A method, device and readable storage medium for identifying high-risk loss users

Abstract

The application provides a high-risk loss user identification method, device and readable storage medium, the method comprising: constructing graph node attributes according to user service data, and dividing the graph node attributes into K categories; constructing the edges of the graph according to the call relationship between users; for the K categories, obtaining K graphs according to the graph node attributes and the edges; inputting the K graphs into K graph neural network algorithms respectively to obtain the vector representation of the user in each category; calculating the weight corresponding to the vector representation of each category through an attention mechanism according to the vector representation of the user in each category, and classifying through a normalization exponential function softmax function to obtain a high-risk loss user list. The method, device and readable storage medium can solve the problem of low recognition accuracy of the existing high-risk loss user identification method due to the inability to fully mine the relationship between users.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer intelligent learning, and in particular to a high-risk loss user identification method and device and readable storage medium. BACKGROUND

[0002] At present, the competition of communication operators is increasingly fierce, and the business development has entered the stock operation stage. In the case of very limited stock users in the communication market, the communication operator business is facing great market competition pressure, and it is more and more difficult for the operator to develop new users, and it is more effective to save enterprise cost expenditure to maintain high-value old customers than to develop new users. On the basis of rich data of the operator, the application of artificial intelligence algorithm can predict and identify high-risk loss users in advance, so that the operator can develop corresponding strategies to retain users.

[0003] However, the existing operator mainly uses the method of machine learning to identify high-risk loss users, wherein the features used for identification are mainly screened by artificial means, which is too dependent on artificial experience, and the relationship between users cannot be fully mined, so the existing method has the problem of low identification accuracy. SUMMARY

[0004] The technical problem to be solved by the present application is to solve the above-mentioned problems of the prior art, and to provide a high-risk loss user identification method, device and readable storage medium, which can solve the problem of low identification accuracy of the existing high-risk loss user identification method due to the inability to fully mine the relationship between users.

[0005] In a first aspect, the present application provides a high-risk loss user identification method, which comprises the following steps:

[0006] The method comprises the following steps:

[0007] Constructing a graph node attribute according to user business data, and dividing the graph node attribute into K categories;

[0008] Constructing the edge of the graph according to the call relationship between users;

[0009] For K categories, obtaining K graphs according to the graph node attribute and the edge;

[0010] Inputting the K graphs into K graph neural network algorithms respectively to obtain the vector representation of the user in each category;

[0011] According to the vector representation of the user in each category, the weight corresponding to the vector representation of each category is calculated through the attention mechanism, and the classification is performed through the normalization exponential function softmax function to obtain a high-risk loss user list.

[0012] Further, the method further comprises:

[0013] Obtaining user service data in a preset time period;

[0014] Taking the mobile phone numbers of each user in the user service data as nodes of a graph, and taking the service data corresponding to each mobile phone number as graph node attributes;

[0015] Dividing the graph node attributes into K categories according to different service conditions.

[0016] Further, the user service data includes user basic attribute information, billing data, voice, traffic, short message usage, mobile terminal condition and user product information.

[0017] Further, the categories include at least one of the following: user basic information category, user attribute category, user voice category, user traffic category, user short message category, user billing category, user stored value category, user terminal category, user product category, and user network state category.

[0018] Further, before the step of constructing edges of the graph according to the call relationship between users, the method further comprises:

[0019] Obtaining user call details in a preset time period;

[0020] Parsing the user call details to obtain the call relationship between users in the preset time period.

[0021] Further, the graph neural network algorithm is a GraphSAGE algorithm.

[0022] Further, after the step of obtaining the high-risk loss user list according to the vector representation of each category of the user through the attention mechanism, the method further comprises:

[0023] Sorting the weight corresponding to the vector representation of each category to obtain a target category corresponding to the highest weight;

[0024] Formulating a corresponding retention strategy according to the business corresponding to the target category.

[0025] In a second aspect, the present application provides a device for identifying high-risk loss users, comprising:

[0026] An attribute division module configured to construct graph node attributes according to user service data, and divide the graph node attributes into K categories;

[0027] An edge construction module, connected to the attribute partitioning module, is used to construct the edges of a graph based on the call relationships between users;

[0028] The graph acquisition module, connected to the edge construction module, is used to obtain K graphs for K categories based on the graph node attributes and edges.

[0029] The vector representation acquisition module, connected to the graph acquisition module, is used to input the K graphs into the K graph neural network algorithm respectively to obtain the user's vector representation for each category;

[0030] The classification and identification module, connected to the vector representation acquisition module, is used to calculate the weight of the vector representation of each category based on the user's vector representation in each category through an attention mechanism, and to classify the data through a normalized exponential function and a softmax function to obtain a list of high-risk churned users.

[0031] Thirdly, the present invention provides a device for identifying high-risk churned users, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the high-risk churned user identification method described in the first aspect above.

[0032] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method for identifying high-risk churned users described in the first aspect.

[0033] This invention provides a method, apparatus, and readable storage medium for identifying high-risk churned users. First, a graph node attribute is constructed based on user business data, and this attribute is divided into K categories. Then, edges are constructed based on the call relationships between users. Next, for each of the K categories, K graphs are obtained based on the graph node attributes and edges. Finally, the K graphs are input into K graph neural network algorithms to obtain the vector representation of each user in each category. Then, based on the vector representation of each user in each category, the weight corresponding to the vector representation of each category is calculated using an attention mechanism, and classification is performed using a normalized exponential function (softmax) to obtain a list of high-risk churned users. This invention utilizes call relationships between users and user business data to construct a graph, and fully considers the influence relationships between users through graph neural network algorithms. Compared to methods that directly use business data for machine learning modeling, this further enhances data features and enables more accurate identification of high-risk churned users. Furthermore, by classifying users and introducing an attention mechanism, the key factors influencing user churn can be more accurately determined, facilitating the development of more reasonable retention strategies. This solves the problem of low accuracy in existing high-risk churned user identification methods due to their inability to fully explore the relationships between users. Attached Figure Description

[0034] Figure 1 A flow chart of a high-risk churn user identification method of Embodiment 1 of the present application;

[0035] Figure 2 A flow chart of another high-risk churn user identification method of Embodiment 1 of the present application;

[0036] Figure 3 A structural schematic diagram of a high-risk churn user identification device of Embodiment 2 of the present application;

[0037] Figure 4 A structural schematic diagram of a high-risk churn user identification device of Embodiment 3 of the present application. DETAILED DESCRIPTION

[0038] In order to make the technical solution of the present application better understood by those skilled in the art, the embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0039] It can be understood that the specific embodiments and the accompanying drawings described herein are only used to explain the present application, but not to limit the present application.

[0040] It can be understood that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0041] It can be understood that, for the convenience of description, only the parts related to the present application are shown in the drawings of the present application, and the parts unrelated to the present application are not shown in the drawings.

[0042] It can be understood that each unit and module involved in the embodiments of the present application can only correspond to one entity structure, or can be composed of multiple entity structures, or multiple units and modules can also be integrated into one entity structure.

[0043] It can be understood that the terms "first", "second" and the like in the embodiments of the present application are used to distinguish different objects, or to distinguish different treatments of the same object, but not to describe a specific order of the object.

[0044] It can be understood that the functions and steps marked in the flow chart and block diagram of the present application can occur in an order different from that marked in the drawings without conflict.

[0045] It can be understood that the flowcharts and block diagrams of the present application show the possible implementation architecture, function and operation of the system, device, equipment, method according to the embodiments of the present application. Each block in the flowchart or block diagram can represent a unit, module, program segment, code, which contains executable instructions for implementing the specified function. Moreover, each block or combination of blocks in the block diagram and flowchart can be implemented by a hardware-based system for implementing the specified function, or by a combination of hardware and computer instructions.

[0046] It can be understood that the units and modules involved in the embodiments of the present application can be implemented by software or by hardware, for example, the units and modules can be located in a processor.

[0047] Embodiment 1:

[0048] The embodiment provides a method for identifying high-risk loss users, as shown in the method, the method comprises the following steps of: Figure 1

[0049] Step S101: constructing graph node attributes according to user service data, and dividing the graph node attributes into K categories.

[0050] In the embodiment, the user service data includes user basic attribute information, billing data, voice, traffic, short message usage, mobile terminal condition and user product information, and K is preferably greater than or equal to 3.

[0051] Optionally, the step of constructing graph node attributes according to user service data and dividing the graph node attributes into K categories comprises the following steps of:

[0052] acquiring user service data in a preset time period;

[0053] taking the mobile phone numbers of each user in the user service data as nodes of a graph, and taking the service data corresponding to each mobile phone number as graph node attributes;

[0054] dividing the graph node attributes into K categories according to different service conditions.

[0055] ​In the embodiment, a mobile phone number of an operator user is used as a unique identifier of each user, the unique identifier is used as a node V of a graph, and user service data in a time period T is used as a graph node attribute. The graph node attribute is divided into K categories according to different service conditions, which can include but are not limited to a user basic information category, a user voice category, a user traffic category, a user short message category, a user bill category, a user terminal category, and a user product category. Each category includes corresponding service data fields, for example, the user basic information category includes user gender and age, the user voice category includes user call duration and call times, and the user traffic category includes online traffic, online duration, and online times. Therefore, the graph node can be represented as Specifically, the nth user has the kth attribute, where n∈{1, 2, 3,...N} represents N different operator users, and k∈{1, 2, 3,...K} represents K attributes of the node n.

[0056] Optionally, the categories include at least one of the following: a user basic information category, a user attribute category, a user voice category, a user traffic category, a user short message category, a user bill category, a user stored value category, a user terminal category, a user product category, and a user network status category.

[0057] In the embodiment, the graph node attribute is preferably divided into 10 categories according to different service conditions, and the categories and corresponding fields are shown in Table 1, including a user basic information category, a user attribute category, a user voice category, a user traffic category, a user short message category, a user bill category, a user stored value category, a user terminal category, a user product category, and a user network status category.

[0058] Table 1: Categories and corresponding fields

[0059] Category Field User basic information class Name, gender, age, zodiac... User attribute class Service type, online time, user level... User voice class Call duration, call times, calling duration... User traffic class Online traffic, online duration, online times... User SMS class SMS number, SMS service days... User billing class Basic monthly rent, outgoing account amount, call fee, traffic fee... User stored value class Stored value amount, stored value times... User terminal class Whether customized, terminal type, terminal hardware condition... User product class App access situation data User network status class Network status data

[0060] Step S102: constructing edges of the graph according to call relationships between users.

[0061] In the embodiment, in order to facilitate mining of relationships between users, edges of the graph are constructed according to call relationships between users.

[0062] Optionally, before the edges of the graph are constructed according to the call relationships between users, the method further includes:

[0063] obtaining user call details in a preset time period;

[0064] parsing the user call details to obtain the call relationships between users in the preset time period.

[0065] In the embodiment, an operator obtains user call details, parses user call relationships, and obtains call relationships between users in a time period T to establish edges E of the graph.

[0066] Step S103: obtaining K graphs according to the graph node attributes and edges for K categories.

[0067] In this embodiment, K graphs (G1, G2,..., G K ) are obtained according to the graph node attributes and edges for K categories. k∈{1,2,3…K}, n∈{1,2,3,…N}.

[0068] Step S104: inputting the K graphs into K graph neural network algorithms respectively to obtain vector representations of the user in each category.

[0069] In this embodiment, the K graphs (G1, G2,..., G K ) are inputted into K graph neural network algorithms respectively to obtain vector representations g=(g1, g2,..., g K ) of the user, k∈{1,2,3…K}.

[0070] Optionally, the graph neural network algorithm is a GraphSAGE (Graph SAmple and aggreGafE) algorithm.

[0071] In this embodiment, GraphSAGE is an iterative algorithm for learning the embedding of each node in a certain graph, and the embedding reflects the transformation of the features of the node. The goal of GraphSAGE is to learn the representation of each node based on the combination of its adjacent nodes to better complete information aggregation.

[0072] Step S105: calculating the weight corresponding to the vector representation of each category through an attention mechanism according to the vector representation of the user in each category, and classifying through a normalization exponential function softmax function to obtain a high-risk loss user list.

[0073] In this embodiment, the obtained vector representations g=(g1, g2,..., g K ), k∈{1,2,3…K} are calculated through an attention mechanism (attention) to obtain the weight α=(α1, α2,..., α K ) corresponding to the vector representation of each category, and then classified through a softmax function to obtain a high-risk loss user list.

[0074] It should be noted that before the above steps, the attention mechanism network model can be trained according to the churned user data counted by the operator. Specifically, according to the churned user data counted by the operator, 100,000 users are labeled, and the label is whether the user is a churned user in the next 3 months, to obtain a training data set. The node attribute of the 100,000 users is set to 6 months of data from January to June, and the user label counts whether the user is a churned user from July to September. Using the training data set, the vector representation g = (g1, g2,..., g of the user is obtained according to the steps S101-S104 10 ), and the weight α = (α1, α2,..., α of the vector representation of each category corresponding to the attention mechanism (attention) learning during the training process 10 ), classification is performed through softmax, a loss function and a back propagation algorithm are set, and the training of the model is completed. Using the trained model, the churned user list for the next 3 months can be predicted using any 6 months of data, and a high-risk churned user list is output, and a corresponding retention strategy is formulated according to the weight α.

[0075] Optionally, after the vector representation of each category is calculated by the attention mechanism according to the vector representation of the user in each category, the weight of the vector representation of each category is calculated, and the high-risk churned user list is obtained by the normalization exponential function softmax function classification, the method further comprises:

[0076] sorting the weight of the vector representation of each category to obtain the target category corresponding to the highest weight;

[0077] formulating a corresponding retention strategy according to the business corresponding to the target category.

[0078] In the embodiment, the weight (α1, α2,..., α of the vector representation of each category K ) is sorted to obtain the target category k corresponding to the highest weight α k , and a corresponding retention strategy is formulated according to the business corresponding to the target category k.

[0079] The application can fully mine the business characteristics of user contact relationship aggregation user contacts, and provides a more accurate and efficient method for high-risk churned user identification, which can improve the accuracy of high-risk churned user identification and analyze the key factors affecting user churn, thereby helping operators to formulate retention strategies.

[0080] In a specific embodiment, as shown in Figure 2 , the high-risk churned user identification method can include the following steps:

[0081] (1) Constructing a graph node attribute of user business data.

[0082] Specifically, the mobile phone number of the operator user is used as the unique identifier of each user, the operator user is taken as the node V of the graph, and the service data of the user in a period of time T is taken as the attribute of the node of the graph.

[0083] In the embodiment, the number of nodes N is 100000, representing 100000 users; and the period of time T can be 6 months or 3 months. The service data of the user includes user basic information, voice traffic and short message usage, bill and the like.

[0084] (2) Division of the attribute of the node of the graph.

[0085] Specifically, the attribute of the node of the graph is divided into K categories according to different service conditions, which can include but are not limited to the user basic information category, the user voice category, the user traffic category, the user short message category, the user bill category, the user terminal category, the user product category and the like, and preferably includes the user basic information category, the user attribute category, the user voice category, the user traffic category, the user short message category, the user bill category, the user value storage category, the user terminal category, the user product category and the user network state category, and each category contains corresponding service data fields, such as the user basic information category containing the user gender, age and the like; the user voice category including the call duration and call times of the user; and the user traffic category including the online traffic, online duration and online times.

[0086] Therefore, the node of the graph can be represented as The specific meaning is the kth attribute of the nth user, wherein n∈{1, 2, 3,...N} represents N different operator users, and k∈{1, 2, 3,...K} represents the K attributes of the node n.

[0087] (3) Construction of the edge of the graph of the user call relationship.

[0088] Specifically, the operator obtains the user call details, analyzes the user call relationship, and obtains the call relationship between the users in a period of time T to establish the edge E of the graph.

[0089] In the embodiment, the edge E of the graph is represented by an N*N-dimensional adjacency matrix A, the value of N is 100000, and in a period of 6 months, if there is a call between two users, the corresponding value of the adjacency matrix A is set to 1, and if there is no call relationship, the corresponding value of the adjacency matrix A is set to 0.

[0090] (4) Extraction of the user vector representation by the graph neural network algorithm.

[0091] Specifically, according to the node attribute and the edge obtained by the above method, the graph can be represented as G=(G1, G2,...,G K ), representing that one graph Gk is obtained for each attribute category k e {1, 2, 3…K}, n e {1, 2, 3…N}. K attribute categories are obtained K graphs (G1, G2,..., G K ) are input into K graph neural network algorithms Graphsage, respectively, to obtain the vector representation g = (g1, g2,..., g K ), k e {1, 2, 3…K}.

[0092] In this embodiment, K = 10, so as to obtain 10 graphs (G1, G2,..., G 10 ) corresponding to 10 categories, and the 10 graphs of the 10 categories are input into 10 graph neural network algorithms Graphsage to obtain the vector representation (g2, g2,..., g 10 ) of the user in each service category, wherein Graphsage adopts a two-layer structure to aggregate two-order neighbor features.

[0093] (5) High-risk churn user identification.

[0094] Specifically, the obtained vector representation g = (g1, g2,..., g K ), k e {1, 2, 3…K} is subjected to attention mechanism to calculate the weight α = (α1, α2,..., α K ) corresponding to each category of vector representation, and then the weight is classified through a softmax function to obtain a high-risk churn user list. The weight (α1, α2,..., α K ) corresponding to each category of vector representation is sorted to obtain the highest weight α k ) corresponding to the feature category k, and then a corresponding retention strategy is formulated according to the service corresponding to the category k.

[0095] The embodiment utilizes the call relationship between users and the service data of the users to build a graph, and identifies high-risk churn users through a graph neural network algorithm, thereby solving the problem of low recognition accuracy. In addition, by classifying the node attributes and giving the weight of the influence of different category service data on user churn through an attention mechanism, the influence factors of user churn can be more accurately given.

[0096] The high-risk loss user identification method provided by the embodiment of the application first constructs graph node attributes according to user service data, and divides the graph node attributes into K categories; then constructs edges of the graph according to call relationships between users; then for the K categories, obtains K graphs according to the graph node attributes and the edges; finally, inputs the K graphs into K graph neural network algorithms respectively to obtain vector representations of users in each category; then calculates weights corresponding to the vector representations of each category through an attention mechanism according to the vector representations of users in each category, and classifies through a normalization exponential function softmax function to obtain a high-risk loss user list. The application constructs a graph by using call relationships between users and user service data, and fully considers the influence relationship between users through a graph neural network algorithm, thereby further enhancing data features compared with a method of directly modeling machine learning by using service data, and more accurate identification of high-risk loss users can be realized. In addition, through the division of categories and the introduction of the attention mechanism, key factors affecting user loss can be more accurately judged to facilitate the development of more reasonable retention strategies. The existing high-risk loss user identification method cannot fully mine the relationship between users, and has the problem of low recognition accuracy.

[0097] Embodiment 2

[0098] As shown in Figure 3 The embodiment provides a high-risk loss user identification device for executing the high-risk loss user identification method, and the device comprises:

[0099] An attribute division module 11 is configured to construct graph node attributes according to user service data, and divide the graph node attributes into K categories.

[0100] An edge construction module 12 is connected with the attribute division module 11 and is configured to construct edges of the graph according to call relationships between users.

[0101] A graph acquisition module 13 is connected with the edge construction module 12 and is configured to obtain K graphs according to the graph node attributes and the edges for the K categories.

[0102] A vector representation acquisition module 14 is connected with the graph acquisition module 13 and is configured to input the K graphs into K graph neural network algorithms respectively to obtain vector representations of users in each category.

[0103] A classification identification module 15 is connected with the vector representation acquisition module 14 and is configured to calculate weights corresponding to the vector representations of each category through an attention mechanism according to the vector representations of users in each category, and classify through a normalization exponential function softmax function to obtain a high-risk loss user list.

[0104] Optionally, the attribute division module 11 comprises:

[0105] a service data obtaining unit, configured to obtain user service data in a preset time period;

[0106] a node attribute extraction unit, configured to take a mobile phone number of each user in the user service data as a node of a graph, and take service data corresponding to each mobile phone number as a graph node attribute;

[0107] a category division unit, configured to divide the graph node attribute into K categories according to different service conditions.

[0108] Optionally, the user service data includes user basic attribute information, bill data, voice, traffic, short message use condition, mobile phone terminal condition and user product information.

[0109] Optionally, the categories include at least one of the following: user basic information category, user attribute category, user voice category, user traffic category, user short message category, user bill category, user stored value category, user terminal category, user product category and user network state category.

[0110] Optionally, the device further includes:

[0111] a call detail obtaining module, configured to obtain user call details in a preset time period;

[0112] a call relationship obtaining module, configured to parse the user call details to obtain the call relationship between users in the preset time period.

[0113] Optionally, the graph neural network algorithm is a GraphSAGE algorithm.

[0114] Optionally, the device further includes:

[0115] a weight ordering module, configured to order the corresponding weight of the vector representation of each category to obtain a target category corresponding to the highest weight;

[0116] a retention strategy making module, configured to make a corresponding retention strategy according to the service corresponding to the target category.

[0117] Embodiment 3:

[0118] Reference Figure 4 The embodiment provides a high-risk loss user identification device, including a memory 21 and a processor 22, the memory 21 stores a computer program, and the processor 22 is configured to run the computer program to execute the high-risk loss user identification method in embodiment 1.

[0119] The memory 21 is connected with the processor 22, the memory 21 can adopt a flash memory or a read-only memory or other memories, and the processor 22 can adopt a central processing unit or a single-chip microcomputer.

[0120] Embodiment 4:

[0121] The embodiment provides a computer readable storage medium, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the method for identifying a high-risk loss user in the above embodiment 1.

[0122] The computer readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information such as computer readable instructions, data structures, computer program modules or other data. The computer readable storage medium includes but is not limited to RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable read only memory), flash memory or other memory technology, CD-ROM (Compact Disc Read-Only Memory), digital versatile disc (DVD) or other optical disc storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device, or any other medium that can be used to store desired information and can be accessed by a computer.

[0123] In summary, the method for identifying a high-risk loss user, the device and the readable storage medium provided by the embodiment of the present application first construct graph node attributes according to user service data, and divide the graph node attributes into K categories; then construct the edges of the graph according to the call relationship between users; then for the K categories, obtain K graphs according to the graph node attributes and the edges; finally, input the K graphs into K graph neural network algorithms respectively to obtain the vector representation of the user in each category; then calculate the weight corresponding to the vector representation of each category through the attention mechanism according to the vector representation of the user in each category, and classify through the normalization exponential function softmax function to obtain a high-risk loss user list. The present application constructs a graph by using the call relationship between users and user service data, and fully considers the influence relationship between users through the graph neural network algorithm, which further enhances the data features compared with the method of directly modeling machine learning using service data, and can realize more accurate identification of high-risk loss users. In addition, through the division of categories and the introduction of the attention mechanism, the key factors affecting user loss can be more accurately judged to facilitate the development of more reasonable retention strategies. The existing method for identifying high-risk loss users has the problem of low recognition accuracy due to the inability to fully mine the relationship between users.

[0124] It is understood that the above embodiments are only exemplary for illustrating the principles of the present application, and the present application is not limited thereto. Various modifications and improvements can be made by those skilled in the art without departing from the spirit and scope of the present application, and these modifications and improvements are also considered as the protection scope of the present application.

Claims

1. A method for identifying high-risk churned users, characterized in that, The method includes: Graph node attributes are constructed based on user business data, and the graph node attributes are divided into K categories; Construct the edges of the graph based on the call relationships between users; For each of the K categories, K graphs are obtained based on the graph node attributes and edges; The K graphs are input into a K-graph neural network algorithm to obtain the vector representation of the user in each category; Based on the user's vector representation in each category, the weight corresponding to the vector representation of each category is calculated through an attention mechanism, and then classified using a normalized exponential function and a softmax function to obtain a list of high-risk churned users.

2. The method according to claim 1, characterized in that, The step of constructing graph node attributes based on user business data and dividing the graph node attributes into K categories specifically includes: Retrieve user business data within a preset time period; The mobile phone numbers of each user in the user business data are used as nodes of the graph, and the business data corresponding to each mobile phone number is used as the graph node attributes. The graph node attributes are divided into K categories according to different business scenarios.

3. The method according to claim 2, characterized in that, The user business data includes basic user attribute information, billing data, voice, data traffic, SMS usage, mobile terminal information, and user product information.

4. The method according to claim 2, characterized in that, The categories include at least one of the following: basic user information, user attributes, user voice, user data usage, user SMS, user bills, user stored value, user terminal, user product, and user network status.

5. The method according to claim 1, characterized in that, Before constructing the edges of the graph based on the call relationships between users, the method further includes: Retrieve detailed call records for users within a preset time period; The user call details are analyzed to obtain the call relationships between users within a preset time period.

6. The method according to claim 1, characterized in that, The graph neural network algorithm is the GraphSAGE algorithm, which is a graph sampling aggregation network.

7. The method according to claim 6, characterized in that, After calculating the weights corresponding to the vector representations of each category based on the user's vector representations in each category using an attention mechanism, and classifying them using a normalized exponential function and a softmax function to obtain a list of high-risk churned users, the method further includes: The weights corresponding to the vector representations of each category are sorted to obtain the target category corresponding to the highest weight; Develop corresponding retention strategies based on the business categories mentioned above.

8. A device for identifying high-risk churned users, characterized in that, include: The attribute partitioning module is used to construct graph node attributes based on user business data and divide the graph node attributes into K categories. An edge construction module, connected to the attribute partitioning module, is used to construct the edges of a graph based on the call relationships between users; The graph acquisition module, connected to the edge construction module, is used to obtain K graphs for K categories based on the graph node attributes and edges. The vector representation acquisition module, connected to the graph acquisition module, is used to input the K graphs into the K graph neural network algorithm respectively to obtain the user's vector representation for each category; The classification and identification module, connected to the vector representation acquisition module, is used to calculate the weight of the vector representation of each category based on the user's vector representation in each category through an attention mechanism, and to classify the data through a normalized exponential function and a softmax function to obtain a list of high-risk churned users.

9. A device for identifying high-risk churned users, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the method for identifying high-risk churned users as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for identifying high-risk churned users as described in any one of claims 1-7.

Images