Method, device, electronic equipment, computer readable storage medium and computer program product for churn prediction

Abstract

The application provides a churn prediction method and device, electronic equipment, computer readable storage medium and computer program product. The method comprises: acquiring first behavior data of a plurality of to-be-predicted objects in real time; for each to-be-predicted object, performing feature extraction on the first behavior data of the to-be-predicted object to obtain behavior features; calling a trained target prediction model to perform churn prediction on the behavior features to obtain a prediction result; determining a target object from the plurality of to-be-predicted objects based on the prediction result of each to-be-predicted object, and determining recommendation information for the target object based on the behavior features; and sending the recommendation information to a terminal corresponding to the target object. Through the application, the accuracy of churn prediction and the retention rate of the target object can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and particularly relates to a churn prediction method and device, electronic equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] With the development of big data and artificial intelligence technology, the research on user churn prediction and retention is of great strategic significance to enterprises. By identifying risk factors that may lead to churn, enterprises can more effectively allocate resources, improve key issues, and improve service quality and customer satisfaction. Data and analysis results of churn prediction can support long-term strategic planning of enterprises, helping enterprises better position the market and develop business plans. Through effective user retention strategies, enterprises can consolidate their market position and achieve sustainable development.

[0003] In the related art, the existing system cannot capture user behavior changes in real time, cannot quickly identify potential churn risks and take appropriate retention measures, and lacks personalized retention strategies, resulting in low accuracy of churn prediction and low retention rate of users. SUMMARY

[0004] The embodiments of the present application provide a churn prediction method, device, electronic equipment, computer readable storage medium and computer program product, which can improve the accuracy of churn prediction and the retention rate of target objects.

[0005] The technical scheme of the embodiments of the present application is as follows:

[0006] The embodiments of the present application provide a churn prediction method, which comprises:

[0007] Real-time acquisition of first behavior data of a plurality of to-be-predicted objects;

[0008] For each to-be-predicted object, feature extraction is performed on the first behavior data of the to-be-predicted object to obtain behavior features;

[0009] Calling a trained target prediction model, performing churn prediction on the behavior features to obtain a prediction result;

[0010] Based on the prediction result of each to-be-predicted object, a target object is determined from the plurality of to-be-predicted objects, and based on the behavior features, recommended information for the target object is determined;

[0011] The recommended information is sent to a terminal corresponding to the target object.

[0012] The embodiments of the present application provide a churn prediction device, which comprises:

[0013] A data acquisition module, used to acquire first behavior data of multiple objects to be predicted in real time;

[0014] A feature extraction module is used to extract features from the first behavior data of each object to be predicted to obtain a behavior feature;

[0015] A behavior prediction module is used to call the trained target prediction model to perform churn prediction on the behavior characteristics and obtain prediction results;

[0016] an information recommendation module, configured to determine a target object from a plurality of objects to be predicted based on a prediction result of each object to be predicted, and determine recommended information for the target object based on the behavior characteristics;

[0017] The information sending module is used to send the recommendation information to the terminal corresponding to the target object.

[0018] An embodiment of the present application provides an electronic device, comprising:

[0019] a memory for storing computer-executable instructions or computer programs;

[0020] The processor is used to implement the churn prediction method provided in the embodiment of the present application when executing the computer executable instructions or computer program stored in the memory.

[0021] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions or a computer program for implementing the churn prediction method provided in the embodiment of the present application when executed by a processor.

[0022] An embodiment of the present application provides a computer program product, including computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the churn prediction method provided in the embodiment of the present application is implemented.

[0023] The embodiments of the present application have the following beneficial effects:

[0024] By applying the embodiment of the present application, the first behavior data of multiple objects to be predicted are acquired in real time. Then, for each object to be predicted, feature extraction is performed on the first behavior data of the object to be predicted to obtain behavioral features. Then, the trained target prediction model is called to perform churn prediction on the behavioral features to obtain prediction results. The behavioral data of the object to be predicted can be acquired and analyzed in real time, so that the target prediction model can dynamically predict churn based on the behavioral changes of the object to be predicted, thereby improving the accuracy of churn prediction. Then, based on the prediction results of each object to be predicted, the target object is determined from the multiple objects to be predicted, and based on the behavioral features, the recommended information for the target object is determined, and then the recommended information is sent to the terminal corresponding to the target object. In this way, personalized recommendation information is sent to the target object based on the behavioral features of the target object, thereby improving the retention rate of the target object. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a schematic diagram of an application mode of the churn prediction method provided in an embodiment of the present application;

[0026] Figure 2 is a schematic structural diagram of an electronic device provided in an embodiment of the present application;

[0027] Figure 3A This is a first flow chart of the churn prediction method provided in an embodiment of the present application;

[0028] Figure 3B This is a second flow chart of the churn prediction method provided in an embodiment of the present application;

[0029] Figure 3C Schematic diagram of the training process of the target prediction model provided in the embodiment of the present application;

[0030] Figure 3D It is a structural diagram of the training process of the target prediction model provided in the embodiment of the present application.

[0031] It should be pointed out that the above-mentioned "first" and "second" are only used to distinguish different solutions, and do not represent the degree of distinction between the advantages and disadvantages of the solutions or the priority in the implementation process. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0033] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0034] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0035] The relevant data collection and processing in the embodiments of this application should be strictly in accordance with the requirements of relevant laws and regulations when applied in examples, and the informed consent or separate consent of the personal information subject should be obtained. Subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the personal information subject.

[0036] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or portion of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal. It can be implemented in whole or in part using software, hardware (such as processing circuits or memory), or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the functionality of the module or unit.

[0037] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meanings as those commonly understood by those skilled in the art. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0038] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0039] 1) Churn prediction: It is the process of analyzing user data to predict which users are likely to stop using a product or service.

[0040] 2) Retention strategy: These are retention measures taken for users who are predicted to churn, such as sending reminders, recommending activities, etc.

[0041] The embodiments of the present application provide a churn prediction method, apparatus, electronic device, computer-readable storage medium, and computer program product, which can improve the accuracy of churn prediction and the retention rate of target objects.

[0042] The following describes exemplary applications of the electronic devices provided in the embodiments of the present application. The electronic devices provided in the embodiments of the present application can be implemented as various types of terminals, such as laptops, tablet computers, desktop computers, set-top boxes, smartphones, smart speakers, smart watches, smart TVs, and in-vehicle terminals. They can also be implemented as servers. The following describes exemplary applications when the device is implemented as a server.

[0043] See also Figure 1 , Figure 1 This is a schematic diagram of the application mode of the churn prediction method provided in the embodiment of the present application, for example, Figure 1 The server 200, the network 300 and the terminal 400 are involved. The terminal 400 is connected to the server 200 via the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0044] When the churn prediction timing is reached, server 200 obtains the first behavior data of multiple objects to be predicted in real time. This can be determined when the current moment and the previous historical moment of churn prediction reach a preset interval, or when the number of active objects is determined to be less than a preset object threshold. For each object to be predicted, feature extraction is performed on the first behavior data of the object to be predicted to obtain behavioral features. A trained target prediction model is invoked to perform churn prediction on the behavioral features to obtain prediction results. Based on the prediction results for each object to be predicted, a target object is determined from the multiple objects to be predicted, and recommendation information for the target object is determined based on the behavioral features. The recommendation information is then sent to the terminal corresponding to the target object, which can be terminal 400. After receiving the recommendation information, terminal 400 displays the recommendation information on its own display device in response to a viewing operation on the recommendation information. Terminal 400 can also log in to a game application to play a game, or purchase virtual items, based on further feedback received regarding the recommendation information.

[0045] In some embodiments, the server (e.g., server 200) can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Terminal 400 can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, in-vehicle terminal, etc., but is not limited to these. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present application.

[0046] See also Figure 2 , Figure 2 is a structural diagram of an electronic device provided in an embodiment of the present application, which may be a terminal or a server. Figure 2 The electronic device 500 shown includes: at least one processor 410, a memory 450, and at least one network interface 420. The various components in the electronic device 500 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 440 is not described in detail. Figure 2 Various buses are labeled as bus system 440 .

[0047] The processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0048] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0049] The memory 450 includes volatile memory or nonvolatile memory, or may include both volatile and nonvolatile memory. The nonvolatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0050] In some embodiments, the memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0051] The operating system 451 includes system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., which are used to implement various basic businesses and process hardware-based tasks.

[0052] The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420 . Exemplary network interfaces 420 include Bluetooth, Wireless LAN (WiFi), and Universal Serial Bus (USB).

[0053] In some embodiments, the apparatus provided in the embodiments of the present application may be implemented in software. Figure 2 A churn prediction device 455 stored in the memory 450 is shown, which can be software in the form of programs and plug-ins, including the following software modules: a data acquisition module 4551, a feature extraction module 4552, a behavior prediction module 4553, an information recommendation module 4554, and an information sending module 4555. These modules are logical, and therefore can be arbitrarily combined or further split according to the functions implemented. The functions of each module will be explained below.

[0054] In some embodiments, the terminal or server can implement the churn prediction method provided in the embodiments of the present application by running various computer-executable instructions or computer programs. For example, the computer-executable instructions can be microprogram-level commands, machine instructions, or software instructions. The computer program can be a native program or software module in the operating system; it can be a native application (APPlication, APP), that is, a program that needs to be installed in the operating system to run, such as a cloud computing APP or an instant messaging APP; it can also be a small program that can be embedded in any APP, that is, a program that only needs to be downloaded to a browser environment to run. In short, the above-mentioned computer-executable instructions can be instructions in any form, and the above-mentioned computer program can be an application, module, or plug-in in any form.

[0055] The churn prediction method provided in the embodiment of the present application will be explained in combination with the exemplary application and implementation of the server device provided in the embodiment of the present application.

[0056] The following describes the churn prediction method provided in the embodiments of the present application. For ease of understanding, the following describes the application scenarios of the churn prediction method provided in the embodiments of the present application. In scenarios such as online shopping and live streaming, it is necessary to predict user churn rates and push recommended information to users. The churn prediction method provided in the embodiments of the present application can improve the accuracy of churn prediction and the retention rate of target users. In the embodiments of the present application, the application of the churn rate prediction scenario for game users is used as an example for explanation.

[0057] As mentioned above, the electronic device that implements the churn prediction method of the embodiment of the present application can be a terminal, a server, or a combination of the two. Next, the churn prediction method provided by the embodiment of the present application is described by taking the electronic device as an example of a server. Figure 3A , Figure 3A This is a first flow chart of the churn prediction method provided in the embodiment of the present application, which will be combined with Figure 3A The steps shown are explained.

[0058] In step 301, first behavior data of a plurality of objects to be predicted are acquired in real time.

[0059] Here, the subject to be predicted can be a game user, and the first behavior data is the behavior data of the subject to be predicted collected at the current moment. The first behavior data includes activity data, consumption data, and social data. Activity data includes login frequency data, average game duration data, and average daily activity data. Consumption data includes total consumption amount data, consumption frequency data, and consumption category data. Social data includes social interaction data and friend interaction data. The first behavior data of multiple subjects to be predicted, collected in real time, is stored in a big data platform (such as Kafka, Flume) or a database (such as MongoDB, HBase).

[0060] In step 302 , for each object to be predicted, feature extraction is performed on the first behavior data of the object to be predicted to obtain behavior features.

[0061] Here, the first behavior data of each target to be predicted is preprocessed by removing duplicates, handling missing values, time series conversion, standardization, and normalization. Feature extraction is then performed on the preprocessed first behavior data to obtain behavioral features. Behavioral features include activity features, consumption features, and social features.

[0062] In some embodiments, see Figure 3B , Figure 3B This is a second flow chart of the churn prediction method provided in an embodiment of the present application. Figure 3A In step 302 shown in FIG. 1 , “feature extraction is performed on the first behavior data of the prediction object to obtain behavior features”, this can be achieved by Figure 3B Steps 3021 to 3024 are implemented as described below.

[0063] In step 3021 , feature extraction is performed on the activity data of the object to be predicted to obtain activity features.

[0064] Here, the activity data can be login frequency data, average game duration data, and daily average activity data. Feature extraction is performed on the login frequency data of the object to be predicted to obtain a login frequency feature. The login frequency feature is used to characterize the number of game logins of the object to be predicted within a certain period of time. Feature extraction is performed on the average game duration data of the object to be predicted to obtain an average game duration feature. The average game duration feature is used to characterize the average game duration of each game login of the object to be predicted. Feature extraction is performed on the daily average activity data of the object to be predicted to obtain a daily average activity feature. The daily average activity feature is used to characterize the number of days the object to be predicted was active within a certain period of time in the past. The login frequency feature, average game duration feature, and daily average activity feature are combined into an activity feature.

[0065] In step 3022, feature extraction is performed on the consumption data of the object to be predicted to obtain consumption features.

[0066] Here, consumption data can include total consumption amount data, consumption frequency data, and consumption category data. Feature extraction is performed on the total consumption amount data of the subject to be predicted to obtain a total consumption amount feature. The total consumption amount feature is used to represent the total consumption amount of the subject to be predicted within a certain time period. Feature extraction is performed on the consumption frequency data of the subject to be predicted to obtain a consumption frequency feature. The consumption frequency feature is used to represent the number of times the subject to be predicted spends money monthly, weekly, or daily. Feature extraction is performed on the consumption category data of the subject to be predicted to obtain a consumption category feature. The consumption category feature is used to represent the subject's consumption preferences (such as in-game props, virtual items, etc.). The total consumption amount feature, consumption frequency feature, and consumption category feature are combined to form a consumption feature.

[0067] In step 3023 , feature extraction is performed on the social data of the prediction object to obtain social features.

[0068] Here, social data can include social interaction data and friend interaction data. Feature extraction is performed on the social interaction data of the target being predicted to obtain social interaction features. These features characterize the frequency with which the target being predicted participates in social activities (such as chatting and teaming). Feature extraction is performed on the friend interaction data of the target being predicted to obtain friend interaction features. These features characterize the frequency and duration of interactions between the target being predicted and its friends. The social interaction features and friend interaction features are combined to form social features.

[0069] In step 3024, the activity characteristics, consumption characteristics, and social characteristics are integrated to obtain behavioral characteristics.

[0070] Here, a neural network model can be used to perform multimodal learning on activity, consumption, and social features. The fully connected layer of the neural network model is used to fuse these features to obtain a multimodal fused feature. The normalization layer of the neural network model is then used to perform state prediction on these fused features to obtain the behavioral characteristics of the object to be predicted.

[0071] In an embodiment of the present application, feature extraction is performed on the first behavior data of the object to be predicted to obtain activity features, consumption features, and social features, and the activity features, consumption features, and social features are integrated to obtain behavior features. This ensures that the behavior features of the object to be predicted include multi-dimensional data features, which can improve the accuracy of churn prediction based on the behavior features of the object to be predicted.

[0072] Continue to refer Figure 3A ,In step 303, the trained target prediction model is called to perform churn prediction on the behavioral features to obtain the prediction results.

[0073] Here, churn refers to the absence of logins or active behavior by the target within a certain period of time. Churn prediction refers to predicting the probability of churn for the target. The target prediction model can be a deep reinforcement learning model, a supervised learning model, or a sequence data model. Deep reinforcement learning models can be models such as the Deep Q-Network (DQN) or the Deep Deterministic Policy Gradient (DDPG). Supervised learning models can be models such as the Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Random Forest. Sequence data models can be models such as the Autoregressive Integrated Moving Average (ARIMA) model and the Long Short-Term Memory (LSTM) model.

[0074] In some embodiments, see Figure 3C , Figure 3C This is a diagram of the training process of the target prediction model provided by the embodiment of the present application. Before step 303, you can also Figure 3C The steps shown are used to train the target prediction model to be trained, and the trained target prediction model is obtained. Figure 3C Provide explanation.

[0075] In step 310 , a sample behavior dataset and a target prediction model to be trained are obtained.

[0076] Here, the sample behavior dataset includes sample behavior data and churn annotation information corresponding to the sample behavior data. The churn annotation information represents the churn rate corresponding to the sample behavior data. The sample behavior data is the collected historical behavior data of the target to be predicted. Each sample behavior data is annotated with a churn rate label to obtain the churn annotation information corresponding to the sample behavior data.

[0077] In step 311, the feature extraction layer in the target prediction model to be trained is called to extract features from the sample behavior data to obtain sample behavior features.

[0078] Here, the sample behavior data undergoes preprocessing, including deduplication, missing value processing, time series processing, standardization, and normalization. Deduplication removes duplicate data records from the sample behavior data. Missing value processing fills in missing fields in the sample behavior data or deletes data records with a high number of missing values. Time series processing sorts the sample behavior data by timestamp. Standardization and normalization standardize and normalize fields such as consumption amount and game duration in the sample behavior data. The feature extraction layer in the target prediction model to be trained is called to extract features from the preprocessed sample behavior data to obtain sample behavior features.

[0079] In step 312, the normalization layer in the target prediction model to be trained is called to perform churn rate prediction on the sample behavior features to obtain the predicted churn rate of the sample behavior data.

[0080] Here, the normalization layer in the target prediction model to be trained is called to normalize the sample behavior features, and the normalization result is used as the predicted churn rate of the sample behavior data.

[0081] In step 313 , the target prediction model to be trained is trained based on the churn annotation information and the predicted churn rate to obtain a trained target prediction model.

[0082] Here, the actual churn rate corresponding to the sample behavior data is determined based on the churn annotation information. A loss function such as a binary cross entropy loss function or a logistic regression loss function can be used. A loss value for the predicted churn rate is determined based on the actual churn rate and the predicted churn rate corresponding to the sample behavior data. The model parameters of the target prediction model to be trained are updated based on the loss value for the predicted churn rate, thereby training the target prediction model and obtaining a trained target prediction model.

[0083] Example, reference Figure 3D , Figure 3DIt is a structural diagram of the training process of the target prediction model provided in an embodiment of the present application. Call the feature extraction layer 603 in the target prediction model to be trained 601, perform feature extraction on the sample behavior data 602, and obtain the sample behavior features. Call the normalization layer 604 in the target prediction model to be trained 601, perform churn rate prediction on the sample behavior features, and obtain the predicted churn rate of the sample behavior data. Based on the churn annotation information 605 and the predicted churn rate, determine the loss value of the predicted churn rate. Update the model parameters of the target prediction model to be trained according to the loss value of the predicted churn rate to train the target prediction model and obtain the trained target prediction model.

[0084] In an embodiment of the present application, the target prediction model to be trained is trained based on the churn annotation information and the predicted churn rate to obtain a trained target prediction model. The trained target prediction model updates the model parameters, thereby improving the accuracy of the trained target prediction model in churn prediction.

[0085] Continue to refer Figure 3A In step 304, based on the prediction results of each object to be predicted, a target object is determined from multiple objects to be predicted, and based on the behavioral characteristics, recommendation information for the target object is determined.

[0086] Here, the target object is an object to be predicted whose churn probability is greater than a preset probability threshold, and the recommended information is information for retaining the target object.

[0087] In some embodiments, the prediction results include churn probability. The above-mentioned step 304 determines the target object from multiple objects to be predicted based on the prediction results of each object to be predicted, and determines the recommendation information for the target object based on the behavioral characteristics. This can be achieved by executing the following steps: for the prediction results of each object to be predicted, when the churn probability is greater than a preset probability threshold, the object to be predicted is determined as the target object; based on the behavioral characteristics of the target object, the target object is classified and processed to obtain the target type of the target object; at least one candidate recommendation information corresponding to the target type and the evaluation index value of each candidate recommendation information are obtained; based on the evaluation index value of each candidate recommendation information, the recommendation information of the target object is determined from at least one candidate recommendation information.

[0088] Here, the churn probability of the target subject is determined based on the prediction results. From multiple targets, a target subject with a churn probability greater than a preset probability threshold is identified and designated as the target subject. Specifically, the target subject with a high churn risk is designated as the target subject. The target subject's behavioral characteristics are input into a classification model, such as a random forest model. The target subject's target type is determined based on the model's output. The target type characterizes the target subject's type, which can include active or new user types.

[0089] The target type can correspond to multiple candidate recommendation information, and at least one candidate recommendation information and an evaluation index value of the candidate recommendation information are obtained from the multiple recommendation information. The evaluation index value is used to evaluate the retention success rate of the recommendation information for the target object. Based on the evaluation index value of each candidate recommendation information, the candidate recommendation information corresponding to the highest evaluation index value is determined, and the candidate recommendation information is determined as the recommendation information for the target object. The evaluation index value of each candidate recommendation information can also be added to obtain the added result, and the evaluation index value of each candidate recommendation information is divided by the added result, and the obtained division result is determined as the selection probability of each candidate recommendation information. Based on the selection probability of each candidate recommendation information, a candidate recommendation information is randomly selected from the multiple candidate recommendation information, and the candidate recommendation information is used as the recommendation information for the target object.

[0090] For example, the target type corresponds to five candidate recommendation information, and the evaluation index values ​​of each candidate recommendation information are 0.8, 0.7, 0.4, 0.2, and 0.4, respectively. The evaluation index values ​​of each candidate recommendation information are added together, resulting in a total of 2.5. The evaluation index value of each candidate recommendation information is divided by the sum, resulting in selection probabilities of 0.8 / 2.5, 0.7 / 2.5, 0.4 / 2.5, 0.2 / 2.5, and 0.4 / 2.5, respectively. Based on the selection probabilities of the five candidate recommendation information, a candidate recommendation information is randomly selected from the multiple candidate recommendation information and is used as the recommendation information for the target object.

[0091] In an embodiment of the present application, the target object is classified based on the behavioral characteristics of the target object to obtain the target type of the target object, and based on at least one candidate recommendation information corresponding to the target type of the target object and the evaluation index value of each candidate recommendation information, the recommendation information of the target object is determined from the at least one candidate recommendation information. Based on the behavioral characteristics of the target object and the evaluation index value of the candidate recommendation information, the personalized recommendation information of the target object can be accurately screened out, thereby improving the retention rate of the target object.

[0092] Continue to refer Figure 3AIn step 305, the recommendation information is sent to the terminal corresponding to the target object.

[0093] Here, recommendation information is sent to the predicted objects with high churn risk.

[0094] In some embodiments, step 305 can be implemented by executing the following steps: obtaining preference information of the target object, and determining a target push mode of the recommendation information based on the preference information; and sending the recommendation information to the terminal corresponding to the target object using the target push mode.

[0095] Here, the target subject's preference information may be channel preference information, which is the target subject's preference for a channel for obtaining recommended information. For example, the channel preference information may include SMS channel preference, push notification channel preference, game terminal reminder channel preference, etc. Based on the target subject's channel preference information, a target push method for the recommended information is determined. The target push method is the method for pushing the target subject's recommended information. Based on the determined target push method, the recommended information is sent to the target subject's corresponding terminal.

[0096] In an embodiment of the present application, a target push method for recommendation information is determined based on the preference information of the target object, and the target push method is used to send the recommendation information to the terminal corresponding to the target object, which can improve the click rate of the target object on the recommendation information.

[0097] In some embodiments, after sending the recommendation information to the terminal corresponding to the target object, the following steps can also be performed: obtaining the second behavior data of the target object within a preset time period; determining the retention result of the target object based on the first behavior data and the second behavior data; and updating the evaluation index value of the recommendation information based on the retention result.

[0098] Here, the preset duration is the time interval between the current moment and the moment the first behavioral data was collected, which is three or five days. The second behavioral data is the behavioral data collected from the target object within the preset duration. Based on the first and second behavioral data of the target object, a retention result for the target object is determined. The retention result indicates whether the target object has been successfully retained. Whether the target object is active is determined based on the activity data in the first and second behavioral data. Specifically, a weighted sum of the login frequency data, average game duration data, and average daily activity data in the activity data is taken to obtain an activity score. When the activity score is greater than a preset activity score threshold, the target object is determined to be active. When the activity score is less than or equal to the preset activity score threshold, the target object is determined to be inactive. If the target object is determined to be active based on the first behavioral data and is also determined to be active based on the second behavioral data, the retention result indicates that the target object has been successfully retained. If the target object is determined to be active based on the first behavioral data and is also determined to be inactive based on the second behavioral data, the retention result indicates that the target object has failed to be retained. An actual retention success rate of the recommended information for the target object is determined based on the retention result, and the actual retention success rate is determined as an evaluation index value of the recommended information to update the evaluation index value of the recommended information.

[0099] In an embodiment of the present application, the retention result of the target object is determined based on the first behavior data and the second behavior data, and the evaluation index value of the recommendation information is updated based on the retention result. The evaluation index value of the recommendation information can be updated in a timely manner within a preset time period, thereby improving the accuracy of determining the recommendation information of the target object.

[0100] In some embodiments, the first behavior data includes at least social data, and the behavior features can be obtained by performing the following steps: constructing a social network graph based on the first behavior data of multiple objects to be predicted; performing graph feature extraction on the social network graph to obtain a graph embedding vector and a node embedding vector corresponding to each object to be predicted; and determining the graph embedding vector and the node embedding vector as the behavior features of the object to be predicted.

[0101] Here, social data includes social interaction data. Multiple objects to be predicted are identified as nodes of a social network graph. Based on the social data in the first behavior data of the multiple objects to be predicted, social interaction data of the multiple objects to be predicted is determined. The social interaction data is used to determine social relationships between the multiple objects to be predicted. The social relationships are then identified as edges of the social network graph. Edges are connections between nodes in the social network graph. Based on the determined nodes and edges of the social network graph, a graph database or graph processing library is used to construct the social network graph.

[0102] Graph embedding methods (such as Node2Vec and Graph Neural Networks) are used to extract graph features from the social network graph, generating a graph embedding vector and a node embedding vector corresponding to each object to be predicted. The graph embedding vector is a low-dimensional vector corresponding to the social network graph, while the node embedding vector is a low-dimensional vector corresponding to each node in the social network graph. For each object to be predicted, the graph embedding vector and the node embedding vector corresponding to the object to be predicted are concatenated to obtain the behavioral characteristics of the object to be predicted. Different weights can also be assigned to the graph embedding vector and the node embedding vector corresponding to the object to be predicted, and the weighted sum of the graph embedding vector and the node embedding vector corresponding to the object to be predicted is performed to obtain the behavioral characteristics of the object to be predicted.

[0103] In an embodiment of the present application, graph features are extracted from a social network graph to obtain a graph embedding vector and a node embedding vector corresponding to each object to be predicted. The graph embedding vector and the node embedding vector are determined as behavioral features of the object to be predicted, thereby achieving determination of the behavioral features of the object to be predicted based on the social network graph, and performing churn prediction based on the behavioral features including social network relationships. This takes into account the impact of social network relationships on churn prediction, thereby improving the accuracy of churn prediction.

[0104] In some embodiments, the "determining recommended information for the target object based on behavioral characteristics" in the above step 304 can be achieved by performing the following steps: determining the target social group to which the target object belongs and the importance value of the target object in the target social group based on behavioral characteristics; determining the group activity of the target social group; and determining the recommended information for the target object based on the group activity and importance value.

[0105] Here, based on the target subject's behavioral characteristics within the social network graph, a graph algorithm (such as the Louvain algorithm) is used to analyze the target subject's social groups. This determines the target subject's social group, which is the target social group. The target subject's social group is also known as the target social group. The target subject's importance within the target social group is also determined. A degree centrality or betweenness centrality algorithm is then used to analyze the target subject's centrality and determine its importance within the target social group. The importance value measures the target subject's importance within the target social group.

[0106] Obtain activity data for each social object in the target social group. The activity data may include login frequency data, average game duration data, and average daily activity data. Calculate the activity score corresponding to each social object's activity data, calculate an average based on each activity score, and determine the average as the group activity of the target social group. Determine the activity level of the target social group based on the group activity. When the group activity is higher than a preset activity threshold, the activity level of the target social group is high. When the group activity is less than or equal to the preset activity threshold, the activity level of the target social group is low.

[0107] When the importance value of the target object in the target social group is higher than the preset threshold, and the activity level of the target social group is high, the target object is determined to be a high-value object type, and the recommendation information corresponding to the high-value object type is determined as the recommended information for the target object. The recommended information corresponding to the high-value object type may be exclusive activity recommendation information, high reward information, etc. When the importance value of the target object in the target social group is less than or equal to the preset threshold, and the activity level of the target social group is low, the target object is determined to be a low-value object type, and the recommendation information corresponding to the low-value object type is determined as the recommended information for the target object. The recommended information corresponding to the low-value object type may be activity reminder information, low reward information, etc.

[0108] In an embodiment of the present application, the recommended information of the target object is determined based on the group activity of the target social group to which the target object belongs and the importance value of the target object in the target social group. The recommended information of the target object can be accurately determined based on the social network relationship of the target object, thereby improving the retention rate of the target object.

[0109] In some embodiments, the churn prediction method provided by the embodiments of the present application can be applied in the field of cloud technology. The first behavior data of multiple objects to be predicted are obtained in real time on the cloud platform, and then for each object to be predicted, the first behavior data of the object to be predicted is subjected to feature extraction to obtain behavioral features. Secondly, the trained target prediction model is called to perform churn prediction on the behavioral features to obtain prediction results. The behavioral data of the object to be predicted can be obtained and analyzed in real time, so that the target prediction model can dynamically predict churn based on the behavioral changes of the object to be predicted, thereby improving the accuracy of churn prediction. Based on the prediction results of each object to be predicted, the target object is determined from the multiple objects to be predicted, and based on the behavioral features, the recommended information for the target object is determined, and then the recommended information is sent to the terminal corresponding to the target object. In this way, the cloud server sends personalized recommendation information to the target object based on the behavioral features of the target object, thereby improving the retention rate of the target object.

[0110] The following describes an exemplary application of the churn prediction method provided in an embodiment of the present application in scenarios of churn rate prediction and retention information recommendation for game users.

[0111] In related technologies, traditional statistical methods and simple machine learning models have low churn prediction accuracy in complex environments, making it difficult to accurately identify potential churned game users. Existing systems do not recommend personalized retention information based on the specific behavior patterns and characteristics of game users, and do not capture changes in game user behavior in real time, resulting in low accuracy in churn prediction and low retention rate of game users.

[0112] In this embodiment, a method for predicting customer churn is proposed to address the problems in related technologies. Compared with related technologies, the method includes the following improvements:

[0113] 1) Using deep reinforcement learning models to analyze game user behavior data in real time and establish a dynamic churn prediction model to improve churn prediction accuracy;

[0114] 2) By analyzing the behavioral characteristics of game users, we recommend personalized retention information to them, detect behavioral changes of game users in real time, quickly identify potential churners, and recommend corresponding retention information.

[0115] The churn prediction method proposed in the embodiment of the present application can be used to design the product interface. The product interface includes a game user dashboard and a retention information recommendation module. The game user dashboard is used to display key game user behavior indicators, such as activity, game time, churn probability and prediction results. The retention information recommendation module is used to provide personalized retention information, such as specific push notification content, social activities, etc. The churn prediction method proposed in the embodiment of the present application can provide data support for game development and operation, and help game operation management to detect game user behavior in real time and optimize retention information. Through personalized retention information, the game user's gaming experience and satisfaction can be improved.

[0116] The system architecture in this application embodiment includes a data collection module, a big data analysis module, a deep reinforcement learning module, and a decision recommendation module. The specific algorithm implementation and data flow links of each module help form a closed loop, which helps to improve the retention rate of game users and the life cycle value of the game.

[0117] The data collection module is used to collect game user behavior data in real time and store it in an appropriate storage system. Game user behavior data includes login frequency, game duration, spending history, and social interaction. Login frequency records the timestamp and interval between each game user login. Game duration records the start and end time of each game session, calculating the duration of each session. Spending history records the items, amounts, and timestamps of each purchase by game users. Social interaction records social activities of game users, such as chatting, team formation, and gift giving. Game user behavior data is collected in real time through backend services and written to big data platforms (such as Kafka and Flume) or databases (such as MongoDB and HBase).

[0118] The big data analysis module is used to analyze large amounts of game user behavior data, extract behavioral characteristics of game users, and support subsequent churn prediction and retention decisions. Game user behavior data undergoes preprocessing such as deduplication, missing value processing, time series processing, standardization, and normalization. Deduplication removes duplicate data records from game user behavior data. Missing value processing fills in missing fields in game user behavior data or deletes data records with a high number of missing values. Time series processing sorts each game user's behavior data by timestamp to facilitate subsequent sequence analysis. Standardization and normalization standardize and normalize fields such as spending amount and game duration in game user behavior data to eliminate dimensional differences.

[0119] Through statistical analysis of game user behavior data, behavioral characteristics of game users are extracted. These behavioral characteristics include activity, consumption, social interaction, and behavioral trend characteristics. Activity characteristics include login frequency, average game duration, and average daily active days. Login frequency refers to the number of times a game user logs in within a certain period. Average game duration refers to the average duration of each login. Average daily active days refers to the number of days a game user was active within a certain period. Consumption characteristics include total spending amount, consumption frequency, and consumption categories. Total spending amount refers to the total amount a game user spends within a certain period. Consumption frequency refers to the number of times a game user spends money per month, week, or day. Consumption categories refer to the game user's consumption preferences (such as in-game props and virtual items). Social characteristics include social interaction frequency and friend interaction. Social interaction frequency refers to the frequency with which a game user participates in social activities (such as chatting and teaming). Friend interaction refers to the frequency and duration of interaction between a game user and their friends. Behavioral trend characteristics include behavioral change trends and churn risk. Behavioral change trends refer to whether a game user's activity, consumption, and social behavior are showing a downward or upward trend. Churn risk is based on historical behavioral data and uses statistical methods to predict the risk of game users churn.

[0120] The big data analysis module also performs cluster analysis and association rule analysis on the behavioral characteristics of game users. Cluster analysis uses clustering algorithms (such as K-means) to categorize game users into different groups. For example, game users can be divided into high-value, average, low-value, and potential churn groups. Association rule analysis uses association rule algorithms (such as Apriori) to discover behavioral correlations among game users and identify potential relationships between their behaviors. For example, game users who frequently purchase virtual items are more likely to participate in social activities.

[0121] The deep reinforcement learning module is used to build a system that provides real-time feedback and continuously optimizes churn prediction and retention information. Retention strategies are continuously adjusted through reinforcement learning to improve user retention. The churn prediction model can be a deep reinforcement learning model, such as a Deep Q-Network (DQN) or Deep Deterministic Policy Gradient (DDPG). A deep reinforcement learning model consists of a state space, an action space, and a reward function. In the state space, each user's state is represented as a vector based on behavioral characteristics (such as login frequency, game duration, and spending behavior). The action space of the deep reinforcement learning model represents the retention strategy, such as pushing event information or personalized advertising. The reward function can be based on the user's retention status or behavioral changes after receiving retention information. Positive rewards are given if the user logs back in or increases spending after receiving retention information; negative rewards are given if the user churns.

[0122] The Q-network of a deep reinforcement learning model is trained using historical user behavior data to estimate the expected rewards for various retention strategies. During training, the model is optimized using an ε-greedy strategy that balances exploration and exploitation. As the model receives behavioral feedback from game users, the Q-network of the deep reinforcement learning model is continuously updated to accurately predict retention information that will effectively retain game users. Every change in user behavior is fed into the deep reinforcement learning model as feedback, further optimizing churn prediction and retention strategies.

[0123] The deep reinforcement learning module is also used for churn prediction and retention strategy adjustments. Churn prediction predicts the probability of user churn based on user behavioral characteristics. For example, if a user's login frequency and game time decrease significantly, they can be identified as a user at risk of churn. Retention strategies are tailored to different users at risk of churn. The system outputs personalized retention information, such as offering more exclusive activities to high-value users and improving retention rates for low-value users through simple reminders and rewards.

[0124] The decision-making recommendation module is used to provide personalized retention information to game users based on the output of the deep reinforcement learning module. Based on the game user's churn risk and value, the system categorizes game users into different groups, such as high-value game users and high-churn risk game users. Based on each game user's type and the model's output, personalized retention information is recommended to game users. For example, for high-value game users, more exclusive event information or in-depth social interaction information can be recommended. For low-value game users, notifications and reminders about upcoming events can be provided. Based on the game user's preferences, the channel for pushing retention information is determined, such as SMS push, push notifications, in-game reminder push, etc.

[0125] The decision-making recommendation module also provides decision support for game operations based on the output of the deep reinforcement learning module. The system can also provide data analysis reports for game operations, demonstrating the effectiveness of different retention strategies. For example, it can track key metrics such as the retention success rate of specific retention information and the user retention rate, helping game operations adjust retention strategies in a timely manner. A / B testing can be used to compare different retention strategies and evaluate which ones are more effective for different game user groups.

[0126] In an embodiment of the present application, the deep reinforcement learning module can use supervised learning methods to predict churn. The churn prediction model can be a supervised learning model, which can be a support vector machine (SVM) model, a multi-layer perceptron (MLP) model, a random forest, or a decision tree model. The support vector machine (SVM) model separates churned game users from non-churned game users by constructing a hyperplane and is suitable for small data sets and high-dimensional data. The multi-layer perceptron (MLP) model can learn complex nonlinear relationships and is suitable for scenarios with large data volumes and complex behavioral characteristics of game users. Random forests or decision trees are tree-based models that process and interpret large data sets.

[0127] Supervised learning methods use labeled historical behavioral data of game users to train a churn prediction model and formulate retention strategies based on the model's predictions. The user behavior data undergoes preprocessing, including numerical standardization, handling missing values, and encoding categorical features. Feature extraction is then performed on the preprocessed user behavior data to obtain behavioral characteristics, such as login frequency, game duration, spending amount, and social behavior. The user's behavior data is labeled with a churn label based on whether the user has churned. Churn can be defined as a user not logging in or being active for a certain period of time. The churn prediction model is trained using labeled historical behavioral data and its performance is evaluated through cross-validation. Evaluation metrics include precision, recall, and the harmonic mean of precision and recall, with recall capturing the majority of churned game users. The churn prediction model outputs a churn probability for each game user. If a user's churn probability exceeds a certain threshold, the user is labeled as a high-churn risk game user. Based on the prediction results for each game user, personalized retention information is recommended for each user.

[0128] In embodiments of the present application, a churn prediction model may be a sequence data model. This model is suitable for processing temporal trends in game user behavior and predicting churn by analyzing user behavior sequences. Sequence data models may include models such as the Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM). The ARIMA is suitable for processing linear time series and predicts churn by performing differential processing on historical user behavior data. The LSTM is suitable for capturing long-term dependencies and can process user behavior data with strong temporal dependencies, predicting whether a user will churn in the future. The user's historical behavior data is divided into time windows, which can be one week or one month long, to generate time series data. This time series data is used as the basis for training a churn prediction model. Feature extraction is performed on the time series data to obtain trend characteristics of the time series data, such as cyclicality, growth or decline, and seasonality. Based on the trend characteristics of the time series data, a churn prediction model is trained to obtain a trained churn prediction model. The trained churn prediction model is used to predict the changing trends of game user behavior. If a user's activity level decreases, it is predicted that the user is likely to churn. Adjust retention strategies based on the predicted trends in game user behavior. If you predict that game users may churn, you can send reminders and recommend activities in advance.

[0129] In embodiments of the present application, social network analysis can be performed on game users to obtain their social relationships. Based on these relationships, social factors influencing their behavior can be identified, thus providing a basis for churn prediction. Game users are represented as nodes in a social network graph, and social relationships (such as friendships, teaming, and other interactions) are represented as edges. A social network graph is constructed based on the nodes and edges of the social network graph. Social behavior characteristics of each game user, such as the number of friends, interaction frequency, and interaction duration, are extracted. Social group analysis of game users is performed using graph algorithms (such as the Louvain algorithm) to determine the social group to which the game user belongs. The risk of churn may be closely related to the overall behavior of the social group. Centrality analysis of game users is performed using degree centrality or betweenness centrality algorithms to measure their importance within the social network. The churn of core game users within the social network may trigger the churn of even more game users. Graph embedding methods (such as Node2Vec and GraphNeural Networks) are used to convert the social network graph into a low-dimensional vector to obtain embedding vector features of the social network graph. These embedding vector features are then input into a churn prediction model. Churn prediction models can identify core game users with high social influence and analyze the impact of their behavior on other game users within their social groups. If core game users churn, this can lead to further churn within the same social group. Retention strategies can be developed for different social groups. For example, providing more social interactions or rewards to game users in highly active social groups can improve user retention. For game users in social groups with high churn risk, targeted retention campaigns can be designed, such as group team-building rewards and group social activity rewards.

[0130] In the aforementioned application scenarios of predicting game user churn and recommending retention information, real-time monitoring and analysis of user behavior provide decision support based on big data analysis results, promoting scientific and data-driven game operations. This enables game operations to quickly respond to user needs and enhance the user experience. Targeted retention strategies and personalized retention information recommendations can effectively improve user retention rates.

[0131] The following continues to describe the exemplary structure of the churn prediction device 455 provided in the embodiment of the present application as a software module. In some embodiments, such as Figure 2As shown, the software modules stored in the churn prediction apparatus 455 of the memory 450 can include: a data acquisition module 4551 configured to acquire first behavior data of a plurality of to-be-predicted objects in real time; a feature extraction module 4552 configured to, for each to-be-predicted object, perform feature extraction on the first behavior data of the to-be-predicted object to obtain behavior features; a behavior prediction module 4553 configured to call a trained target prediction model to perform churn prediction on the behavior features to obtain a prediction result; an information recommendation module 4554 configured to determine a target object from the plurality of to-be-predicted objects based on the prediction result of each to-be-predicted object, and determine recommendation information for the target object based on the behavior features; and an information sending module 4555 configured to send the recommendation information to a terminal corresponding to the target object.

[0132] In some embodiments, the behavior prediction module 4553 is further configured to acquire a sample behavior data set and a to-be-trained target prediction model, the sample behavior data set including sample behavior data and churn label information corresponding to the sample behavior data; call a feature extraction layer in the to-be-trained target prediction model to perform feature extraction on the sample behavior data to obtain sample behavior features; call a normalization layer in the to-be-trained target prediction model to perform churn rate prediction on the sample behavior features to obtain a predicted churn rate of the sample behavior data; and train the to-be-trained target prediction model based on the churn label information and the predicted churn rate to obtain the trained target prediction model.

[0133] In some embodiments, the first behavior data includes activity data, consumption data, and social data, and the feature extraction module 4552 is further configured to perform feature extraction on the activity data of the to-be-predicted object to obtain activity features, perform feature extraction on the consumption data of the to-be-predicted object to obtain consumption features, and perform feature extraction on the social data of the to-be-predicted object to obtain social features; and fuse the activity features, the consumption features, and the social features to obtain the behavior features.

[0134] In some embodiments, the prediction result includes a churn probability, and the information recommendation module 4554 is further configured to, for the prediction result of each to-be-predicted object, determine the to-be-predicted object as the target object when the churn probability is greater than a preset probability threshold; perform classification processing on the target object based on the behavior features of the target object to obtain a target type of the target object; acquire at least one candidate recommendation information corresponding to the target type and an evaluation index value of each candidate recommendation information; and determine the recommendation information of the target object from the at least one candidate recommendation information based on the evaluation index value of each candidate recommendation information.

[0135] In some embodiments, the information sending module 4555 is further configured to acquire preference information of the target object, and determine a target pushing manner of the recommendation information based on the preference information; and send the recommendation information to the terminal corresponding to the target object by using the target pushing manner.

[0136] In some embodiments, the information sending module 4555 is also used to obtain the second behavior data of the target object within a preset time period after sending the recommendation information to the terminal corresponding to the target object; determine the retention result of the target object based on the first behavior data and the second behavior data; and update the evaluation index value of the recommendation information based on the retention result.

[0137] In some embodiments, the first behavior data includes at least social data, and the feature extraction module 4552 is further used to construct a social network graph based on the first behavior data of multiple objects to be predicted; perform graph feature extraction on the social network graph to obtain a graph embedding vector and a node embedding vector corresponding to each object to be predicted; and determine the graph embedding vector and the node embedding vector as the behavior features of the object to be predicted.

[0138] In some embodiments, the information recommendation module 4554 is also used to determine the target social group to which the target object belongs and the importance value of the target object in the target social group based on behavioral characteristics; determine the group activity of the target social group; and determine recommended information for the target object based on the group activity and importance value.

[0139] The present invention provides a computer program product comprising computer-executable instructions or a computer program stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions or the computer program from the computer-readable storage medium and executes the computer-executable instructions or the computer program, causing the electronic device to perform the churn prediction method described in the present invention.

[0140] The embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions or computer programs are stored. When the computer-executable instructions or computer programs are executed by a processor, the processor will execute the churn prediction method provided by the embodiment of the present application, for example, Figure 3A The churn prediction method is shown.

[0141] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or various devices including one or any combination of the above memories.

[0142] In some embodiments, computer-executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0143] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0144] By way of example, computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed across multiple sites and interconnected by a communication network.

[0145] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. A method for predicting churn, characterized in that: The method comprises: Acquiring first behavior data of a plurality of objects to be predicted in real time, wherein the first behavior data at least includes social data; constructing a social network graph based on the first behavior data of the plurality of objects to be predicted, wherein the nodes of the social network graph are the objects to be predicted, and the edges are the social relationships between the objects to be predicted; Performing graph feature extraction on the social network graph to obtain a graph embedding vector and a node embedding vector corresponding to each of the objects to be predicted, wherein the graph feature extraction adopts a graph embedding method; Determining the graph embedding vector and the node embedding vector as behavioral features of the object to be predicted; Calling the trained target prediction model to perform churn prediction on the behavioral features to obtain a prediction result; Determining a target object from the plurality of objects to be predicted based on the prediction result of each of the objects to be predicted; Based on the behavioral characteristics, determining a target social group to which the target object belongs and an importance value of the target object in the target social group; Determining the group activity of the target social group, where the group activity is an average value calculated based on activity scores corresponding to activity data of each social object in the target social group, where the activity data includes login frequency data, average game duration data, and average daily activity data; Determining recommendation information for the target object based on the group activity and the importance value; The recommendation information is sent to a terminal corresponding to the target object.

2. The method according to claim 1, characterized in that The method further comprises: Obtaining a sample behavior data set and a target prediction model to be trained, wherein the sample behavior data set includes sample behavior data and churn annotation information corresponding to the sample behavior data; Calling a feature extraction layer in the target prediction model to be trained to perform feature extraction on the sample behavior data to obtain sample behavior features; Calling the normalization layer in the target prediction model to be trained to predict the churn rate of the sample behavior features to obtain the predicted churn rate of the sample behavior data; Based on the churn annotation information and the predicted churn rate, the target prediction model to be trained is trained to obtain a trained target prediction model.

3. The method according to claim 1, characterized in that The first behavior data includes activity data, consumption data, and social data. The feature extraction of the first behavior data of the object to be predicted to obtain behavior features includes: Extracting features from the activity data of the object to be predicted to obtain activity features; Extracting features from the consumption data of the object to be predicted to obtain consumption features; Extracting features from the social data of the object to be predicted to obtain social features; The activity characteristics, consumption characteristics and social characteristics are integrated to obtain behavioral characteristics.

4. The method according to any one of claims 1 to 3, characterized in that The prediction result includes the churn probability, and the determining of the recommended information of the target object includes: For each prediction result of the object to be predicted, when the churn probability is greater than a preset probability threshold, the object to be predicted is determined as a target object; Based on the behavioral characteristics of the target object, the target object is classified to obtain the target type of the target object; Obtain at least one candidate recommendation information corresponding to the target type and an evaluation index value of each candidate recommendation information; Based on the evaluation index value of each candidate recommendation information, the recommendation information of the target object is determined from the at least one candidate recommendation information.

5. The method according to any one of claims 1 to 3, characterized in that The sending the recommendation information to the terminal corresponding to the target object includes: Obtaining preference information of the target object, and determining a target push mode of the recommendation information based on the preference information; The recommendation information is sent to the terminal corresponding to the target object by using the target push method.

6. The method according to any one of claims 1 to 3, characterized in that After sending the recommendation information to the terminal corresponding to the target object, the method further includes: Obtaining second behavior data of the target object within a preset time period; Determining a retention result of the target object based on the first behavior data and the second behavior data; Based on the retention result, the evaluation index value of the recommendation information is updated.

7. The method according to claim 1, characterized in that The determining of the recommended information of the target object based on the group activity and the importance value includes: When the group activity is higher than a preset activity threshold and the importance value is higher than a preset threshold, the target object is determined to be a high-value object type, and the recommended information of the target object is determined to be exclusive activity information; When the group activity is lower than or equal to a preset activity threshold and the importance value is lower than or equal to a preset threshold, the target object is determined to be a low-value object type, and the recommended information of the target object is determined as basic activity reminder information.

8. A churn prediction device, characterized in that: The device comprises: A data acquisition module, configured to acquire first behavior data of a plurality of objects to be predicted in real time, wherein the first behavior data at least includes social data; constructing a social network graph based on the first behavior data of the plurality of objects to be predicted, wherein the nodes of the social network graph are the objects to be predicted, and the edges are the social relationships between the objects to be predicted; Performing graph feature extraction on the social network graph to obtain a graph embedding vector and a node embedding vector corresponding to each of the objects to be predicted, wherein the graph feature extraction adopts a graph embedding method; Determining the graph embedding vector and the node embedding vector as behavioral features of the object to be predicted; A behavior prediction module is used to call the trained target prediction model to perform churn prediction on the behavior characteristics and obtain prediction results; An information recommendation module, configured to determine a target object from a plurality of objects to be predicted based on a prediction result of each object to be predicted; Based on the behavioral characteristics, determining a target social group to which the target object belongs and an importance value of the target object in the target social group; Determining the group activity of the target social group, where the group activity is an average value calculated based on activity scores corresponding to activity data of each social object in the target social group, where the activity data includes login frequency data, average game duration data, and average daily activity data; Determining recommendation information for the target object based on the group activity and the importance value; The information sending module is used to send the recommendation information to the terminal corresponding to the target object.

9. An electronic device, characterized in that: The electronic device comprises: a memory for storing computer-executable instructions; A processor, configured to implement the method according to any one of claims 1 to 7 when executing the computer-executable instructions stored in the memory.

10. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that: When the computer executable instructions or computer program are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

11. A computer program product comprising computer executable instructions or a computer program, characterized in that When the computer executable instructions or computer program are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images