window mechanism, provides daily updated user behavior data for the fusion model, speeds up the model update speed, increases the time of business intervention retention, and reduces the probability that the user has become a loss user.

[0037] (3) The present application constructs an RFL model based on the characteristics of value-added services to obtain user value indicators, and establishes a Kmeans clustering model for high-risk loss users in combination with user loss probability to refine the grouping, can formulate personalized retention schemes according to the group portrait, reduces the retention cost, and improves the retention success rate. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is the step flow chart of the high-risk user loss early warning and retention method of fusing CNN and Informer of the present application;

[0039] Figure 2 is the process flow chart of constructing user static features, daily dynamic features and time series feature matrix of the present application;

[0040] Figure 3 is the process flow chart of the CNN and Informer fusion deep learning model for high-risk loss user prediction of the present application;

[0041] Figure 4is a flow chart of grouping users by constructing an RFL model and Kmeans clustering according to the application. DETAILED DESCRIPTION

[0042] The application will be further described below with reference to the drawings. The following examples are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application. It should be noted that the terms "comprising" and any variations thereof in the specification and claims of the application and the above drawings are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0043] Embodiment 1

[0044] As shown in the figure, a high-risk user loss early warning and retention method fusing CNN and Informer is provided, including the following steps: Figure 1 S1: Obtain multi-source data of each user, and construct static features, daily dynamic features and time sequence feature matrix of the user by using the multi-source data.

[0045]

[0046] As shown in the figure, the multi-source data includes: user basic attribute data, real-time browsing data of the user in the APP, click data, behavior data of the user subscribing, using and unsubscribing value-added services and associated services in the APP, and user complaint work order and traffic usage data. Figure 2

[0047] After the multi-source data is obtained, preprocessing is needed, and basic analysis of the distribution, type and mean value of the multi-source data is needed, and significant outliers are removed, and then missing value filling, format unification, coding and normalization operations are performed to generate a standardized data set.

[0048] The daily dynamic features include: user online behavior, user browsing, click log, user subscription behavior, whether the user newly subscribes value-added services, whether the user uses value-added services, whether the user uses value-added services, the type of value-added services used, and APP active information.

[0049] The static features include: user basic information (user age, gender and location), user traffic usage data, complaint service type in the user complaint work order, and preferred click and browsing data of the user in the APP.

[0050] ​​In some other embodiments, static features and dynamic features can be cross-combined to generate cross features, such as subscription frequency of users in different age groups, etc.; or derivative features can be generated according to existing features, such as user usage growth rate in three months, number of service types subscribed by the user, etc., and the cross features and the derivative features are taken as static features.

[0051] A time sequence feature matrix of a user setting time period is constructed, specifically, the daily dynamic features extracted by the user are sorted in time sequence, and a sliding window mechanism is used to construct the time sequence feature matrix of the user setting time period with a window length of a set number of days, that is, the daily dynamic features of the latest day are updated in the time sequence feature matrix every day, and the daily dynamic features corresponding to the earliest day in the time sequence feature matrix are removed; as an option, a T+30 time sequence feature matrix is constructed, that is, historical data of the user's APP browsing, clicking, etc. for one month is extracted, 8 features such as the most browsed module and the most clicked module of the user are constructed, and the feature data of each user within 30 days is sorted in time sequence, and a sliding window mechanism is used to construct a user T+30 time sequence matrix with a window length of a day, and the latest one-day data is added to the matrix every day, and the earliest one-day data is removed.

[0052] S2: A deep learning model combining CNN and Informer is constructed, and based on the static features, daily dynamic features and time sequence feature matrix of the user setting time period of the user, the trained deep learning model is used to predict the churn probability of the user.

[0053] The deep learning model combining CNN and Informer is a series structure, and the deep learning model includes a CNN model layer, an Informer model layer and a classifier, that is, the convolutional layer of CNN is used to extract local effective features, remove redundant features and reduce calculation amount, and the Informer encoder is used to capture long-term dependence in time sequence data, and the local features extracted by CNN and the overall information of time sequence are considered for more comprehensive analysis.

[0054] CNN (Convolutional Neural Network) is a deep learning model that can efficiently extract local features and improve the generalization ability of the model by using the weight sharing mechanism. CNN has excellent performance on large-scale data sets due to its powerful feature learning ability and efficient parallel computing characteristics, and runs fast. CNN has excellent performance in prediction tasks due to its comprehensive technical advantages.

[0055] Informer is a long time series prediction deep learning model based on Transformer architecture, which introduces sparse self-attention mechanism, focuses on data segments with key information, reduces the time and space complexity of traditional self-attention mechanism, and can efficiently process long sequence data. It uses bidirectional attention to consider all information, captures short-term fluctuations while focusing on long-term trends, and has an advantage in long sequence time series tasks.

[0056] In this embodiment, as an option, a CNN model with 4 convolutional layers and max pooling is established using PyTorch. The important features of the user's full static and dynamic features are automatically learned through convolutional operations, and effective features are selected. Then an Informer model with 4 encoder layers, each with 6 heads, a prediction length of 1, and an encoding length of 30 is established. The input of the model is the feature matrix selected by CNN combined with the time series feature matrix. It not only has a rich data source, including high-quality features selected by CNN and time series information, but also can learn user behavior trajectory changes and improve prediction accuracy. In addition, it eliminates noise and redundant features, has faster training speed and better generalization ability.

[0057] The positive and negative sample datasets are obtained based on whether the user is lost or not, and SMOTE sampling is used to solve the problem of too large negative sample ratio. The CNN model training set and test set are generated using static features and daily dynamic features.

[0058] The effective feature vector extracted by the CNN network is spliced with the time series feature matrix obtained in step S1 through the attention guide mechanism to construct the training set and test set of the Informer model.

[0059] As shown in Figure 3 , the training of the CNN model layer and the Informer model can refer to the prior art. Specifically, the CNN training set can be input into the CNN model for training, and the accuracy and other indicators can be optimized by adjusting the parameters. After training, the test set is input into the model to complete the prediction process. The dimensions of the CNN features are expanded, and the 25 features are converted into a (1, 25) matrix to match the time series dimensions. Then, the 30x8 dimensional time series feature matrix of step 3 is used for attention guide splicing, and the Informer training set and test set are established based on whether the user is lost or not. The training set of the Informer model is input into the Informer model, and the model performance is observed. The model parameters are adjusted in real time for training optimization. The Informer test set is input into the trained model for prediction, and the user's loss or loss probability is output.

[0060] The loss function used by the CNN and Informer model of the present application is MSE (Mean Squared Error), and the optimizer is Adam.

[0061] The specific step of step S2 of making a prediction is: inputting the static features and daily dynamic features of the user into the CNN model layer to extract effective features; after aligning and fusing the effective features extracted by the CNN model layer and the time series feature matrix of the user in the set time period, inputting them into the Informer model layer, and predicting whether the user will churn and the probability of churning through the classifier.

[0062] The CNN features are dimensionally expanded to convert 25 features into a (1, 25) matrix to match the time series dimension. Then, the 30x8 dimensional time series feature matrix of step 3 is taken and attention-guided splicing is performed, specifically: a 25x8 trainable parameter matrix W is established, and Xavier uniform initialization is adopted; the CNN features are multiplied by the W matrix to obtain a projection from 25 dimensions to 8 dimensions, the projection result is multiplied by the transpose of the time series feature matrix, and the output 30-dimensional vector is scaled and then softmax normalized; feature reconstruction is performed, the 30-dimensional weight vector output by the softmax is multiplied by the time series feature matrix, each 8-dimensional time series feature channel is multiplied by the weight of the corresponding time step to obtain the final 8-dimensional time series representation; finally, the original 25-dimensional CNN features and the 8-dimensional weighted time series features are spliced, and the 33-dimensional fused features are processed by LayerNorm.

[0063] S3: Constructing an RFL model and using the RFL model and multi-source data to obtain a user value score.

[0064] As shown in Figure 4 Based on the RFM model, the frequency of using value-added services (representing user activity), the number of renewals (representing user loyalty), and the subscription of value-added services (representing user consumption ability in value-added services) are taken as key indicators to establish an RFL model; a weighted average method is used to distribute weights and perform weighted summation and averaging on the key indicators to obtain a value score of the user in the value-added service field.

[0065] S4: Using the Kmenas model to group all users in combination with all user value scores and churn probabilities, outputting user grouping results and grouping portraits, and performing early warning and retention.

[0066] Step S4, specifically:

[0067] S41: According to the value score of the user, the user is divided into a high-value, medium-value, or low-value user; according to the churn probability of the user, the user is divided into a high-churn, medium-churn, or low-churn user; according to the frequency of using value-added services by the user, the user is divided into a high-frequency, medium-frequency, or low-frequency user;

[0068] S42: Obtain all user types based on value, churn and frequency of use, and select one typical user for each type as the initial center point of the Kmeans model;

[0069] S43: Obtain the distance between other users and all center points based on user value score, user frequency of use and user churn probability, and assign each user to the nearest center point;

[0070] S44: Update the center point of the user group using the mean of the number of subscriptions, frequency of use and churn probability of all users in each user group;

[0071] S45: Continuously update the center point to obtain the final user grouping result, and output all user groups after grouping and user group portraits.

[0072] S46: Issue a preset level of warning according to the value, churn and frequency of use of the user group, and output a retention plan that meets the current user group portrait according to the preset retention strategy.

[0073] Kmeans is a clustering analysis method that divides a data set into K clusters, so that the data points in each cluster are as similar as possible to each other, and the data points in different clusters are as different as possible, which is suitable for customer classification and grouping. The Kmeans algorithm can refer to the prior art.

[0074] The warning level can be warned according to the value and churn of the user group as the main factor, the type of warning can be adjusted according to the actual type, the retention plan can be specified according to the expert experience, and the specific can refer to the prior art.

[0075] At present, the mode of value-added products for user subscription retention is to obtain user groups through data analysis or mining according to business operation experience, which is far behind the business decision demand and user actual demand in terms of data providing speed, quality and range. The present application combines CNN and Informer deep learning model to predict the subscription tendency of users to value-added products, improves the prediction accuracy through the fusion model, reduces the artificial input cost, and improves the data update speed. The present application constructs a user T+30 time sequence feature matrix, combines user daily and monthly dimension feature data for training and prediction, and updates the prediction fluctuation more quickly, effectively solves the problem of long waiting window period caused by the timeliness of the update result of the traditional data mining method, and gives business personnel more time to intervene in retention, improves the retention success rate. The present application adopts a customized RFL model to evaluate user value indicators for value-added products, uses a Kmeans clustering algorithm to group users in combination with indicators and churn probability, formulates individualized retention strategies for user group portraits, reduces retention cost, and improves operation efficiency.

[0076] Example 2

[0077] A high-risk user loss warning and retention system fusing CNN and Informer, comprising:

[0078] A data acquisition module acquires multi-source data of each user, and constructs static features, daily dynamic features and time sequence feature matrices of a set time period of the user by using the multi-source data;

[0079] A loss prediction module constructs a deep learning model fusing CNN and Informer, and predicts the loss probability of the user by using the trained deep learning model based on the static features, daily dynamic features and time sequence feature matrices of the set time period of the user;

[0080] A value evaluation module constructs an RFL model, and acquires the value score of the user by using the RFL model and the multi-source data;

[0081] A grouping warning module groups all users by using a Kmenas model in combination with the value scores and loss probabilities of all users, outputs the user grouping result and grouping portrait, and performs warning and retention.

[0082] The more specific process of the above method can refer to the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0083] Embodiment 3

[0084] The application provides a computer device, comprising a processor and a memory; wherein the processor realizes the steps of the high-risk user loss warning and retention method fusing CNN and Informer when executing the computer program saved in the memory.

[0085] The more specific process of the above method can refer to the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0086] Embodiment 4

[0087] The application provides a computer readable storage medium for storing a computer program; the computer program realizes the steps of the high-risk user loss warning and retention method fusing CNN and Informer when executed by the processor.

[0088] The more specific process of the above method can refer to the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0089] The various embodiments described in the specification are progressive in nature, and each embodiment highlights the differences from other embodiments. The same or similar parts among the various embodiments can be mutually referred to. For the system, device and storage medium disclosed by the embodiments, since they correspond to the method disclosed by the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0090] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present application can be realized by means of software and necessary general hardware platforms. Based on such understanding, the technical solutions in the embodiments of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in the various embodiments or some parts of the embodiments of the present application.

[0091] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for those skilled in the art, some improvements and refinements without departing from the principles of the present application shall be considered as the protection scope of the present application.

Claims

1. A high-risk user loss warning and retention method fusing a CNN and an Informer, characterized in that, The application relates to a user churn prediction method and device. The application comprises the following steps: acquiring multi-source data of each user and constructing static features, daily dynamic features and time sequence feature matrices of the user in a set time period by using the multi-source data; constructing a deep learning model combining CNN and Informer, and predicting the user churn probability by using the trained deep learning model based on the static features, daily dynamic features and time sequence feature matrices of the user in a set time period; constructing an RFL model and acquiring the user value score by using the RFL model and the multi-source data; 2. The high-risk user churn early warning and retention method of fusing CNN and Informer according to claim 1, characterized in that, combining all user value scores and churn probabilities, grouping all users by using a Kmenas model, outputting user grouping results and grouping portraits, and performing early warning and retention. The multi-source data comprises user basic attribute data, real-time browsing data, click data of the user in an APP, behavior data of the user subscribing, using and unsubscribing value-added services and associated services in the APP, and user complaint work orders and traffic usage data. The daily dynamic features comprise user online behavior, user browsing, click logs, user subscription behavior, whether the user newly subscribes value-added services, whether the user uses value-added services, whether the user uses value-added services, the type of value-added services used by the user, and APP active information.

3. The high-risk user churn early warning and retention method of fusing CNN and Informer according to claim 1, characterized in that, The static features comprise user basic information, user traffic usage data, complaint service types in user complaint work orders and preferred click and browsing data of the user in the APP.

4. The high-risk user churn early warning and retention method of fusing CNN and Informer according to claim 1, characterized in that, The time sequence feature matrix of the user in a set time period is constructed by sorting the daily dynamic features of the user in time sequence and using a sliding window mechanism to construct the time sequence feature matrix of the user in a set time period with a set number of days as the window length.

5. The high-risk user churn early warning and retention method of fusing CNN and Informer according to claim 4, characterized in that, The deep learning model combining CNN and Informer is in a series structure, and the deep learning model comprises a CNN model layer, an Informer model layer and a classifier. The deep learning model is trained based on the static features, daily dynamic features and time sequence feature matrices of the user in a set time period, and the user churn probability is predicted by using the trained deep learning model, specifically as follows: the static features and daily dynamic features of the user are input into the CNN model layer to extract effective features; 6. The high-risk user churn early warning and retention method of fusing CNN and Informer according to claim 1, characterized in that, the effective features extracted by the CNN model layer and the time sequence feature matrix of the user in a set time period are aligned and fused, and then input into the Informer model layer, and the classifier is used to predict whether the user churns and the probability of the user churning. The RFL model is constructed, and the user value score is acquired by using the RFL model and the multi-source data, specifically as follows: based on the RFM model, the frequency of the user using value-added services, the number of renewals and the subscription of value-added services are taken as key indicators to establish the RFL model; 7. The high-risk user churn early warning and retention method of fusing CNN and Informer according to claim 1, characterized in that, the key indicators are weighted and summed by using a weighted average method to obtain the value score of the user in the value-added service field. The Kmenas model is used to group all users by combining all user value scores and churn probabilities, output user grouping results and grouping portraits, and perform early warning and retention, specifically as follows: According to the user value score, the user is divided into a high-value, medium-value or low-value user; according to the user churn probability, the user is divided into a high-churn, medium-churn or low-churn user; According to the frequency of the user using the value-added service, the user is divided into a high-frequency, medium-frequency or low-frequency user; Based on the value, churn and frequency of use, all user types are obtained, and one typical user of each type is selected as the initial center point of the Kmenas model; Based on the user value score, user usage frequency and user churn probability, the distance between other users and all center points is obtained, and each user is assigned to the nearest center point; The center point of each user group is updated using the mean of the number of subscriptions, usage frequency and churn probability of all users in the user group; The center point is continuously updated to obtain the final user grouping result, and all user groups and user group portraits after grouping are output; According to the value, churn and usage frequency of the user group, a preset level of early warning is issued, and a retention scheme that meets the current user group portrait is output according to the preset retention strategy.

8. A high-risk user loss early warning and retention system fusing a CNN and an Informer, characterized in that, It includes: A data acquisition module acquires multi-source data of each user, and constructs static features, daily dynamic features and time series feature matrices of a set time period of the user using the multi-source data; A churn prediction module constructs a deep learning model combining CNN and Informer, and predicts the user's churn probability based on the user's static features, daily dynamic features and time series feature matrices of a set time period of the user using the trained deep learning model; A value assessment module constructs an RFL model and obtains a user value score using the RFL model and multi-source data; A grouping early warning module groups all users using the Kmenas model based on all user value scores and churn probabilities, outputs user grouping results and grouping portraits, and performs early warning and retention.

9. A computer device, comprising: It includes a processor and a memory; wherein the processor executes the computer program saved in the memory to realize the steps of the high-risk user churn early warning and retention method combining CNN and Informer of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program for storing; the computer program is executed by the processor to realize the steps of the high-risk user churn early warning and retention method combining CNN and Informer of any one of claims 1-7.

Images