CN112070310A

CN112070310A - Loss user prediction method and device based on artificial intelligence and electronic equipment

Info

Publication number: CN112070310A
Application number: CN202010947833.4A
Authority: CN
Inventors: 钟子宏
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-09-10
Filing date: 2020-09-10
Publication date: 2020-12-11

Abstract

The application provides a method and a device for predicting lost users based on artificial intelligence, electronic equipment and a computer-readable storage medium; data calculation in the technical field of big data is involved; the method comprises the following steps: classifying whether a user sample participating in a service is a lost user of the service or not through a loss parameter and a loss rate classification threshold of a user loss model, and updating the loss parameter and the loss rate classification threshold of the user loss model according to a difference between an obtained prediction category and a mark category of the user sample; and classifying whether the target user is a potential lost user of the service or not according to the updated loss parameters of the user loss model and the loss rate classification threshold. Through the application, the accuracy of predicting lost users can be improved.

Description

Loss user prediction method and device based on artificial intelligence and electronic equipment

Technical Field

The present application relates to artificial intelligence and big data technologies, and in particular, to a method and an apparatus for predicting lost users based on artificial intelligence, an electronic device, and a computer-readable storage medium.

Background

Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. Machine Learning (ML) is the core of artificial intelligence, and is used for specially researching how a computer simulates or realizes human Learning behaviors to acquire new knowledge or skills and reorganize an existing knowledge structure to continuously improve the performance of the computer.

The loss user prediction is an important branch of machine learning, relates to big data processing, and can be applied to service scenes such as social service, game service and e-commerce service. In the scheme provided by the related art, for the model for predicting the attrition users, in the training stage and the application stage of the model, a fixed attrition rate classification threshold value is adopted for classification processing. Due to frequent updating of the service scene, when a fixed attrition rate classification threshold is used, the accuracy of predicting the attrition users is low, and the service operation strategy cannot be effectively adjusted according to the predicted attrition users.

Disclosure of Invention

The embodiment of the application provides a lost user prediction method and device based on artificial intelligence, an electronic device and a computer-readable storage medium, and the accuracy of predicting lost users can be improved.

The technical scheme of the embodiment of the application is realized as follows:

the embodiment of the application provides a loss user prediction method based on artificial intelligence, which comprises the following steps:

classifying whether a user sample participating in a service is a lost user of the service or not through a loss parameter and a loss rate classification threshold of a user loss model, and updating the loss parameter and the loss rate classification threshold of the user loss model according to a difference between an obtained prediction category and a mark category of the user sample;

and classifying whether the target user is a potential lost user of the service or not according to the updated loss parameters of the user loss model and the loss rate classification threshold.

The embodiment of the application provides a loss user prediction device based on artificial intelligence, includes:

the training module is used for classifying whether a user sample participating in a service is a lost user of the service or not through a loss parameter and a loss rate classification threshold of a user loss model, and updating the loss parameter and the loss rate classification threshold of the user loss model according to the difference between the obtained prediction category and the mark category of the user sample;

and the application module is used for classifying whether the target user is a potential lost user of the service or not according to the updated loss parameters and the loss rate classification threshold of the user loss model.

An embodiment of the present application provides an electronic device, including:

a memory for storing executable instructions;

and the processor is used for realizing the artificial intelligence-based loss user prediction method provided by the embodiment of the application when the executable instructions stored in the memory are executed.

The embodiment of the application provides a computer-readable storage medium, which stores executable instructions for causing a processor to execute the method for predicting the lost user based on artificial intelligence provided by the embodiment of the application.

The embodiment of the application has the following beneficial effects:

according to the training set extracted from a plurality of user samples, the loss parameters and the loss rate classification threshold of the user loss model are updated, and model application is performed according to the updated user loss model, namely whether the target user is a potential loss user is predicted, so that even if the service scene is continuously changed, the accuracy of predicting the loss user can be improved through the dynamically updated loss rate classification threshold.

Drawings

FIG. 1 is a diagram illustrating user churn model training, testing, and applications provided in the related art;

FIG. 2 is an alternative architectural diagram of an artificial intelligence based attrition user prediction system provided in an embodiment of the present application;

fig. 3 is an alternative architecture diagram of a terminal device provided in the embodiment of the present application;

FIG. 4A is a schematic flow chart diagram illustrating an alternative method for artificial intelligence based attrition user prediction in accordance with an embodiment of the present application;

FIG. 4B is a schematic flow chart diagram illustrating an alternative method for artificial intelligence based attrition user prediction in accordance with an embodiment of the present application;

FIG. 4C is a schematic flow chart diagram illustrating an alternative method for artificial intelligence based attrition user prediction as provided in an embodiment of the present application;

FIG. 4D is a schematic flow chart diagram illustrating an alternative method for artificial intelligence based attrition user prediction as provided in an embodiment of the present application;

FIG. 4E is a schematic flow chart diagram illustrating an alternative method for artificial intelligence based attrition user prediction as provided in the embodiments of the present application;

FIG. 5 is an alternative diagram of user sample construction in a pre-attrition scenario provided by an embodiment of the present application;

FIG. 6 is an alternative diagram of user sample construction in an attrition recovery scenario provided by an embodiment of the present application;

FIG. 7 is an alternative schematic diagram of user churn model training, testing, and application in a pre-churn scenario as provided by an embodiment of the present application;

fig. 8 is an alternative schematic diagram of user churn model training, testing and application in a churn recovery scenario according to an embodiment of the present application.

Detailed Description

In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be described in further detail with reference to the attached drawings, the described embodiments should not be considered as limiting the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is 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.

In the following description, references to the terms "first \ second \ third" are only to distinguish similar objects and do not denote a particular order, but rather the terms "first \ second \ third" are used to interchange specific orders or sequences, where appropriate, so as to enable the embodiments of the application described herein to be practiced in other than the order shown or described herein. In the following description, the term "plurality" referred to means at least two.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of the present application only and is not intended to be limiting of the application.

Before further detailed description of the embodiments of the present application, terms and expressions referred to in the embodiments of the present application will be described, and the terms and expressions referred to in the embodiments of the present application will be used for the following explanation.

1) User churn model: the two-classification model is constructed based on a machine learning principle, the type of the user attrition model is not limited in the embodiment of the application, and the model may be a Logistic Regression (LR) model, for example. Here, the categories to which the user churn model corresponds may include a churn category and a non-churn category (retention category).

2) Loss parameters: i.e., the weight parameters of the user churn model, can be updated in conjunction with the mechanism of back propagation and gradient descent.

3) Attrition rate classification threshold: in the process of classifying the user loss model, the probabilities of the user belonging to the loss category and the non-loss category are obtained, and the final prediction category can be obtained after the obtained probabilities are compared with the loss rate classification threshold.

4) Loss user classification effect index: the effect index used for evaluating the performance of the user churn model is not limited in the embodiment of the present application, and may be, for example, a recall ratio, a precision ratio, an F1 score, or an Area Under a Curve (AUC), where a Curve refers to a Receiver Operating Characteristic (ROC) Curve.

5) User characteristics: the user characteristics are usually collected in a period unit, for example, 7 days, and can be set according to the actual situation of the service. For example, user characteristics include, but are not limited to, the number of clicks the user makes on the service, the number of collections, the amount paid, the number of payments, the length of the activity period, and the number of days active. The type of the service is not limited in the embodiment of the present application, and may be, for example, a social service (such as an instant messaging application), a game service, or an e-commerce service.

6) The label category: when a user has behavior data in a certain period, namely is in an active state, determining that the marking category of the user in the period is a non-attrition category; when the behavior data does not exist in the user in a certain period, the mark category of the user in the period is determined to be the attrition category.

7) Big Data (Big Data): the data set which can not be captured, managed and processed by a conventional software tool in a certain time range is a massive, high-growth-rate and diversified information asset which can have stronger decision-making power, insight discovery power and flow optimization capability only by a new processing mode. Big data requires special technology to effectively process a large amount of data within a tolerable elapsed time, and technologies suitable for big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the internet and scalable storage systems. In embodiments of the present application, big data technology may be employed to collect user samples.

For the prediction of the attrition users, in the solution provided by the related art, the training, testing and application of the classification model are usually performed according to a fixed classification threshold, wherein the fixed classification threshold is usually set manually according to experience. As an example, a schematic diagram of predicting the attrition users in the related art shown in fig. 1 is provided, in fig. 1, the user characteristics of the T-1 th period and the classification labels (i.e., label categories) of the T-th period are used as constituent elements of a user sample, so as to construct a training user sample and a testing user sample, perform model training according to the training user sample and a fixed classification threshold, and determine a training effect of the model according to the testing user sample and the fixed classification threshold. And when the training effect reaches the standard, processing the user characteristics of the T-th period according to the trained classification model and the fixed classification threshold value so as to determine the prediction category of the user corresponding to the user characteristics.

The solutions provided by the related art have the following problems:

1) the fixed classification threshold value cannot automatically adapt to the influence brought by the numerical value updating of the user features, and under the condition that the types of the features included in the user features are the same, the influence degrees of the feature values of the user features on the categories in different periods are different. For example, in the game business, the game active duration of the user in the T-th period is generally longer than that of the T-1 th period, the recharge and the consumption are generally more, and the like, possibly due to the influence of the factors of updating the game version, being more friendly to experience, intervening activities and the like. If a fixed classification threshold is adopted, the influence caused by updating of the user characteristic numerical value can be ignored, and the accuracy of predicting lost users is low.

2) For example, after an original LR model is replaced with a Support Vector Machine (SVM) model, if the same fixed classification threshold as the LR model is used, the classification effect of the SVM model is poor.

3) The fixed classification threshold is also not suitable for updating the service scene, for example, the service scene includes a pre-churning scene and a churning recovery scene, where the pre-churning scene is to predict whether a target user who is not churning will churn or not, and the churning recovery scene is to predict whether a target user who has churned will continue churning or not, and if the two scenes use the same fixed classification threshold, the predicted churning user will be inaccurate because the user groups targeted by the two scenes are not the same.

The embodiment of the application provides a lost user prediction method and device based on artificial intelligence, an electronic device and a computer-readable storage medium, which can improve the accuracy of predicting lost users and facilitate more effective development of business operation activities. An exemplary application of the electronic device provided in the embodiment of the present application is described below, and the electronic device provided in the embodiment of the present application may be implemented as various types of terminal devices such as a notebook computer, a tablet computer, a desktop computer, a set-top box, a mobile device, and the like, and may also be implemented as a server.

Referring to fig. 2, fig. 2 is an alternative architecture diagram of the artificial intelligence-based attrition user prediction system 100 according to the embodiment of the present application, in which a terminal device 400 is connected to a server 200 through a network 300, and the server 200 is connected to a database 500, where the network 300 may be a wide area network or a local area network, or a combination of both.

In some embodiments, taking the electronic device as a terminal device as an example, the artificial intelligence-based churn user prediction method provided by the embodiments of the present application may be implemented by the terminal device. For example, the terminal device 400 trains and tests the user churn model stored locally through a plurality of user samples stored locally, and when the obtained churn user classification effect index is greater than the churn user index threshold, classifies the target user through the churn parameter and the churn rate classification threshold of the updated user churn model. The terminal device 400 may construct the user sample according to the locally stored user log file related to the service, which certainly does not limit the embodiment of the present application, for example, the user sample may also be directly obtained by the terminal device 400 from the outside (e.g., in the internet).

In some embodiments, taking the electronic device as a server as an example, the artificial intelligence-based attrition user prediction method provided in the embodiments of the present application may also be implemented by the server. For example, the server 200 trains and tests the user churn model according to a plurality of user samples acquired from the database 500, and when the obtained churn user classification effect index is greater than the churn user index threshold, classifies the target user according to the churn parameter and the churn rate classification threshold of the updated user churn model. Here, the server 200 may send the prediction category to which the target user belongs to the terminal device 400, or may send the updated churn parameter and churn rate classification threshold of the user churn model to the terminal device 400, so that the terminal device 400 performs the classification process on the target user locally. It should be noted that, in the embodiment of the present application, the storage location of the user sample is not limited to the database, and for example, the user sample may also be a location such as a distributed storage system of the server 200 or a block chain, and the user churn model is the same.

Terminal device 400 is configured to display various results and final results of the attrition user prediction process in graphical interface 410. In fig. 2, by taking a pre-churn scenario in a game service as an example, a user sample obtained by the server 200 from the database 500 includes a user feature of a tagged user in a T-1 th period and a tag category of the T-th period, the server 200 performs training and testing of a user churn model according to a plurality of user samples, and the terminal device 400 sends the user feature of a target user (for example, a user in a login state in a game application of the terminal device 400) in the T-th period to the server 200. If the server 200 classifies the target user according to the updated churn parameter and churn rate classification threshold of the user churn model, and the target user is obtained to belong to a churn category, the server 200 allocates a game gift bag (including virtual game resources in the game application, such as equipment, props, coins, and the like) to the target user according to the set operation policy, so that the terminal device 400 displays the allocated game gift bag in the game application of the graphical interface 410. Therefore, the probability that the target user actually becomes the lost user can be reduced, and the user viscosity of the service is enhanced.

In some embodiments, the server 200 may be an independent physical server, or may be a server cluster or a distributed system formed by a plurality of physical servers, or may be a cloud server providing basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a network service, cloud communication, a middleware service, a domain name service, a security service, a CDN, and a big data and artificial intelligence platform, where the cloud service may be an attrition user prediction service, which is called by the terminal device 400, to obtain a prediction category to which a target user belongs by classifying according to a user characteristic of the target user sent by the terminal device 400, and send the prediction category to the terminal device 400, so that the terminal device 400 determines an operation manner for the target user according to the prediction category and a business operation policy, and of course, the server 200 may also determine the operation manner directly according to the prediction category to which the target user belongs, and transmits the operation manner to the terminal device 400. The terminal device 400 may be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, and the like. The terminal device and the server may be directly or indirectly connected through wired or wireless communication, and the embodiment of the present application is not limited.

Taking the electronic device provided in the embodiment of the present application as an example for illustration, it can be understood that, for the case where the electronic device is a server, parts (such as the user interface, the presentation module, and the input processing module) in the structure shown in fig. 3 may be default. Referring to fig. 3, fig. 3 is a schematic structural diagram of a terminal device 400 provided in an embodiment of the present application, where the terminal device 400 shown in fig. 3 includes: at least one processor 410, memory 450, at least one network interface 420, and a user interface 430. The various components in the terminal device 400 are coupled together by a bus system 440. It is understood that the bus system 440 is used to enable communications among the components. The bus system 440 includes a power bus, a control bus, and a status signal bus in addition to a data bus. For clarity of illustration, however, the various buses are labeled as bus system 440 in FIG. 3.

The Processor 410 may be an integrated circuit chip having Signal processing capabilities, such as a general purpose Processor, a Digital Signal Processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, or the like, wherein the general purpose Processor may be a microprocessor or any conventional Processor, or the like.

The user interface 430 includes one or more output devices 431, including one or more speakers and/or one or more visual displays, that enable the presentation of media content. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid state memory, hard disk drives, optical disk drives, and the like. Memory 450 optionally includes one or more storage devices physically located remote from processor 410.

The memory 450 includes either volatile memory or nonvolatile memory, and may include both volatile and nonvolatile memory. The nonvolatile Memory may be a Read Only Memory (ROM), and the volatile Memory may be a Random Access Memory (RAM). The memory 450 described in embodiments herein is intended to comprise any suitable type of memory.

In some embodiments, memory 450 is capable of storing data, examples of which include programs, modules, and data structures, or a subset or superset thereof, to support various operations, as exemplified below.

An operating system 451, including system programs for handling various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and handling hardware-based tasks;

a network communication module 452 for communicating to other computing devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 including: bluetooth, wireless compatibility authentication (WiFi), and Universal Serial Bus (USB), etc.;

a presentation module 453 for enabling presentation of information (e.g., user interfaces for operating peripherals and displaying content and information) via one or more output devices 431 (e.g., display screens, speakers, etc.) associated with user interface 430;

an input processing module 454 for detecting one or more user inputs or interactions from one of the one or more input devices 432 and translating the detected inputs or interactions.

In some embodiments, the apparatus provided by the embodiments of the present application may be implemented in software, and fig. 3 illustrates an artificial intelligence based attrition user prediction apparatus 455 stored in a memory 450, which may be software in the form of programs and plug-ins, and the like, and includes the following software modules: a training module 4551 and an application module 4552, which are logical and thus may be arbitrarily combined or further split depending on the functions implemented. The functions of the respective modules will be explained below.

In other embodiments, the apparatus provided in this Application may be implemented in hardware, and for example, the artificial intelligence based attrition user prediction apparatus provided in this Application may be a processor in the form of a hardware decoding processor, which is programmed to execute the artificial intelligence based attrition user prediction method provided in this Application, for example, the processor in the form of a hardware decoding processor may be one or more Application Specific Integrated Circuits (ASICs), DSPs, Programmable Logic Devices (PLDs), Complex Programmable Logic Devices (CPLDs), Field Programmable Gate Arrays (FPGAs), or other electronic components.

The artificial intelligence based attrition user prediction method provided by the embodiment of the present application will be described in conjunction with exemplary applications and implementations of the electronic device provided by the embodiment of the present application. According to the above description, the churn user prediction method below may be performed by a terminal device or a server.

Referring to fig. 4A, fig. 4A is an alternative flowchart of an artificial intelligence-based attrition user prediction method provided in this embodiment of the present application, which will be described with reference to the steps shown in fig. 4A.

In step 101, a user sample participating in a service is classified according to a loss parameter and a loss rate classification threshold of a user loss model, and the loss parameter and the loss rate classification threshold of the user loss model are updated according to a difference between an obtained prediction category and a labeled category of the user sample.

In the embodiment of the present application, the user churn model includes two types of parameters, one type is churn parameter, which may be, for example, a weight parameter of the user churn model, and the other type is churn rate classification threshold. After the user samples participating in the service are obtained, the user samples are classified through the loss parameters and the loss rate classification threshold of the user loss model to obtain the prediction categories. Then, updating the attrition parameter of the user attrition model and the attrition rate classification threshold value according to the difference between the prediction category and the mark category included in the user sample, for example, determining a loss value according to the difference between the prediction category and the mark category of the user sample, and then updating the attrition parameter according to the loss value; the attrition rate classification threshold may be updated based on the classification error rates of the plurality of user samples, as described in more detail below.

In some embodiments, the user sample includes user characteristics and tag categories for the tagged user; the categories corresponding to the user churn model comprise a churn category and a non-churn category; the above-mentioned classification process of whether the user sample participating in the service is the lost user of the service or not can be realized by the loss parameter and the loss rate classification threshold of the user loss model in this way: performing probability mapping processing on user characteristics included in a user sample participating in a service through loss parameters of a user loss model to obtain the probability that a marked user corresponding to the user sample belongs to a loss category; when the probability of belonging to the loss category is larger than the loss rate classification threshold of the user loss model, taking the loss category as the prediction category of the marked user corresponding to the user sample; and when the probability of belonging to the loss category is less than or equal to the loss rate classification threshold of the user loss model, taking the non-loss category as the prediction category of the marked user corresponding to the user sample.

Here, when classifying user samples participating in a service, probability mapping processing may be performed on user characteristics included in the user samples through churn parameters of the user churn model, so as to obtain probabilities that the labeled users corresponding to the user samples belong to churn categories, where a sum of the probability that the user belongs to a churn category and the probability that the user belongs to a non-churn category is 1, that is, the probability that the user belongs to a non-churn category may also be obtained at the same time.

The case of the probability of belonging to the attrition category corresponding to the attrition rate classification threshold is taken as an example. When the probability of belonging to the loss category is larger than the loss rate classification threshold of the user loss model, taking the loss category as the prediction category of the marked user corresponding to the user sample; and when the probability of belonging to the loss category is less than or equal to the loss rate classification threshold of the user loss model, taking the non-loss category as the prediction category of the marked user corresponding to the user sample.

Alternatively, the attrition rate classification threshold is the probability of corresponding belonging to a non-attrition category. In this case, when the probability of belonging to the non-attrition category is greater than the attrition rate classification threshold, the non-attrition category is used as the prediction category of the labeled user corresponding to the user sample; and when the probability of belonging to the non-attrition category is less than or equal to the attrition rate classification threshold, taking the attrition category as the prediction category of the marked user corresponding to the user sample.

In some embodiments, the user sample includes user characteristics (including) of tagged users belonging to the non-attrition category at the T-1 th cycle, and a tagged category of tagged users at the T-th cycle; the marker category is one of a non-attrition category and an attrition category; or the user sample comprises the user characteristics of the marked users belonging to the non-attrition category in the T-2 th period and the marked categories of the marked users in the T-2 th period; marking users belonging to the attrition category in the T-1 period; the marker class is one of a non-attrition class and an attrition class.

As an example, the user characteristics include at least one of: basic attribute data such as user gender, age, region and the like, active attribute data representing the activity degree of the service, recharging attribute data recharged for using the service, behavior attribute data of stage behaviors in the process of using the service, and commodity prop attributes of commodities related in the process of using the service.

The embodiment of the application provides two scenes, namely a pre-attrition scene and an attrition retrieval scene, wherein for the pre-attrition scene, a user sample comprises user characteristics of a marked user belonging to a non-attrition category in the T-1 th period and a marked category of the marked user in the T-th period, wherein, the marking user belongs to the non-loss category in the T-1 period means that the marking user participates in the service in the T-1 period, namely, the behavior data is formed in the service, the behavior data comprises the relevant record data of the behavior implemented by the user in the process of using the service, such as the number of people/times an online social communication was made in a social business, the duration of each communication, or the type and duration of game play in the game service, or the type of merchandise involved in merchandise click or merchandise purchase in the e-commerce service, etc. Correspondingly, if a certain user belongs to the churn category in the T-1 th period, the user does not participate in the service in the T-1 th period, that is, no behavior data is formed. The mark category refers to a determined category, specifically to one of an attrition category and a non-attrition category, and can be labeled by human. In the pre-churn scenario, when the subsequent model is applied, classification processing of the target user can be performed according to the user characteristics of the target user in the T-th period, that is, the prediction category of the target user in the T + 1-th period is determined, so that the operation strategy of the target user is determined in a targeted manner.

For the attrition retrieval scenario, the user sample comprises user characteristics of the tagged users belonging to the non-attrition category at the T-2 th period and the tagged categories of the tagged users at the T-1 th period, wherein the tagged users belong to the attrition category at the T-1 th period. In the loss recovery scenario, when the subsequent model is applied, classification processing of the target user can be performed according to the user characteristics of the target user in the T-1 th period (on the premise that the target user belongs to the loss category in the T-1 th period), that is, the prediction category of the target user in the T +1 th period is determined, so that the operation strategy for the target user is determined in a targeted manner. By the method, the flexibility of user sample construction is improved, and the applicability of the embodiment of the application to different service scenes is also improved.

In step 102, a classification process is performed on whether the target user is a potentially traffic churning user according to the churning parameter and the churning rate classification threshold of the updated user churning model.

After the loss parameters and the loss rate classification threshold of the user loss model are updated through the user sample, the target user can be classified according to the updated loss parameters and the updated loss rate classification threshold of the user loss model, and whether the target user is a service potential loss user or not is determined. If the target user belongs to the loss category, the target user is proved to be a potential loss user in the service, and the target user has a higher probability of losing; if the target user belongs to the non-loss category, the target user is proved to have a higher probability of being retained in the service.

After the prediction category to which the target user belongs is obtained, service operation can be performed in a targeted manner, so that the probability that the target user actually becomes a lost user is reduced. For example, in a social service based on a social application, a prompt message (sending method such as a short message or a mail) may be sent to a target user belonging to the attrition category, for example, the prompt message may include the number of unread messages of the target user in the social application, so that the user may be prompted to open the social application to participate in the social service; in the game business based on the game application, a game gift package can be distributed to target users belonging to the attrition category in the game application, and the game gift package comprises virtual game resources, such as equipment, props, gold coins and the like, which can be used by the target users, so that the target users are attracted to continue to remain in the game application; in the e-commerce business based on the e-commerce application, a commodity coupon or a shopping red package, etc. may be allocated to a target user belonging to the attrition category in the e-commerce application to incentivize the target user to purchase goods using the commodity coupon or the shopping red package in the e-commerce application. The above is only some examples of the service operation policy, and other service operation policies may also be adopted according to different actual application scenarios.

As shown in fig. 4A, in the embodiment of the present application, the churn parameter and churn rate classification threshold of the user churn model are dynamically updated according to the user sample participating in the service, so that even in the case of user feature value update, model version update, or service scene update, it can be ensured that the accuracy of predicting churn users is high enough, and effective development of service operation is facilitated.

In some embodiments, referring to fig. 4B, fig. 4B is an optional flowchart of the artificial intelligence-based attrition user prediction method provided in this embodiment of the present application, before step 101 shown in fig. 4A, in step 201, a plurality of user samples participating in a business may be extracted to obtain a training set and a test set; the training set is used for updating the attrition parameters and the attrition rate classification threshold of the user attrition model.

Here, the extraction ratios corresponding to the training set and the test set, respectively, may be preset, for example, the extraction ratio corresponding to the training set is 4/5, and the extraction ratio corresponding to the test set is 1/5; or only setting the extraction ratio corresponding to the training set, and then determining the number of the test user samples included in the test set (i.e. determining the extraction ratio corresponding to the test set) according to the sample number ratio between the training set and the test set, where the sample number ratio may be set according to an actual application scenario, for example, set to 8: 2.

and extracting a plurality of user samples participating in the service according to the extraction ratio to obtain a training set and a test set, wherein the extraction process can be a random extraction process with a release.

In some embodiments, after step 201, further comprising: when the matching of the quantity proportion and the proportion condition between the training user samples of different label types in the training set fails, a plurality of user samples are extracted, and the extracted user samples are added to the training set, so that the matching of the quantity proportion and the proportion condition in the new training set is successful.

After the training set is obtained through the extraction process, the training set may be subjected to a sample equalization process. For example, the proportion condition is set as the number of training user samples in the training set including the attrition category: the number of training user samples in the training set including the non-attrition category is 1: 1, that is, the number of positive and negative samples is the same, however, this does not limit the embodiments of the present application, and other proportional conditions may be set according to the actual application scenario. When the matching of the quantity proportion between the training user samples of different label types in the training set and the proportion condition fails, the plurality of user samples are extracted, for example, the extraction with the replacement is carried out, and the extracted user samples are added to the training set, so that the matching of the quantity proportion in the new training set and the proportion condition is successful. For example, the number of training user samples in the training set including the attrition category is 100, the number of training user samples in the training set including the non-attrition category is 95, and the ratio of the number in the training set to 1: 1, determining the missing training user samples of the non-loss categories if the matching of the proportion conditions fails, performing replaced random extraction processing on the original multiple user samples including the non-loss categories, and adding the extracted 5 user samples to a training set so that the proportion of the number in a new training set meets the proportion conditions. By the mode, sample balance of the training set is realized, and the subsequent effect of model training can be effectively improved.

In some embodiments, the user sample includes user characteristics and tag categories for the tagged user; after step 201, the method further includes: performing principal component analysis processing on user characteristics included in a training user sample in a training set, and taking the obtained principal component characteristics as new user characteristics; wherein the principal component features include a feature type that is less than a feature type included in the user features.

Because the user features in the training user sample may include more feature types, in this embodiment of the present application, Principal Component Analysis (PCA) processing may be performed on the user features included in the training user sample in the training set, and the Principal Component features obtained after the PCA processing may be used as new user features. Therefore, the feature types can be reduced, and the convergence rate of the model can be effectively improved during subsequent model training, namely the training efficiency is improved.

In fig. 4B, step 101 shown in fig. 4A may be updated to step 202, in step 202, a classification process of whether the training user samples in the training set are the loss users of the service is performed according to the loss parameters and the loss rate classification thresholds of the user loss model, and the loss parameters and the loss rate classification thresholds of the user loss model are updated according to the difference between the obtained prediction categories and the labeled categories of the training user samples.

After the training set and the testing set are obtained through step 201, the attrition parameters and the attrition rate classification threshold of the user attrition model are updated through the training user samples in the training set.

In fig. 4B, before step 102 shown in fig. 4A, in step 203, a classification process is performed on the test user sample in the test set according to the updated churn parameter and churn rate classification threshold of the user churn model to determine whether the test user sample is a churn user of the service, and an churn user classification effect index is determined according to a difference between the obtained prediction category and the mark category of the test user sample.

After model training by the training set, model testing is performed by the test set. For example, the user loss model is updated according to the loss parameters and the loss rate classification threshold, and the user samples in the test set are classified to obtain the prediction categories. And then, determining a loss user classification effect index according to the difference between the prediction type and the mark type included in the test user sample, wherein the larger the loss user classification effect index is, the better the training effect of the user loss model is. The index of the attrition user classification effect may be recall, precision, F1 score or AUC value, but this does not limit the embodiment of the present application.

After step 203, in step 204, it is determined that the user churn model update is complete when the churn user classification effect index is greater than the churn user index threshold.

Here, a loss user index threshold is preset, and if the loss user classification effect index obtained in step 203 is less than or equal to the loss user index threshold, the next iteration may be performed based on the existing training set and test set, or based on the training set and test set obtained by re-extraction, until the loss user classification effect index is greater than the loss user index threshold.

And if the loss user classification effect index is larger than the loss user index threshold, determining that the training stage of the user loss model is finished, and entering a stage of model application, namely classifying the target users through the updated loss parameters and the loss rate classification threshold of the user loss model to obtain the prediction categories of the target users.

In some embodiments, the number of the user churn models is multiple, and the model structures of different user churn models are different; the above-mentioned classification process of whether the target user is a potentially traffic-losing user or not through the loss parameters and the loss rate classification thresholds of the updated user loss model can be realized in this way: taking the user loss model with the largest loss user classification effect index in the plurality of updated user loss models as a target user loss model; and classifying whether the target user is a service potential lost user or not according to the loss parameters and the loss rate classification threshold of the target user loss model.

In the embodiment of the application, a plurality of user churn models can be trained and tested simultaneously, and model structures of different user churn models are different, for example, if the user churn models include an LR model and an SVM model, steps 201 to 203 can be executed on the LR model, and steps 201 to 203 can be executed on the SVM model, and processing processes of the two models are isolated from each other and do not interfere with each other; as another example, the user churn model includes a neural network model with a network layer number of 100 and a neural network model with a network layer number of 200.

And when the loss user classification effect index of each updated user loss model is greater than the loss user index threshold, taking the user loss model with the largest loss user classification effect index in all the updated user loss models as a target user loss model, and performing model application according to the target user loss model, namely performing classification processing on whether the target user is a loss user with potential service or not. Through the mode, the model structure with the best effect can be screened out, and the accuracy of predicting lost users is further improved.

As shown in fig. 4B, in the embodiment of the present application, by constructing the training set and the test set, a user loss model meeting an expected effect can be obtained, and the accuracy of the model in application is further improved.

In some embodiments, referring to fig. 4C, fig. 4C is an optional flowchart of the artificial intelligence-based attrition user prediction method provided in this embodiment of the present application, in fig. 4C, step 201 shown in fig. 4B may be updated to step 301, and in step 301, a plurality of user samples participating in a business are extracted according to an extraction ratio, so as to obtain a training set and a test set.

The extraction ratio can be preset according to the actual application scenario.

In fig. 4C, step 202 shown in fig. 4B can be updated to step 302, in step 302, a training user sample in the training set is classified as a loss user of the service according to the loss parameter and the loss rate classification threshold of the user loss model, and the loss parameter, the loss rate classification threshold and the extraction ratio of the user loss model are updated according to the difference between the obtained prediction category and the labeled category of the training user sample.

On the basis of obtaining the training set and the test set by extraction according to the extraction proportion, in the stage of model training, besides updating the loss parameters and the loss rate classification threshold of the user loss model, the extraction proportion can be updated according to the difference between the prediction class and the label class of the training user samples, so that in the next iteration, the training set with more proper number of the training user samples and the test set with more proper number of the test user samples are extracted, and the training effect of the user loss model is favorably improved.

In fig. 4C, after step 203 shown in fig. 4B, in step 303, when the attrition user classification effect index is smaller than or equal to the attrition user index threshold, the extraction processing and training are performed again on the plurality of user samples according to the updated extraction ratio until the obtained new attrition user classification effect index is larger than the attrition user index threshold.

When the loss user classification effect index is smaller than or equal to the loss user index threshold value, the next iteration is started, firstly, a plurality of user samples are extracted again according to the updated extraction proportion, then, model training is carried out according to the obtained new training set, model testing is carried out according to the obtained new testing set, and finally, a new loss user classification effect index is obtained. And iterating until the obtained new attrition user classification effect index is larger than the attrition user index threshold value.

As shown in fig. 4C, the embodiment of the present application, by updating the extraction ratio, helps to accelerate convergence of the user churn model in the subsequent iteration process.

In some embodiments, referring to fig. 4D, fig. 4D is an optional flowchart of the artificial intelligence-based attrition user prediction method provided in this embodiment of the present application, and step 302 shown in fig. 4C may be implemented through steps 401 to 403, which will be described in detail with reference to these steps.

In step 401, a training user sample in the training set is classified according to the loss parameter and the loss rate classification threshold of the user loss model to obtain a prediction category.

In step 402, for each training user sample in the training set, a loss calculation process is performed on a difference between a prediction category and a label category of the training user sample, a loss value obtained by the loss calculation process is propagated backwards in the user churn model, and a churn parameter of the user churn model is updated in the process of the backward propagation.

In the updating process of the loss parameters, for each training user sample in the training set, loss calculation processing is performed on the difference between the prediction category and the label category of the training user sample through a loss function of the user loss model to obtain a loss value, wherein the type of the loss function is not limited, and may be, for example, a cross entropy loss function. And then updating the loss parameters of the user loss model according to the obtained loss values, wherein in the embodiment of the application, the loss parameters of the user loss model can be updated based on the loss values of all training user samples in the training set by combining a reverse propagation mechanism and a gradient descent mechanism. The Gradient decrease may be a Batch Gradient Decrease (BGD), a random Gradient decrease (SGD), or a Mini-Batch Gradient decrease (MBGD).

In step 403, a ratio between the number of training user samples with different prediction categories and labeling categories in the training set and the total number of training user samples in the training set is determined as a classification error rate, and a churn rate classification threshold and an extraction ratio of the user churn model are updated according to the classification error rate.

In the updating process of the attrition rate classification threshold and the extraction ratio of the user attrition model, firstly, the ratio of the number of training user samples with different prediction classes and label classes (namely, classification errors) in the training set to the total number of training user samples in the training set is determined as the classification error rate. Then, the attrition rate classification threshold and the extraction ratio are updated according to the classification error rate, and the updating manner can be adjusted according to the actual application scenario, for example, a negative correlation relationship is set between the updated extraction ratio and the classification error rate, and a positive correlation relationship is set between the updated attrition rate classification threshold and the classification error rate.

In some embodiments, the above-mentioned updating of the attrition rate classification threshold and the extraction ratio of the user attrition model according to the classification error rate may be implemented by: determining a division result between a result obtained by subtracting the classification error rate from 1 and the classification error rate; carrying out logarithm processing on the division result, and taking a first product result between the result of the logarithm processing and a set coefficient as an updated extraction proportion; and determining a second product result between the attrition rate classification threshold and the updated extraction proportion, and subtracting the second product result from the attrition rate classification threshold to obtain the updated attrition rate classification threshold.

The embodiment of the present application provides an example of updating the attrition rate classification threshold and the extraction ratio, and first, the classification error rate is subtracted from 1, and the obtained result is divided by the classification error rate itself to obtain a division result. Then, the division result is subjected to logarithm processing, and the result of multiplication between the result of logarithm processing and a setting coefficient (named as a first multiplication result for convenience of distinction) is used as an updated extraction ratio, wherein the setting coefficient can be specifically set according to the actual application scenario, for example, the setting coefficient can be 1/2 on the basis of the result of logarithm processing on the basis of a natural constant on the division result. After the updated extraction proportion is obtained, determining a product result (named as a second product result for convenience of distinguishing) between the attrition rate classification threshold and the updated extraction proportion, and subtracting the second product result from the attrition rate classification threshold to obtain the updated attrition rate classification threshold. Therefore, according to the updated extraction proportion and the updated attrition rate classification threshold, the effect of model training can be effectively improved in the next iteration.

As shown in fig. 4D, in the embodiment of the present application, the churn parameter of the user churn model is updated based on the loss value obtained by the loss calculation processing; based on the classified error rate obtained by statistics, the loss rate classification threshold and the extraction proportion of the user loss model are updated, so that the parameter updating is accurately and effectively realized, and the rapid convergence of the user loss model is facilitated.

In some embodiments, referring to fig. 4E, fig. 4E is an optional flowchart of the artificial intelligence-based attrition user prediction method provided in this embodiment of the present application, and step 301 shown in fig. 4C may be implemented by steps 501 to 502, which will be described in conjunction with the steps.

In step 501, the reciprocal of the number of the set training sets is used as an extraction proportion, and a plurality of user samples are subjected to replaced random extraction processing to obtain training sets meeting the number of the set training sets; wherein the number of training sets is set to be an integer greater than 1.

In the embodiment of the present application, the number of training sets may be multiple, for example, the total number of training sets in each iteration may be set to be constant, that is, the number of training sets is set to be an integer greater than 1. In addition, each training set corresponds to a decimation ratio, and the initial decimation ratio (i.e., the decimation ratio in the first iteration) is the inverse of the number of the set training sets.

And when the extraction is carried out, respectively carrying out replaced random extraction on a plurality of user samples according to a plurality of extraction ratios in the iteration to obtain training sets in accordance with the set training set quantity. For example, if the number of training sets is set to 10 and the total number of original user samples is set to 1000, then in the first iteration, the corresponding extraction ratio of each training set is 1/10.

In step 502, for each training set, the number of test user samples is determined according to the number of training user samples included in the training set and the sample number ratio between the training set and the test set, and the replaced random extraction processing is performed on a plurality of user samples according to the number of test user samples, so as to obtain a test set corresponding to the training set.

Here, the ratio of the number of samples between the training set and the test set may be preset, for example, to 8: 2. after the training sets meeting the set training set number are obtained through step 501, for each training set, the number of the test user samples is determined according to the number of the training user samples in the training set and the sample number ratio between the training set and the test set, and then the replaced random extraction processing is performed on the plurality of user samples according to the number of the test user samples, so that the test set corresponding to the training set is obtained. Finally, a test set meeting the set training set number can be obtained.

In fig. 4E, step 302 shown in fig. 4C may be updated to step 503, in step 503, a training user sample in the target training set is classified according to the churn parameter and churn rate classification threshold of the user churn model, and the churn parameter, churn rate classification threshold, and extraction ratio of the user churn model are updated according to the difference between the obtained prediction category and the labeled category of the training user sample; wherein, the target training set is any one of the training sets which accord with the set training set quantity.

For example, if the number of training sets is set to 10, the 1 st training set of the 10 training sets may be set as the target training set.

In fig. 4E, step 203 shown in fig. 4C may be updated to step 504, in step 504, the testing user sample in the target testing set is subjected to classification processing on whether the testing user sample is a lost user of a service according to the updated loss parameter and the loss rate classification threshold of the user loss model, and an index of the classification effect of the lost user is determined according to the difference between the obtained prediction category and the mark category of the testing user sample; and the target test set and the target training set have a corresponding relation.

Since the extracted training set and the test set have a one-to-one correspondence relationship, after model training is performed through the target training set, model testing is performed according to the test set corresponding to the target training set (named as a target test set for convenience of distinction). If the model training is performed according to the 1 st training set of the 10 training sets in step 503, the model testing is performed according to the 1 st test set of the 10 test sets.

In fig. 4E, after step 504, in step 505, when the attrition user classification effect index corresponding to the target test set is less than or equal to the attrition user index threshold and there is a training set that is not subjected to classification processing, the training user samples in the training set that is not subjected to classification processing are classified according to the updated attrition parameter and attrition rate classification threshold of the user attrition model, so as to obtain the attrition user classification effect index corresponding to the test set that is not subjected to classification processing.

Taking the target test set as the 1 st test set of the 10 test sets for example, when the loss user classification effect index corresponding to the target test set is less than or equal to the loss user index threshold, because there is a training set (i.e. the 2 nd to 9 th training sets of the 10 training sets) that is not classified, classifying the training user samples in the 2 nd training set by the updated loss parameters and loss rate classification threshold of the user loss model, i.e. performing model training, then performing model testing on the user loss model trained based on the 2 nd training set by the 2 nd test set to obtain the loss user classification effect index corresponding to the 2 nd test set, and so on until obtaining the loss user classification effect index greater than the loss user index threshold.

In fig. 4E, step 303 shown in fig. 4C may be updated to step 506, and in step 506, when the attrition user classification effect indexes corresponding to all the test sets are less than or equal to the attrition user index threshold, the replaced random extraction processing and training are performed again on the plurality of user samples according to the updated extraction ratio until the obtained new attrition user classification effect index is greater than the attrition user index threshold.

For example, if the classification effect indexes of the lost users corresponding to 10 test sets are all less than or equal to the threshold of the index of the lost user, the next iteration is performed, that is, the random extraction processing with the replace is performed on a plurality of user samples respectively according to the updated extraction proportion corresponding to each training set to obtain new training sets meeting the set training set quantity, and new test sets meeting the set training set quantity are obtained through corresponding extraction, model training is performed according to the new training sets, and model testing is performed according to the new test sets until the classification effect indexes of the lost users corresponding to a certain new test set are greater than the threshold of the index of the lost user.

As shown in fig. 4E, in each iteration, the embodiment of the application performs model training through multiple training sets, and performs model testing through multiple testing sets, so that effective convergence of the user loss model can be realized, and the model training effect is improved.

Next, an exemplary application of the embodiments of the present application in an actual application scenario will be described. The embodiment of the application can be applied to a pre-attrition scenario and an attrition retrieval scenario in digital marketing, and the digital marketing can be a marketing activity in applications such as a social application, a game application or an e-commerce application, but is not limited thereto, and the example is given by the digital marketing in the game application.

In the pre-churn scenario, if a user is in an active state (i.e., belongs to a non-churn category) in a T-1 th period but in an inactive state (i.e., belongs to a churn category) in a T-1 th period, the user is marked to churn in the T-th period (a corresponding classification label is marked as 1); and if the user is in an active state in both the T-1 th period and the Tth period, marking the user to remain in the Tth period (the corresponding classification label is marked as 0). Because the user marked as attrition has no user characteristics in the Tth period, a user sample in the pre-attrition scene is constructed according to the user characteristics of the user in the T-1 th period and the classification label of the user in the Tth period, wherein the classification label is the attrition-retention classification label. Wherein the classification label corresponds to the above label category.

In the loss recovery scene, if a user is in an active state in a T-2 th period, in an inactive state in a T-1 th period, and in an active state in a T-th period, marking the user to reflow in the T-th period (a corresponding classification label is marked as 1); and if the user is in an active state in the T-2 th period, and is in an inactive state in the T-1 th period and the Tth period, marking that the user does not reflow in the Tth period (the corresponding classification label is marked as 0). Because the user marked as no backflow has no user characteristics in the T-1 th period and the T-th period, a user sample in the loss retrieval scene is constructed according to the user characteristics of the user in the T-2 th period and the classification label of the user in the T-th period, wherein the classification label is the backflow-no backflow classification label.

Next, a training process of the classification model (corresponding to the user churn model above) in the embodiment of the present application is described by taking the pre-churn scenario as an example. The embodiment of the present application provides a schematic diagram of training, testing and applying a classification model in a pre-churn scene as shown in fig. 7, which mainly includes 5 steps of user sample preprocessing, user sample resampling, weight parameter (corresponding to the churn parameter above) updating, dynamic threshold updating and classification prediction, and will be described according to each step:

1) and (4) preprocessing a user sample. In this step, a user sample is extracted from a data source (e.g., a database storing log files of game applications), that is, for each user, a user sample is constructed using the user characteristics in the T-1 th period and the classification tags of the T-th period, so as to obtain an original user sample set.

The user characteristics include but are not limited to basic attribute data such as gender, age and region of the user, active attribute data such as active days, active duration, active game quantity, time interval between registration time and current time, charge amount, consumption amount, charge times, charge days, time interval between first charge time and current time, behavior attribute data such as user clicking, installing, uninstalling and uninstalling, and commodity property attributes such as gift bag/gift certificate type (quantity, times and value), use gift bag/gift certificate type (quantity and value) and expiration gift bag/gift certificate type (quantity and value). In the pre-churn scene, when the classification label of the user in the Tth period is 1, the user is churn in the Tth period; when the classification label of the user in the Tth period is 0, the user is indicated to be retained in the Tth period.

2) The user samples are resampled. Here, E is a sample weight (corresponding to the above decimation ratio)_l-1And performing random extraction processing with putting back on the user sample set to obtain N training sets, wherein TR is { train ═ train₁,train₂,...,train_NN is an integer greater than 1, which can be set according to the actual application scenario, E_l-1Wherein, one iteration refers to the process of carrying out model training according to all training sets and carrying out model testing according to all testing sets.

Each training set corresponds to a sample weight, and all the training sets correspond to an initial sample weight E₀Are all made of

I.e. in the first iteration, in accordance with

The extraction proportion of (2) is that random extraction processing with replacement is carried out on the user sample set to obtain train₁I.e. train₁The ratio between the number of included training user samples and the total number of user samples included in the set of user samples is

Then according to

The extraction proportion of (2) is that random extraction processing with replacement is carried out on the user sample set to obtain train₂And so on.

In addition, for each obtained training set, the number of the tested user samples is determined according to the sample number proportion between the training set and the testing set, and then the user sample set is subjected to replaced random extraction processing according to the number of the tested user samples, so that the testing set corresponding to the training set is obtained. Thus, TE ═ test can be obtained₁,test₂,...,test_NIn which test₁Corresponding train₁，test₂Corresponding train₂And so on. Wherein, the sample quantity proportion can be set according to the practical application scene, for example, set as 8: 2.

3) and updating the weight parameters. Here, sample equalization may be performed on each obtained training set, for example, for a training set, the corresponding proportion condition is set as: the number ratio between the training user sample with the classification label of 1 and the training user sample with the classification label of 0 is 1: 1. if the matching of the quantity proportion in the training set and the proportion condition fails, performing extraction processing with replacement on the user sample set, and adding the extracted user samples to the training set until the quantity proportion in the training set is successfully matched with the proportion condition, wherein when the training user samples with the classification labels of 1 in the training set are insufficient, performing random extraction processing with replacement on a plurality of user samples with the classification labels of 1 in the user sample set; and when the training user sample with the classification label of 0 in the training set is insufficient, performing replaced random extraction processing on a plurality of user samples with the classification label of 0 in the user sample set.

In addition, principal component analysis processing can be carried out on the user features included in the training concentrated training user samples, and the obtained principal component features are used as new user features, so that the feature types in the training samples can be effectively reduced, and the convergence rate of the classification model can be improved conveniently during model training.

After sample equalization and principal component analysis processing are carried out on the N training sets, the classification model is trained according to the training sets so as to update the weight parameters of the classification model, wherein model training can be carried out by combining a back propagation mechanism and a gradient descent mechanism. For easy understanding, the updated weight parameter obtained after the classification model is trained according to the ith training set is denoted as M_iThe weight parameter before updating is M_i-1. In addition, the type of the classification model is not limited in the embodiments of the present application, and may be, for example, a Logistic Regression (LR) model.

4) And updating the dynamic threshold value. In the weight parameter updating step, for the ith training set, the classification threshold β is passed_i-1To obtain a predicted classification label, such as:

wherein, X represents the user characteristic, p represents the probability that the prediction classification label of the user corresponding to the user characteristic is 1, g (X) represents the prediction classification label, the classification threshold value corresponds to the above attrition rate classification threshold value, and the prediction classification label corresponds to the above prediction category. In the first iteration, the initial classification threshold β₀The setting can be made according to the actual application scenario, for example, the setting is 0.5.

While updating the weight parameters of the classification model through the ith training set, calculating the classification error rate corresponding to the training set

Wherein K represents the total number of training user samples in the ith training set, x_kRepresents the user characteristics, y, included in the kth training user sample in the ith training set_kAnd the classification label included in the kth training user sample in the ith training set is represented. Function I () is an indicative function when g (x)_k)≠y_kWhen true, I (g (x)_k)≠y_k) Has a value of 1; when g (x)_k)≠y_kIf not, I (g (x)_k)≠y_k) The value of (d) is 0.

Then, calculating the corresponding misclassification factor of the ith training set

The misclassification factor E_lAnd the updated sample weight corresponding to the ith training set is also used for carrying out replaced random extraction processing on the user sample set in the next iteration so as to obtain a new ith training set. According to the misclassification factor E_lClassification threshold β for classification model_i-1Is updated, i.e. beta_i＝β_i-1-E_l×β_i-1。

For the ith training set, the weight parameter M is saved_iClassification threshold beta_iAnd sample weight E_lAnd applying the weight parameter M to the ith test set_iAnd a classification threshold β_iThe classification model is tested to obtain an effect index R_i(corresponding to the foregoing attrition user classification effect index), the type of the effect index may be a recall ratio, an accuracy ratio, or an AUC value, which is not limited herein.

When the effect index R_iLess than or equal to the effect index threshold (corresponding to the above lost user index threshold), and there is a training set without model trainingContinuing to perform model training according to the training set which is not subjected to model training, and performing model testing according to the testing set which is not subjected to model testing; when the effect index R_iWhen the value is less than or equal to the effect index threshold value and no training set which is not subjected to model training exists, the iteration of the round is finished, and the next iteration is carried out according to the updated sample weight corresponding to each training set, namely, the steps 2) to 4) are carried out again; when the effect index R_iStopping iteration when the weight parameter is larger than the effect index threshold value, and enabling the weight parameter M to be_iAs an optimal weight parameter M_endClassifying the threshold value beta_iAs an optimal classification threshold β_end。

5) And (3) classification prediction: according to the application weight parameter M_endAnd a classification threshold β_endThe classification model of (2) processes the user characteristics of the target user in the T-th period to obtain the predicted classification label of the target user in the T + 1-th period, such as:

where p denotes a weight parameter M according to the application_endThe classification model is used for performing probability mapping processing on the user characteristics of the target user in the T-th period to obtain the probability that the predicted classification label of the target user in the T + 1-th period is 1.

In the pre-churn scenario, the digital marketing campaign of each period can relate to the change of the feature value and the change of the classification label caused by the scene data update, for example, in the marketing campaign of the T-1 period, the user sample comprises the user features in the T-2 period and the churn-retention classification label of the T-1 period; in the Tth cycle marketing campaign, the user sample includes user features for the T-1 th cycle, and the attrition-retention category labels for the Tth cycle. Therefore, the classification threshold is dynamically updated according to the change of the period, and the accuracy of classification processing in the pre-churn scene can be improved.

Under the same used classification model, the inventor performs experimental verification on the scheme of the fixed classification threshold provided by the related art and the scheme of dynamically updating the threshold provided by the embodiment of the present application, and obtains the effect indexes (here, the effect indexes in the actual application of the model) in the marketing campaign of the T-1 th cycle and the marketing campaign of the T-th cycle in the pre-churn scenario, and the experimental results are as follows:

the embodiment of the present application further provides a schematic diagram of training, testing and applying a classification model in a loss recovery scenario as shown in fig. 8, and it can be seen that in the loss recovery scenario, the training step of the classification model is similar to that in a pre-loss scenario, except that in step 1) in the loss recovery scenario, when constructing a user sample, the user sample is constructed by using the user characteristics of the user in the T-2 th period and the classification label of the T-1 th period, where the user is in an inactive state in the T-1 th period, and the classification label is a reflow-non-reflow classification label. The other difference is that in the step 5) of the loss recovery scene, the user characteristics of the target user in the T-1 th period are processed to obtain the predicted classification label of the target user in the T +1 th period,

in the digital marketing activity, the user characteristics and the classification labels used in different scenes are different, for example, in the marketing activity of the Tth period in the pre-churning scene, the user sample comprises the user characteristics in the T-1 th period and the churning-retention classification labels of the Tth period; and in the marketing campaign of the Tth period in the loss recovery scene, the user sample comprises the user characteristics in the T-2 th period and the reflow-no-reflow classification label of the Tth period, and the actual probability of the reflow of the user in the loss recovery scene is generally lower than that in the pre-loss scene, so that the fixed classification threshold cannot be simultaneously applied to different scenes. In the embodiment of the application, the classification threshold is dynamically updated according to the actual service scene, so that the classification threshold can be well adapted to the user characteristics and the classification labels used in the service scene.

Under the same used classification model, the inventor performs experimental verification on a scheme of a fixed classification threshold provided by the related art and a scheme of dynamically updating the threshold provided by the embodiment of the present application, and obtains an effect index (here, an effect index when the model is actually applied) of a marketing campaign in a certain period in a pre-churning scene and a churning recovery scene, where the experimental result is as follows:

through this application embodiment, can realize following technological effect: 1) the classification threshold is automatically updated in the model training process, and the classification error caused by manual experience is reduced; 2) when the classification threshold is updated, a gradient descent algorithm is used, so that the optimal classification threshold can be quickly converged; 3) the classification threshold value can be adjusted in a self-adaptive manner according to the actual user characteristics and the actual service scene, so that the applicability is improved; 4) the method has good expandability, can be integrated in various types of classification models, such as an LR model, an SVM model or a Random Forest (RF) model, and can be used in business scenes of various classification models, such as a pre-churn scene, a churn recovery scene or a payment scene.

Continuing with the exemplary structure in which artificial intelligence based attrition user prediction means 455 provided by embodiments of the present application is implemented as software modules, in some embodiments, as shown in fig. 3, the software modules stored in artificial intelligence based attrition user prediction means 455 of memory 450 may include: a training module 4551, configured to perform classification processing on whether a user sample participating in a service is a service loss user according to a loss parameter and a loss rate classification threshold of a user loss model, and update the loss parameter and the loss rate classification threshold of the user loss model according to a difference between an obtained prediction category and a labeled category of the user sample; an application module 4552, configured to perform classification processing on whether the target user is a potentially traffic-losing user according to the updated loss parameter and the loss rate classification threshold of the user loss model.

In some embodiments, artificial intelligence based attrition user prediction means 455 further comprises: the extraction module is used for extracting a plurality of user samples participating in the business to obtain a training set and a test set; the training set is used for updating the attrition parameters and the attrition rate classification threshold of the user attrition model.

In some embodiments, artificial intelligence based attrition user prediction means 455 further comprises: the testing module is used for carrying out classification processing on whether the testing user sample in the testing set is a loss user of a service or not through the updated loss parameters and the loss rate classification threshold of the user loss model, and determining a loss user classification effect index according to the difference between the obtained prediction class and the marking class of the testing user sample; and when the attrition user classification effect index is larger than the attrition user index threshold value, determining that the user attrition model is updated.

In some embodiments, the training module 4551 is further configured to: classifying whether the training user samples in the training set are service loss users or not, and updating loss parameters, loss rate classification threshold values and extraction proportions of a user loss model according to the difference between the obtained prediction classes and the labeled classes of the training user samples; the extraction proportion is used for extracting a plurality of user samples; artificial intelligence based attrition user prediction means 455 further comprises: and the iteration module is used for extracting and training a plurality of user samples again according to the updated extraction proportion when the loss user classification effect index is smaller than or equal to the loss user index threshold value until the obtained new loss user classification effect index is larger than the loss user index threshold value.

In some embodiments, the training module 4551 is further configured to: carrying out loss calculation processing on the difference between the prediction category and the mark category of each training user sample in a training set, carrying out backward propagation on a loss value obtained by the loss calculation processing in a user loss model, and updating the loss parameters of the user loss model in the backward propagation process; and determining the ratio of the number of training user samples with different prediction types and mark types in the training set to the total number of the training user samples in the training set as a classification error rate, and updating the loss rate classification threshold and the extraction proportion of the user loss model according to the classification error rate.

In some embodiments, the training module 4551 is further configured to: determining a division result between a result obtained by subtracting the classification error rate from 1 and the classification error rate; carrying out logarithm processing on the division result, and taking a first product result between the result of the logarithm processing and a set coefficient as an updated extraction proportion; and determining a second product result between the attrition rate classification threshold and the updated extraction proportion, and subtracting the second product result from the attrition rate classification threshold to obtain the updated attrition rate classification threshold.

In some embodiments, the extraction module is further configured to: taking the reciprocal of the set training set quantity as an extraction proportion, and performing replaced random extraction processing on a plurality of user samples to obtain a training set according with the set training set quantity; setting the number of training sets as an integer greater than 1; and for each training set, determining the number of the test user samples according to the number of the training user samples in the training set and the sample number proportion between the training set and the test set, and performing replaced random extraction processing on a plurality of user samples according to the number of the test user samples to obtain the test set corresponding to the training set.

In some embodiments, the training module 4551 is further configured to: classifying whether the training user samples in the target training set are lost users of the service or not; the target training set is any one of training sets which accord with the set training set number; the test module is further configured to: classifying whether the test user sample in the target test set is a service loss user or not through the updated loss parameters of the user loss model and the loss rate classification threshold; and the target test set and the target training set have a corresponding relation.

In some embodiments, artificial intelligence based attrition user prediction means 455 further comprises: the continuous classification module is used for classifying the training user samples in the training set which is not subjected to classification processing through the updated loss parameters and the loss rate classification threshold of the user loss model when the loss user classification effect index corresponding to the target test set is less than or equal to the loss user index threshold and the training set which is not subjected to classification processing exists, so as to obtain the loss user classification effect index corresponding to the test set which is not subjected to classification processing; and the test set which is not subjected to classification processing and the training set which is not subjected to classification processing have a corresponding relation.

In some embodiments, the iteration module is further to: and when the loss user classification effect indexes corresponding to all the test sets are smaller than or equal to the loss user index threshold value, performing replaced random extraction processing and training on the plurality of user samples again according to the updated extraction proportion.

In some embodiments, artificial intelligence based attrition user prediction means 455 further comprises: and the sample balancing module is used for extracting a plurality of user samples when the matching of the quantity proportion between the training user samples of different label types in the training set and the proportion condition fails, and adding the extracted user samples to the training set so as to ensure that the matching of the quantity proportion in the new training set and the proportion condition succeeds.

In some embodiments, the user sample includes user characteristics and tag categories for the tagged user; artificial intelligence based attrition user prediction means 455 further comprises: the principal component analysis module is used for carrying out principal component analysis processing on the user characteristics included in the training user samples in the training set and taking the obtained principal component characteristics as new user characteristics; wherein the principal component features include a feature type that is less than a feature type included in the user features.

In some embodiments, the number of the user churn models is multiple, and the model structures of different user churn models are different; the application module 4552 is further configured to: taking the user loss model with the largest loss user classification effect index in the plurality of updated user loss models as a target user loss model; and classifying whether the target user is a service potential lost user or not according to the loss parameters and the loss rate classification threshold of the target user loss model.

In some embodiments, the user sample includes user characteristics and tag categories for the tagged user; the categories corresponding to the user churn model comprise a churn category and a non-churn category; a training module 4551, further configured to: performing probability mapping processing on user characteristics included in a user sample participating in a service through loss parameters of a user loss model to obtain the probability that a marked user corresponding to the user sample belongs to a loss category; when the probability of belonging to the loss category is larger than the loss rate classification threshold of the user loss model, taking the loss category as the prediction category of the marked user corresponding to the user sample; and when the probability of belonging to the loss category is less than or equal to the loss rate classification threshold of the user loss model, taking the non-loss category as the prediction category of the marked user corresponding to the user sample.

In some embodiments, the user sample includes user characteristics of the tagged users belonging to the non-attrition category at the T-1 th period, and the tagged categories of the tagged users at the T-th period; the marker category is one of a non-attrition category and an attrition category; or the user sample comprises the user characteristics of the marked users belonging to the non-attrition category in the T-2 th period and the marked categories of the marked users in the T-2 th period; marking users belonging to the attrition category in the T-1 period; the marker class is one of a non-attrition class and an attrition class.

Embodiments of the present application provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to execute the artificial intelligence based attrition user prediction method described above in the embodiments of the present application.

Embodiments of the present application provide a computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform a method provided by embodiments of the present application, for example, an artificial intelligence based attrition user prediction method as illustrated in fig. 4A, 4B, 4C, 4D or 4E. Note that the computer includes various computing devices including a terminal device and a server.

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

In some embodiments, executable instructions may be written in any form of programming language (including compiled or interpreted languages), in the form of programs, software modules, scripts or code, 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.

By way of example, executable instructions may correspond, but do not necessarily have to correspond, to files in a file system, and may be stored in a portion of a file that holds 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 coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code).

By way of example, executable instructions may be deployed to be executed on one computing device or on multiple computing devices at one site or distributed across multiple sites and interconnected by a communication network.

In summary, the following technical effects can be achieved through the embodiments of the present application:

1) the loss rate classification threshold is updated in the model training process, so that the obtained loss rate classification threshold can be suitable for practice even if the characteristic value of the user characteristic is changed, the model version is changed or the service scene is changed, the accuracy of predicting the loss user can be improved, and the classification error caused by manual experience is reduced; meanwhile, synchronous updating of the weight parameters can be realized.

2) The method and the system can be suitable for various service scenes such as a pre-loss scene, a loss recovery scene and the like, can carry out targeted service operation according to the predicted potential loss users, and improve the viscosity of the users to the service.

3) The extracted training set can be subjected to sample equalization and principal component analysis, so that the quality of a training user sample can be improved, and the convergence speed of a user loss model is effectively improved; in each iteration, model training and model testing are respectively carried out through a plurality of training sets and the test set corresponding to each training set, and on the other hand, the speed of model convergence is improved.

4) Under the condition that the number of the user loss models is multiple, the optimal model structure can be screened out according to the loss user classification effect indexes, and the accuracy of predicting the loss users is further improved.

5) The method and the device have good expandability, and the scheme provided by the embodiment of the application can be integrated in various binary models to realize accurate prediction.

The above description is only an example of the present application, and is not intended to limit the scope of the present application. Any modification, equivalent replacement, and improvement made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. An artificial intelligence-based attrition user prediction method is characterized by comprising the following steps:

2. The attrition user prediction method of claim 1 further comprising:

extracting a plurality of user samples participating in the business to obtain a training set and a testing set;

the training set is used for updating the attrition parameters and the attrition rate classification threshold of the user attrition model;

before the classifying process of whether the target user is a potential lost user of the service according to the updated loss parameter and the updated loss rate classification threshold of the user loss model, the method further includes:

classifying whether the testing user sample in the testing set is the service loss user or not through the updated loss parameters and the loss rate classification threshold of the user loss model, and determining a loss user classification effect index according to the difference between the obtained prediction category and the marking category of the testing user sample;

and when the attrition user classification effect index is larger than the attrition user index threshold value, determining that the user attrition model is updated.

3. The attrition user prediction method of claim 2 wherein,

the classifying whether the user sample participating in the service is a lost user of the service is performed, and updating a loss parameter and a loss rate classification threshold of the user loss model according to a difference between an obtained prediction category and a mark category of the user sample comprises:

classifying whether the training user samples in the training set are loss users of the service or not, and updating loss parameters, loss rate classification thresholds and extraction ratios of the user loss model according to the difference between the obtained prediction categories and the marking categories of the training user samples;

wherein the extraction proportion is used for extracting the plurality of user samples;

the churn user prediction method further comprises:

and when the attrition user classification effect index is smaller than or equal to the attrition user index threshold value, extracting and training the plurality of user samples again according to the updated extraction proportion until the obtained new attrition user classification effect index is larger than the attrition user index threshold value.

4. The attrition user prediction method of claim 3 wherein the updating of the attrition parameters, the attrition rate classification threshold, and the extraction ratio of the user attrition model based on the difference between the obtained prediction category and the labeled category of the training user sample comprises:

for each training user sample in the training set, performing loss calculation processing on the difference between the prediction category and the mark category of the training user sample, performing back propagation in the user loss model according to a loss value obtained by the loss calculation processing, and performing back propagation on the loss value

Updating the loss parameters of the user loss model in the process of back propagation;

determining a ratio between the number of training user samples in the training set with different prediction classes and label classes and the total number of training user samples in the training set as a classification error rate, and

and updating the attrition rate classification threshold value and the extraction proportion of the user attrition model according to the classification error rate.

5. The attrition user prediction method of claim 4 wherein updating the attrition rate classification threshold and the extraction ratio of the user attrition model based on the classification error rate comprises:

determining a division result between a result of subtracting the classification error rate from 1 and the classification error rate;

carrying out logarithm processing on the division result, and taking a first product result between a result of the logarithm processing and a set coefficient as the updated extraction proportion;

and determining a second product result between the attrition rate classification threshold and the updated extraction proportion, and subtracting the second product result from the attrition rate classification threshold to obtain the updated attrition rate classification threshold.

6. The attrition user prediction method of claim 3 wherein the extracting a plurality of user samples participating in a business to obtain a training set and a test set comprises:

taking the reciprocal of the number of the set training sets as an extraction proportion, and performing replaced random extraction processing on the plurality of user samples to obtain the training sets according with the number of the set training sets; wherein the number of the set training sets is an integer greater than 1;

for each training set, determining the number of samples of the test user according to the number of samples of the training set including the training user and the number proportion of the samples between the training set and the test set, and

and performing replaced random extraction processing on the plurality of user samples according to the number of the test user samples to obtain a test set corresponding to the training set.

7. The attrition user prediction method of claim 6 wherein,

the classifying whether the training user sample in the training set is the lost user of the service or not includes:

classifying whether the training user sample in the target training set is the lost user of the service;

the target training set is any one of training sets which accord with the set training set quantity;

the classifying processing whether the testing user sample in the testing set is the lost user of the service or not is performed through the updated loss parameter and the loss rate classification threshold of the user loss model, and the classifying processing includes:

classifying whether the test user sample in the target test set is the lost user of the service or not through the updated loss parameters and the loss rate classification threshold of the user loss model;

and the target test set and the target training set have a corresponding relation.

8. The attrition user prediction method of claim 7 further comprising:

when the loss user classification effect index corresponding to the target test set is smaller than or equal to the loss user index threshold and a training set which is not subjected to classification processing exists, classifying the training user samples in the training set which is not subjected to classification processing through the updated loss parameters and loss rate classification threshold of the user loss model so as to classify the training user samples to obtain the target test set

Obtaining a loss user classification effect index corresponding to a test set which is not subjected to classification processing;

the test set which is not subjected to classification processing and the training set which is not subjected to classification processing have a corresponding relation;

when the attrition user classification effect index is smaller than or equal to the attrition user index threshold value, performing extraction processing and training again on the plurality of user samples according to the updated extraction proportion, including:

and when the attrition user classification effect indexes corresponding to all the test sets are smaller than or equal to the attrition user index threshold value, performing replaced random extraction processing and training on the plurality of user samples again according to the updated extraction proportion.

9. The attrition user prediction method of claim 2, wherein after the extracting the plurality of user samples participating in the service to obtain the training set and the test set, further comprising:

when the matching of the quantity proportion and the proportion condition between the training user samples of different label types in the training set fails, the plurality of user samples are extracted and processed, and

and adding the extracted user sample to the training set so as to enable the number proportion in the new training set to be successfully matched with the proportion condition.

10. The attrition user prediction method of claim 2 wherein,

the user sample comprises user characteristics and mark categories of mark users;

after the plurality of user samples participating in the service are extracted and processed to obtain the training set and the test set, the method further includes:

performing principal component analysis processing on user characteristics included in the training user samples in the training set, and taking the obtained principal component characteristics as new user characteristics;

wherein the principal component feature comprises a feature type smaller than a feature type comprised by the user feature.

11. The attrition user prediction method of claim 1 wherein,

the user sample comprises user characteristics and mark categories of mark users; the categories corresponding to the user churn model comprise a churn category and a non-churn category;

the classifying processing whether the user sample participating in the service is the lost user of the service or not through the loss parameter and the loss rate classification threshold of the user loss model comprises the following steps:

performing probability mapping processing on user characteristics included in a user sample participating in a service through loss parameters of the user loss model to obtain the probability that a marked user corresponding to the user sample belongs to the loss category;

when the probability of belonging to the loss category is larger than the loss rate classification threshold of the user loss model, taking the loss category as the prediction category of the marked user corresponding to the user sample;

and when the probability of belonging to the loss category is smaller than or equal to the loss rate classification threshold of the user loss model, taking the non-loss category as the prediction category of the marked user corresponding to the user sample.

12. The attrition user prediction method of any one of claims 1 to 11 wherein,

the user sample comprises user characteristics of the marked users belonging to the non-attrition category in the T-1 th period and the marked categories of the marked users in the T-1 th period; the marker class is one of the non-attrition class and the attrition class; alternatively, the first and second electrodes may be,

the user sample comprises user characteristics of the marked users belonging to the non-attrition category in the T-2 th period and the marked categories of the marked users in the T-2 th period; the marked users belong to the attrition category in the T-1 th period; the marker class is one of the non-attrition class and the attrition class.

13. An attrition user prediction device based on artificial intelligence, comprising:

14. An electronic device, comprising:

a memory for storing executable instructions;

a processor configured to implement the artificial intelligence based attrition user prediction method of any one of claims 1 to 12 when executing executable instructions stored in the memory.

15. A computer-readable storage medium having stored thereon executable instructions for, when executed by a processor, implementing the artificial intelligence based attrition user prediction method of any one of claims 1 to 12.