Complaint refund prediction method and device and computer readable storage medium

By acquiring and integrating the historical and future behavioral characteristics of users, using the complaint refund prediction model to predict the amount of refunds for users, the problem of the inability to predict whether the complaints can generate refunds is solved, and the refund users are realized before the complaint occurs intervention to reduce losses and improve user satisfaction.

CN119027131APending Publication Date: 2024-11-26CHINA UNITED NETWORK COMM GRP CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410994060.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing complaint prediction methods can only predict whether users complain, and cannot predict whether user complaints will result in refunds, resulting in the inability to curb the occurrence of complaints and refunds at the source.

Method used

By obtaining the user's historical behavior data, extracting historical behavior characteristics and future behavior characteristics, and fusing them into the complaint refund prediction model to predict the amount of user complaint refund, so that manual intervention is carried out before the complaint occurs.

Benefits of technology

It realizes an accurate prediction of whether a refund occurs for user complaints, reduces company losses, and improves user perception and operational efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027131A_ABST
    Figure CN119027131A_ABST
Patent Text Reader

Abstract

The invention provides a complaint refund prediction method and device and a computer readable storage medium. The method comprises the steps of obtaining historical behavior data of a user; obtaining historical behavior characteristics and future behavior characteristics of the user according to the historical behavior data of the user; fusing the historical behavior characteristics and the future behavior characteristics of the user to obtain data characteristics of the user; and inputting the data features of the user into a preset complaint and refund prediction model, and predicting the complaint and refund amount of the user through the complaint and refund prediction model to obtain a prediction result. According to the method, the device and the computer readable storage medium, the problem that an existing complaint prediction method can only predict whether the user complains or not and cannot predict whether the user complains or not to generate refund or not can be solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a complaint refund prediction method, device and computer-readable storage medium. Background Art

[0002] With the development of network technology, there are more and more mobile Internet users and broadband users, and users are pursuing higher and higher network quality. The amount of complaints and refunds caused by various faults is getting larger and larger. At present, the way the customer service industry responds to complaints is often based on a single dimension and is a post-event remedial treatment. As a service provider, operators can collect data on the poor quality of users' daily Internet calls. How to make full use of this poor quality data, predict users who may complain before user complaints lead to refunds, curb the occurrence of complaints and refunds from the source, and prevent problems before they happen, is particularly important.

[0003] However, existing complaint prediction methods can only predict whether a user will file a complaint, but cannot predict whether a user complaint will result in a refund. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a complaint refund prediction method, device and computer-readable storage medium in response to the above-mentioned shortcomings of the prior art, so as to solve the problem that the existing complaint prediction method can only predict whether a user will file a complaint, but cannot predict whether a refund will be generated by the user's complaint.

[0005] In a first aspect, the present invention provides a complaint refund prediction method, comprising:

[0006] Obtain user's historical behavior data;

[0007] Obtaining historical behavior characteristics and future behavior characteristics of the user according to the historical behavior data of the user;

[0008] The historical behavior characteristics and future behavior characteristics of the user are integrated to obtain the data characteristics of the user;

[0009] The data features of the user are input into a preset complaint refund prediction model, and the complaint refund amount of the user is predicted by the complaint refund prediction model to obtain a prediction result.

[0010] Furthermore, obtaining the historical behavior characteristics and future behavior characteristics of the user based on the historical behavior data of the user specifically includes:

[0011] Extracting features from the user's historical behavior data to obtain basic features, contact features, preference features, and perception features of the user's history;

[0012] Obtaining historical behavior characteristics of the user according to basic characteristics, contact characteristics, preference characteristics and perception characteristics of the user history;

[0013] The user's historical behavior characteristics are input into a preset generative adversarial network model to obtain the user's future behavior characteristics.

[0014] Furthermore, the user's historical behavior data includes user basic data, user contact data, user preference data and user perception data, and the feature extraction of the user's historical behavior data to obtain the basic features, contact features, preference features and perception features of the user's history specifically includes:

[0015] Calculating basic features of the user history based on features of each user's basic information in the user basic data;

[0016] Calculating the contact characteristics of the user history according to the characteristics of each user contact information in the user contact data;

[0017] Calculating the preference characteristics of the user history according to the characteristics of each user preference information in the user preference data;

[0018] The perception feature of the user history is calculated according to the feature of each user perception information in the user perception data.

[0019] Furthermore, the calculation formula of the basic characteristics of the user history is:

[0020]

[0021] Among them, A is the basic feature of user history, a m is the feature of the mth user’s basic information in the user’s basic data, w m is the weight corresponding to the mth user basic information in the user basic data, and M is the number of user basic information;

[0022] The calculation formula of the contact characteristics of the user history is:

[0023]

[0024] Among them, B is the user's historical contact characteristics, b j is the feature of the jth user contact information in the user contact data, w j is the weight corresponding to the jth user contact information in the user contact data, and J is the number of user contact information;

[0025] The calculation formula of the user's historical preference characteristics is:

[0026]

[0027] Among them, C is the user's historical preference feature, c k is the feature of the kth user preference information in the user preference data, w k is the weight corresponding to the kth user preference information in the user preference data, and K is the number of user preference information;

[0028] The calculation formula of the perception feature of the user history is:

[0029]

[0030] Where D is the perceived feature of the user's history, d l is the feature of the lth user perception information in the user perception data, w l is the weight corresponding to the lth user perception information in the user perception data, and L is the number of user perception information.

[0031] Furthermore, the generative adversarial network model includes a generator and a discriminator, the generator is used to generate the user's future behavior characteristics, and the discriminator is used to discriminate the feature similarity between the user's historical behavior characteristics and the user's future behavior characteristics.

[0032] Furthermore, the fusing of the historical behavior characteristics and future behavior characteristics of the user to obtain the data characteristics of the user specifically includes:

[0033] The data characteristics of the user are obtained by the following formula:

[0034]

[0035] Among them, U is the user's data characteristics, H is the user's historical behavior characteristics, and W h is the weight corresponding to the user's historical behavior characteristics, F is the user's future behavior characteristics, and W f The weight corresponding to the user's future behavior characteristics.

[0036] Furthermore, the complaint refund prediction model includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer and a SoftMax layer connected in sequence, wherein the fully connected layer is used to fully connect the features processed by the second pooling layer with the contact features and perception features of the user's history.

[0037] In a second aspect, the present invention provides a complaint refund prediction device, comprising:

[0038] The acquisition module is used to obtain the user's historical behavior data;

[0039] An obtaining module, connected to the acquisition module, for obtaining the historical behavior characteristics and future behavior characteristics of the user according to the historical behavior data of the user;

[0040] A fusion module, connected to the obtaining module, for fusing the historical behavior characteristics and future behavior characteristics of the user to obtain the data characteristics of the user;

[0041] A refund prediction module is connected to the fusion module and is used to input the user's data features into a preset complaint refund prediction model, predict the user's complaint refund amount through the complaint refund prediction model, and obtain a prediction result.

[0042] In a third aspect, the present invention provides a complaint refund prediction device, comprising 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 complaint refund prediction method described in the first aspect above.

[0043] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the complaint refund prediction method described in the first aspect is implemented.

[0044] The complaint refund prediction method, device and computer-readable storage medium provided by the present invention first obtain the user's historical behavior data; then obtain the user's historical behavior characteristics and future behavior characteristics based on the user's historical behavior data; then fuse the user's historical behavior characteristics and future behavior characteristics to obtain the user's data characteristics; finally, input the user's data characteristics into a preset complaint refund prediction model, and predict the user's complaint refund amount through the complaint refund prediction model to obtain a prediction result. The present invention generates the user's historical behavior characteristics and future behavior characteristics through the user's historical behavior data, and based on the user's data characteristics obtained by fusing the user's historical behavior characteristics and future behavior characteristics, uses the complaint refund prediction model to predict the amount of refunds generated by user complaints, thereby manually intervening in the predicted refund users before user complaints lead to refunds, reducing company losses, improving user perception, and solving the problem that existing complaint prediction methods can only predict whether users will complain, but cannot predict whether user complaints will result in refunds. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flowchart of a complaint refund prediction method according to Embodiment 1 of the present invention;

[0046] Figure 2 A flowchart of generating future behavior data of a user according to an embodiment of the present invention;

[0047] Figure 3 A schematic diagram of splicing historical behavior characteristics of a user according to an embodiment of the present invention;

[0048] Figure 4 A schematic diagram of the architecture of a generative adversarial network model according to an embodiment of the present invention;

[0049] Figure 5 A flowchart of predicting the refund amount for user complaints according to an embodiment of the present invention;

[0050] Figure 6 A schematic diagram of the structure of a complaint refund prediction model according to an embodiment of the present invention;

[0051] Figure 7 This is a schematic diagram of the structure of a complaint refund prediction device according to Embodiment 2 of the present invention;

[0052] Figure 8 This is a structural diagram of a complaint refund prediction device according to Example 3 of the present invention. DETAILED DESCRIPTION

[0053] In order to enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0054] It should be understood that the specific embodiments and drawings described herein are only used to explain the present invention rather than to limit the present invention.

[0055] It can be understood that, in the absence of conflict, the various embodiments of the present invention and the various features in the embodiments can be combined with each other.

[0056] It can be understood that, for the convenience of description, the drawings of the present invention only show the parts related to the present invention, while the parts irrelevant to the present invention are not shown in the drawings.

[0057] It can be understood that each unit and module involved in the embodiments of the present invention may correspond to only one physical structure, or may be composed of multiple physical structures, or multiple units and modules may be integrated into one physical structure.

[0058] It can be understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of the present invention may occur in an order different from that marked in the drawings.

[0059] It is understood that the flowcharts and block diagrams of the present invention illustrate the possible architectures, functions, and operations of the systems, devices, equipment, and methods according to the various embodiments of the present invention. Each box in the flowchart or block diagram may represent a unit, module, program segment, or code, which contains executable instructions for implementing the specified functions. Moreover, each box or combination of boxes in the block diagram and flowchart may be implemented by a hardware-based system that implements the specified functions, or may be implemented by a combination of hardware and computer instructions.

[0060] It can be understood that the units and modules involved in the embodiments of the present invention can be implemented by software or hardware. For example, the units and modules can be located in a processor.

[0061] Application Overview

[0062] With the development of network technology, there are more and more mobile Internet users and broadband users. Users are pursuing higher and higher network quality, and the amount of complaints and refunds caused by various faults is getting larger and larger. At present, the way the customer service industry responds to complaints is often based on a single dimension and is a post-event remedial treatment. This single and lagging treatment method has the following disadvantages: 1. Low efficiency in complaint handling: Complaint handling relies on manual processing. If the number of complaints is large, it will lead to untimely complaint handling. 2. Unable to intervene in advance: When a complaint is generated, it has already caused a bad experience for users, and the financial losses caused to the company cannot be avoided. 3. The complaint handling results are not accurate enough: Complaint handling depends on the professional quality of customer service personnel, resulting in the complaint handling results not being able to satisfy users and thus generating secondary complaints. As a service provider, operators can collect users' daily Internet call quality data. How to make full use of this quality data to predict users who may complain before user complaints are generated and refunds are caused, and curb the occurrence of complaints and refunds from the source, and prevent problems before they happen, is particularly important.

[0063] Existing methods for complaint prediction include:

[0064] 1. A prediction method and prediction system for potential complaint users of home broadband (application number CN202310935830.2), which constructs a feature pool based on home broadband batch data and home feature data and obtains a training sample set; inputs the training sample set into a complaint prediction model constructed based on a random forest algorithm for training to obtain a trained complaint prediction model; finally, the trained complaint prediction model is used to determine the complaint prediction results of the potential complaint users.

[0065] 2. A complaint prediction processing method and device (application number CN202311037320.X), the method comprising: obtaining customer full journey data; standardizing the customer full journey data to obtain customer standard feature data; inputting the customer standard feature data into a complaint prediction model to perform complaint prediction and obtain complaint prediction results; and generating complaint prevention measures based on the complaint prediction results.

[0066] 3. User complaint prediction method, device and storage medium (application number CN202310675071.0), the method includes: determining user complaint data; inputting the user portrait data of the user to be predicted into the first model for hidden layer feature extraction to determine the first feature data; the first model is used to perform at least one hidden layer feature dimensionality increase and hidden layer feature dimensionality reduction on the data; the first feature data is input into the second model for convolutional layer feature extraction to determine the second feature data; the second model is used to perform at least one convolutional layer feature extraction on the data; the second feature data is input into the third model for complaint prediction to determine the user complaint probability, the third model is used to perform at least one hidden layer feature dimensionality reduction on the data to determine the third feature data, and predict the user's complaint probability based on the third feature data.

[0067] However, the existing complaint prediction methods use data modeling, which can only predict whether users will complain, but cannot predict whether user complaints will result in refunds. In essence, they are classification methods based on rule customization, with low accuracy and many limiting factors. Some complaint prediction methods use the random forest model in machine learning to predict the complaint rate, which is inefficient and cannot fully extract user behavior characteristics to accurately predict the refund amount.

[0068] In response to the above technical problems, the present application provides a complaint refund prediction method, device and computer-readable storage medium, which generates the user's historical behavior characteristics and future behavior characteristics through the user's historical behavior data, and uses the user's data characteristics obtained by fusing the user's historical behavior characteristics and future behavior characteristics, and uses a complaint refund prediction model to predict the amount of refund generated by user complaints, thereby manually intervening on the predicted refund users before the user complaint leads to a refund, reducing the company's losses and improving user perception, so as to at least solve the problem that the existing complaint prediction method can only predict whether the user will complain, but cannot predict whether the user complaint will result in a refund.

[0069] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0070] Embodiment 1:

[0071] This embodiment provides a complaint refund prediction method, such as Figure 1As shown, the method includes:

[0072] Step S101: Obtain the user's historical behavior data.

[0073] It should be noted that the user's historical behavior data includes basic user data, user contact data, user preference data and user perception data. Among them, basic user data includes user gender, age, region, mobile network package monthly rent, broadband package monthly rent, router model, network access time, device working time and other data information. User contact data includes the number of user complaints, complaint reasons, complaint classification, number of repeated complaints, number of failures, number of repeated failures, return visit satisfaction and other data information. User preference data includes user daily average traffic data, user Internet habits, commonly used APP categories and other data information. User perception data includes user network uplink rate, user network downlink rate, broadband uplink rate, broadband downlink rate, broadband quality data, language call quality data, SMS data, device online and offline times, network packet loss rate, game delay ping and other data information.

[0074] Step S102: Obtain the historical behavior characteristics and future behavior characteristics of the user based on the historical behavior data of the user.

[0075] In this embodiment, in order to accurately predict whether a user complaint will result in a refund, the user's historical behavior characteristics and future behavior characteristics are generated based on the user's historical behavior data.

[0076] In an optional embodiment, obtaining the historical behavior characteristics and future behavior characteristics of the user according to the historical behavior data of the user specifically includes:

[0077] Extracting features from the user's historical behavior data to obtain basic features, contact features, preference features, and perception features of the user's history;

[0078] Obtaining historical behavior characteristics of the user according to basic characteristics, contact characteristics, preference characteristics and perception characteristics of the user history;

[0079] The user's historical behavior characteristics are input into a preset generative adversarial network model to obtain the user's future behavior characteristics.

[0080] Specifically, the specific method for extracting features from the user's historical behavior data to obtain the basic features, contact features, preference features, and perception features of the user's history may be:

[0081] (1) calculating the basic features of the user history according to the features of each user's basic information in the user basic data;

[0082] It should be noted that, considering that the basic data of users include user gender, age, region, monthly rent of mobile network package, monthly rent of broadband package, router model, network access time, device working time and other data information, the basic features of user history should include the features of all the above basic information of users. In order to extract the basic features of user history from all the above basic information of users, firstly, feature extraction is performed on each type of basic information of users, and then weighted summation is performed on each type of feature information to obtain the basic features of user history;

[0083] Specifically, the characteristics of each type of user basic information are defined as a m , a m ∈R 1×n , through weighted summation, the basic features of user history A, A∈R 1×n ;

[0084] (2) calculating the contact characteristics of the user history according to the characteristics of each user contact information in the user contact data;

[0085] It should be noted that considering that user contact data includes data information such as the number of user complaints, complaint reasons, complaint classification, number of repeated complaints, number of failures, number of repeated failures, and return visit satisfaction, we first extract features from various types of user contact information, and then perform weighted summation on various types of feature information to obtain the user's historical contact features;

[0086] Specifically, the characteristics of each type of user contact information are defined as b j , b j ∈R 1×n , through weighted summation, the user's historical contact feature B can be calculated, B∈R 1×n ;

[0087] (3) calculating the preference characteristics of the user history according to the characteristics of each user preference information in the user preference data;

[0088] It should be noted that considering that user preference data includes user daily average traffic data, user Internet habits, commonly used APP categories and other data information, we first extract features of various types of user preference information, and then perform weighted summation on various types of feature information to obtain user historical preference features;

[0089] Specifically, the characteristics of each type of user preference information are defined as c k , c k ∈R 1×n , through weighted summation, the user's historical preference feature C can be calculated, C∈R 1×n ;

[0090] (4) calculating the perception feature of the user history according to the feature of each user perception information in the user perception data;

[0091] It should be noted that, considering that user perception data includes user network uplink rate, user network downlink rate, broadband uplink rate, broadband downlink rate, broadband quality data, voice call quality data, SMS data, device online and offline times, network packet loss rate, game latency ping and other data information, firstly, feature extraction is performed on various types of user perception information, and then weighted summation is performed on various types of feature information to obtain user historical perception features;

[0092] Specifically, the characteristics of each type of user perception information are defined as d l , d l ∈R 1×n , through weighted summation, the user's historical perception feature D can be calculated, D∈R 1×n .

[0093] Specifically, the calculation formula of the basic characteristics of the user history is:

[0094]

[0095] Among them, A is the basic feature of user history, a m is the feature of the mth user’s basic information in the user’s basic data, w m is the weight corresponding to the mth user basic information in the user basic data, and M is the number of user basic information;

[0096] The calculation formula of the contact characteristics of the user history is:

[0097]

[0098] Among them, B is the user's historical contact characteristics, b j is the feature of the jth user contact information in the user contact data, w j is the weight corresponding to the jth user contact information in the user contact data, and J is the number of user contact information;

[0099] The calculation formula of the user's historical preference characteristics is:

[0100]

[0101] Among them, C is the user's historical preference feature, c k is the feature of the kth user preference information in the user preference data, w k is the weight corresponding to the kth user preference information in the user preference data, and K is the number of user preference information;

[0102] The calculation formula of the perception feature of the user history is:

[0103]

[0104] Where D is the perceived feature of the user's history, d l is the feature of the lth user perception information in the user perception data, w l is the weight corresponding to the lth user perception information in the user perception data, and L is the number of user perception information.

[0105] Specifically, by combining the basic features of the user's history, the contact features of the user's history, the preference features of the user's history, and the perception features of the user's history, we can obtain the user's historical behavior features H, H∈R 4×n .

[0106] Specifically, the user's historical behavior features H are input into the preset generative adversarial network model, and the user's future behavior features F can be obtained by using the generative adversarial network model, F∈R 4×n .

[0107] It should be noted that the generative adversarial network model includes a generator and a discriminator. The generator is used to generate the user's future behavior characteristics, and the discriminator is used to determine the feature similarity between the user's historical behavior characteristics and the user's future behavior characteristics. The discriminator's judgment results include two situations:

[0108] (1) If the feature similarity between the user's future behavior feature and the user's historical behavior feature is greater than or equal to 85%, it means that the user's future behavior feature is available, and the user's future behavior feature data is recorded and stored in the local database;

[0109] (2) If the feature similarity between the user's future behavior features and the user's historical behavior features is less than 85%, the generator is returned to regenerate the user's future behavior features. Unlike the first generation of the user's future behavior features, the subsequent generation of the user's future behavior features is generated based on the previous generation of the user's future behavior features.

[0110] Step S103: Fusing the historical behavior characteristics and future behavior characteristics of the user to obtain data characteristics of the user.

[0111] Specifically, through the user's historical behavior weight W h and the user's future behavior weight W f The user's historical behavior features H and the user's future behavior features F are combined to obtain the user's data features U, where W h ∈R n×4 , W f ∈Rn×4 , U∈R n ×n .

[0112] Specifically, the data characteristics of the user are obtained by the following formula:

[0113]

[0114] Among them, U is the user's data characteristics, H is the user's historical behavior characteristics, and W h is the weight corresponding to the user's historical behavior characteristics, F is the user's future behavior characteristics, and W f The weight corresponding to the user's future behavior characteristics.

[0115] Step S104: inputting the data features of the user into a preset complaint refund prediction model, and predicting the complaint refund amount of the user through the complaint refund prediction model to obtain a prediction result.

[0116] In this embodiment, the complaint refund prediction model is used to predict the refund amount generated by user complaints, wherein the input data of the complaint refund prediction model is the user's data feature U.

[0117] In an optional embodiment, the complaint refund prediction model includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer and a SoftMax layer connected in sequence, wherein the fully connected layer is used to fully connect the features processed by the second pooling layer with the contact features and perception features of the user's history.

[0118] Specifically, the complaint refund prediction model can be divided into a six-layer architecture: convolution layer 1 (also known as the first convolution layer), pooling layer 1 (also known as the first pooling layer), convolution layer 2 (also known as the second convolution layer), pooling layer 2 (also known as the second pooling layer), fully connected layer, and SoftMax layer. At the same time, considering that the contact feature B and perception feature D in the user's historical behavior characteristics have a direct impact on the user's complaint refund, the user's historical contact feature B and the user's historical perception feature D are introduced in the fully connected layer.

[0119] It is worth mentioning that the existing technology predicts whether the user will complain, which is a binary classification problem. It cannot predict whether the user complaint will result in a refund, and it cannot predict the specific refund amount. The present invention can accurately predict the probability of refunds generated by user complaints and the amount of refunds generated based on the user's historical behavior data, which belongs to a multi-classification problem. In terms of complaint prediction methods, the existing technology adopts methods for user complaint prediction, including 1. Traditional rule-based customization method 2. Machine learning random forest method 3. Convolutional neural network method. The above methods extract user feature information based on user historical data, which is somewhat one-sided and easy to lose certain user feature information. The complaint refund prediction method provided by the present invention generates the user's historical behavior characteristics and future behavior characteristics based on the user's historical behavior data, and then predicts the user's complaint refund amount based on the user's historical behavior characteristics and future behavior characteristics. The complaint refund prediction model provided by the present invention can predict the amount of refunds generated by user complaints, rather than simply predicting whether a refund will be generated due to user complaints. In addition, in the complaint refund prediction model provided by the present invention, the fully connected layer is transformed, and the contact characteristics and perception characteristics of the user's history are introduced, which further improves the accuracy and generalization ability of the model prediction.

[0120] In a specific embodiment, in order to accurately predict the probability of refunds resulting from user complaints, manually intervene in advance on users with high refund rates, reduce company losses, and improve user perception, the present invention provides a complaint refund prediction method. The method can generate user behavior data that may occur in the future based on user historical behavior data such as user basic data, user contact data, user preference data, and user perception data, and accurately predict user complaints and the amount of refunds based on the user's historical behavior data and future behavior data. The content of the present invention includes two parts: one is to generate users' future behavior data, and the other is to predict the refund amount for user complaints.

[0121] 1. Generate user’s future behavior data

[0122] In daily production and life, user behavior data implies the user's daily usage experience. How to mine the user behavior characteristics in the user behavior data is the key to accurately predicting refunds caused by user complaints. However, usually, companies collect user behavior data after the user behavior has occurred, and the data characteristics are lagging. If complaint refund predictions are made based on the lagging user behavior characteristics, the prediction effect will be unsatisfactory. Considering that the user's future behavior data will directly affect whether the user complains, the present invention creatively generates the user's future behavior data based on the user's historical behavior data, and improves the accuracy of predicting the user complaint refund amount by extracting the user behavior characteristics in the user's historical data and the user's future data.

[0123] like Figure 2 As shown in FIG, the specific method for generating the user's future behavior data is: obtaining the user's historical behavior data, preprocessing the user's historical behavior data, extracting features from the preprocessed user's historical behavior data, and generating the user's future behavior data through a generative adversarial network model based on the extracted features (i.e., the user's historical basic features, contact features, preference features, and perception features). The above mainly includes four parts: data collection (i.e., obtaining the user's historical behavior data), data preprocessing, data feature extraction, and generative adversarial network model training. The specific implementation details and implementation methods of each part are introduced in detail below.

[0124] 1. Data Collection

[0125] 1. Determine the type of user data

[0126] In order to ensure the scientific nature of user data types, the user data types that need to be used are determined by studying the correlation between user data and complaints. The specific situation is as follows:

[0127] (1) Basic user data. This includes user gender, age, region, monthly mobile network package rental, monthly broadband package rental, router model, network access time, device operating time, and other data information;

[0128] (2) User contact data. This includes the number of user complaints, reasons for complaints, complaint classification, number of repeated complaints, number of failures, number of repeated failures, and satisfaction with follow-up visits;

[0129] (3) User preference data. This includes user daily average traffic data, user Internet habits, commonly used APP categories, and other data information;

[0130] (4) User perception data. This includes user network uplink rate, user network downlink rate, broadband uplink rate, broadband downlink rate, broadband quality data, voice call quality data, SMS data, device online and offline times, network packet loss rate, game latency ping, and other data information.

[0131] 2. Collect user data

[0132] The above four types of user data are collected through scheduled crawling and real-time updating:

[0133] (1) Use Python to develop a crawler program to automatically crawl user complaint data, network quality data, etc., and store it in the local database system;

[0134] (2) Write SQL stored procedures and execute them at a scheduled time to perform simple deduplication processing on the crawled data and compile them into a data base to provide data support for subsequent applications.

[0135] (II) Data preprocessing

[0136] The purpose of data preprocessing is to clean, process, and encode the collected user behavior data to provide data support for data feature extraction:

[0137] (1) Non-numeric data

[0138] Since user basic data, user contact data, user preference data, and user perception data contain a large number of non-numeric types, the collected user behavior data is first converted to different types. The rules for type conversion are as follows: exhaustively enumerate the values ​​of various ordered non-numeric indicators, and then sort them in descending order according to the number of times each value appears. The value with the most occurrences is assigned to 0, and the value with the least occurrences is assigned to n. The values ​​of various unordered non-numeric indicators, such as "gender" and "complaint category", are assigned to 0 and 1.

[0139] (2) Numerical data

[0140] Considering that the collected user data also contains a large amount of numerical data, in order to ensure the availability of data, the null values ​​of numerical data are first filled with zero values, and then normalized according to each category. At the same time, the noise data is cleaned and removed in a timely manner to prevent the noise data from affecting the generalization ability of the model during training.

[0141] Through data preprocessing, the characteristics of various types of user information can be obtained.

[0142] (III) Data feature extraction

[0143] (1) Basic characteristics of user history

[0144] Considering that the basic data of users include user gender, age, region, monthly rent of mobile network package, monthly rent of broadband package, router model, network access time, device working time and other data information, the basic features of user history should include the features of all the above basic information of users. In order to extract the basic features of user history from all the above basic information of users, firstly, feature extraction is performed on each type of basic information of users, and then weighted summation is performed on each type of feature information to obtain the basic features of user history;

[0145] Specifically, the characteristics of each type of user basic information are defined as a m , a m ∈R 1×n , through weighted summation, the basic features of user history A, A∈R 1×n ;

[0146] The calculation formula for the basic characteristics of user history is:

[0147]

[0148] Among them, A is the basic feature of user history, a m is the feature of the mth user’s basic information in the user’s basic data, w m is the weight corresponding to the mth user basic information in the user basic data, and M is the amount of user basic information.

[0149] (2) User’s historical contact characteristics

[0150] Considering that user contact data includes user complaint times, complaint reasons, complaint classification, repeated complaint times, failure times, repeated failure times, return visit satisfaction and other data information, we first extract features from various types of user contact information, and then perform weighted summation on various types of feature information to obtain the user's historical contact features;

[0151] Specifically, the characteristics of each type of user contact information are defined as b j , b j ∈R 1×n , through weighted summation, the user's historical contact feature B can be calculated, B∈R 1×n ;

[0152] The calculation formula for the user's historical contact characteristics is:

[0153]

[0154] Among them, B is the user's historical contact characteristics, b j is the feature of the jth user contact information in the user contact data, w j is the weight corresponding to the jth user contact information in the user contact data, and J is the number of user contact information.

[0155] (3) User historical preference characteristics

[0156] Considering that user preference data includes user daily average traffic data, user Internet habits, commonly used APP categories and other data information, we first extract features of various types of user preference information, and then perform weighted summation on various types of feature information to obtain the user's historical preference features;

[0157] Specifically, the characteristics of each type of user preference information are defined as c k , c k ∈R 1×n , through weighted summation, the user's historical preference feature C can be calculated, C∈R 1×n ;

[0158] The calculation formula of user history preference characteristics is:

[0159]

[0160] Among them, C is the user's historical preference feature, c k is the feature of the kth user preference information in the user preference data, w k is the weight corresponding to the kth user preference information in the user preference data, and K is the number of user preference information.

[0161] (4) Perceptual characteristics of user history

[0162] Considering that user perception data includes user network uplink rate, user network downlink rate, broadband uplink rate, broadband downlink rate, broadband quality data, voice call quality data, SMS data, device online and offline times, network packet loss rate, game latency ping and other data information, we first extract features from various types of user perception information, and then perform weighted summation on various types of feature information to obtain the user's historical perception features;

[0163] Specifically, the characteristics of each type of user perception information are defined as d l , d l ∈R 1×n , through weighted summation, the user's historical perception feature D can be calculated, D∈R 1×n ;

[0164] The calculation formula of the perceptual characteristics of user history is:

[0165]

[0166] Where D is the perceived feature of the user's history, d l is the feature of the lth user perception information in the user perception data, w l is the weight corresponding to the lth user perception information in the user perception data, and L is the number of user perception information.

[0167] (5) User’s historical behavior characteristics

[0168] like Figure 3 As shown in Figure 2, by combining the basic features of the user's history, the contact features of the user's history, the preference features of the user's history, and the perception features of the user's history, we can get the user's historical behavior features H, H∈R 4×n .

[0169] (IV) Generative Adversarial Network Model Training

[0170] The purpose of model training is to optimize the parameters of each network layer of the generative adversarial network model and improve the accuracy of the model in generating future data for users.

[0171] Specifically, the architecture of the generative adversarial network model is as follows Figure 4As shown in Figure 1, the generative adversarial network model mainly consists of two parts: the generator and the discriminator. The generator uses the initial random matrix to generate the user's future behavior characteristics, while the discriminator determines the similarity between the user's historical behavior characteristics and the user's future behavior characteristics generated by the generator. The discriminator's judgment results include two situations:

[0172] (1) If the similarity between the user's future behavior characteristics and the user's historical behavior characteristics is greater than or equal to 85%, it means that the user's future behavior characteristics are available, and the user's future behavior characteristic data is recorded and stored in the local database.

[0173] (2) If the similarity between the user's future behavior characteristics and the user's historical behavior characteristics is less than 85%, the generator is returned to regenerate the user's future behavior characteristics. Unlike the first generation of the user's future behavior characteristics, the subsequent generation of the user's future behavior characteristics is generated based on the user's previous future behavior characteristics. This is a process of continuous optimization and continuous approach to the most real data, thereby ensuring the accuracy and availability of the generated user's future behavior characteristics.

[0174] Specifically, the generator includes an input layer, an upsampling layer, a convolution layer, and a fully connected layer. The input layer is a random matrix with a dimension of 100; the upsampling layer upsamples the random matrix and increases the matrix dimension to 128; the convolution layer uses 32 3×3 filters with a step size of 1 to convolve the random matrix to 64 and uses Relu activation; finally, the user's future behavior characteristics are output through the fully connected layer, and Tanh is used as the activation function to limit the value between -1 and 1.

[0175] The discriminator includes convolution layer 1, convolution layer 2, and SoftMax layer. Convolution layer 1 uses 64 3x3 filters with a step size of 2 and ReLU as the activation function; convolution layer 2 uses 128 3x3 filters with a step size of 2 and ReLU as the activation function; the SoftMax layer outputs the similarity between the user's future behavior characteristics and the user's historical behavior characteristics.

[0176] Using the trained generative adversarial network model, we can obtain the user's future behavior characteristics F, F∈R 4×n .

[0177] So far, the user's future behavior data has been predicted through the user's historical behavior data. Next, the second part will be introduced in detail, predicting the refund amount of user complaints.

[0178] 2. Predicting the refund amount of user complaints

[0179] Considering that different users will receive different refund amounts if their complaints result in refunds, predicting the refund amount for user complaints is essentially a multi-classification task.

[0180] like Figure 5 As shown in the figure, the specific method for predicting the refund amount of user complaints is: obtaining the user's historical behavior data and generating the user's future behavior data, preprocessing the user's historical behavior data and future behavior data, extracting the user's data features from the preprocessed user's historical behavior data and future behavior data, and predicting the user's complaint refund amount through the complaint refund prediction model based on the extracted user's data features. The above includes four parts: obtaining data, preprocessing data, extracting user data features, and training the complaint refund prediction model. The data acquisition and data preprocessing work has been completed in the first part, so the following will introduce in detail how to extract user data features and train the complaint refund prediction model.

[0181] 1. Extracting User Data Features

[0182] In the first part, the basic features A of the user's history, the contact features B of the user's history, the preference features C of the user's history, and the perception features D of the user's history have been extracted, and these four types of features have been fused into the user's historical behavior features H. At the same time, the user's historical behavior features H are used to generate the user's future behavior features F. In order to improve the accuracy of the complaint refund prediction model, the user's historical behavior features H and the user's future behavior features F are fused to obtain the user's data features U:

[0183]

[0184] Among them, U is the user's data characteristics, H is the user's historical behavior characteristics, and W h is the weight corresponding to the user's historical behavior characteristics, F is the user's future behavior characteristics, and W f is the weight corresponding to the user's future behavior characteristics, W h ∈R n×4 , W f ∈R n ×4 , U∈R n×n .

[0185] At this point, the basic characteristics of user historical behavior and user future behavior, contact characteristics, preference characteristics, and perception characteristics are integrated.

[0186] (II) Training the complaint refund prediction model

[0187] In order to predict the refund amount generated by user complaints, the present invention provides a complaint refund prediction model. Figure 6As shown, the model is divided into a six-layer architecture of convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, fully connected layer, and SoftMax layer. The input data of the model is the user data feature U extracted in the previous step. At the same time, considering that the contact feature B and the perception feature D in the user's historical behavior characteristics have a direct impact on the user's complaint refund, the present invention creatively introduces the user's historical contact feature B and the user's historical perception feature D in the fully connected layer.

[0188] Specifically, in the complaint refund prediction model, the parameters of each layer are:

[0189] (1) Convolutional layer 1 uses 32 3×3 filters with a stride of 1 and ReLU as the activation function.

[0190] (2) Pooling layer 1 uses maximum pooling with a window of 2×2 and a stride of 2.

[0191] (3) Convolutional layer 2 uses 64 3×3 filters with a stride of 1 and also uses ReLU as the activation function.

[0192] (4) Pooling layer 2 also uses maximum pooling with a window of 2×2 and a stride of 2.

[0193] (5) The fully connected layer consists of 128 neurons and uses ReLU as the activation function.

[0194] At this point, we have achieved the prediction of the refund amount for users who complain about refunds.

[0195] By training the complaint refund prediction model to predict the amount of refunds generated by user complaints, the company can better manage risks, improve customer satisfaction, optimize operating strategies, enhance financial planning capabilities, enhance market competitiveness, and improve decision-making efficiency, thereby providing strong support for the company's development.

[0196] The complaint refund prediction method provided by the embodiment of the present invention first obtains the user's historical behavior data; then obtains the user's historical behavior characteristics and future behavior characteristics based on the user's historical behavior data; then fuses the user's historical behavior characteristics and future behavior characteristics to obtain the user's data characteristics; finally, inputs the user's data characteristics into a preset complaint refund prediction model, and predicts the user's complaint refund amount through the complaint refund prediction model to obtain a prediction result. The present invention generates the user's historical behavior characteristics and future behavior characteristics through the user's historical behavior data, and based on the user's data characteristics obtained by fusing the user's historical behavior characteristics and future behavior characteristics, uses the complaint refund prediction model to predict the amount of refund generated by the user's complaint, thereby manually intervening in the predicted refund user before the user complaint leads to a refund, reducing the company's losses, improving user perception, and solving the problem that the existing complaint prediction method can only predict whether the user will complain, but cannot predict whether the user complaint will result in a refund.

[0197] Embodiment 2:

[0198] like Figure 7 As shown, this embodiment provides a complaint refund prediction device, which is used to execute the above-mentioned complaint refund prediction method, including:

[0199] An acquisition module 11 is used to acquire historical behavior data of a user;

[0200] An obtaining module 12, connected to the obtaining module 11, is used to obtain the historical behavior characteristics and future behavior characteristics of the user according to the historical behavior data of the user;

[0201] A fusion module 13, connected to the obtaining module 12, is used to fuse the historical behavior characteristics and future behavior characteristics of the user to obtain the data characteristics of the user;

[0202] The refund prediction module 14 is connected to the fusion module 13 and is used to input the data features of the user into a preset complaint refund prediction model, and predict the complaint refund amount of the user through the complaint refund prediction model to obtain a prediction result.

[0203] Furthermore, the obtaining module 12 specifically includes:

[0204] A feature extraction unit, used to extract features from the historical behavior data of the user to obtain basic features, contact features, preference features and perception features of the user's history;

[0205] an obtaining unit, configured to obtain the historical behavior characteristics of the user according to the basic characteristics, contact characteristics, preference characteristics and perception characteristics of the user history;

[0206] The input obtaining unit is used to input the historical behavior characteristics of the user into a preset generative adversarial network model to obtain the future behavior characteristics of the user.

[0207] Furthermore, the user's historical behavior data includes user basic data, user contact data, user preference data and user perception data, and the feature extraction unit specifically includes:

[0208] A first calculation unit, configured to calculate basic features of the user history according to features of each user's basic information in the user basic data;

[0209] A second calculation unit, configured to calculate the contact characteristics of the user history according to the characteristics of each user contact information in the user contact data;

[0210] a third calculation unit, configured to calculate the preference feature of the user history according to the feature of each user preference information in the user preference data;

[0211] The fourth calculation unit is used to calculate the perception feature of the user history according to the feature of each user perception information in the user perception data.

[0212] Furthermore, the calculation formula of the basic characteristics of the user history is:

[0213]

[0214] Among them, A is the basic feature of user history, a m is the feature of the mth user’s basic information in the user’s basic data, w m is the weight corresponding to the mth user basic information in the user basic data, and M is the number of user basic information;

[0215] The calculation formula of the contact characteristics of the user history is:

[0216]

[0217] Among them, B is the user's historical contact characteristics, b j is the feature of the jth user contact information in the user contact data, w j is the weight corresponding to the jth user contact information in the user contact data, and J is the number of user contact information;

[0218] The calculation formula of the user's historical preference characteristics is:

[0219]

[0220] Among them, C is the user's historical preference feature, c kis the feature of the kth user preference information in the user preference data, w k is the weight corresponding to the kth user preference information in the user preference data, and K is the number of user preference information;

[0221] The calculation formula of the perception feature of the user history is:

[0222]

[0223] Where D is the perceived feature of the user's history, d l is the feature of the lth user perception information in the user perception data, w l is the weight corresponding to the lth user perception information in the user perception data, and L is the number of user perception information.

[0224] Furthermore, the generative adversarial network model includes a generator and a discriminator, the generator is used to generate the user's future behavior characteristics, and the discriminator is used to discriminate the feature similarity between the user's historical behavior characteristics and the user's future behavior characteristics.

[0225] Furthermore, the fusion module 13 is specifically used for:

[0226] The data characteristics of the user are obtained by the following formula:

[0227]

[0228] Among them, U is the user's data characteristics, H is the user's historical behavior characteristics, and W h is the weight corresponding to the user's historical behavior characteristics, F is the user's future behavior characteristics, and W f The weight corresponding to the user's future behavior characteristics.

[0229] Furthermore, the complaint refund prediction model includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer and a SoftMax layer connected in sequence, wherein the fully connected layer is used to fully connect the features processed by the second pooling layer with the contact features and perception features of the user's history.

[0230] Embodiment 3:

[0231] refer to Figure 8 This embodiment provides a complaint refund prediction device, including a memory 21 and a processor 22, wherein the memory 21 stores a computer program, and the processor 22 is configured to run the computer program to execute the complaint refund prediction method in Example 1.

[0232] The memory 21 is connected to the processor 22. The memory 21 may be a flash memory, a read-only memory or other memory. The processor 22 may be a central processing unit or a single-chip microcomputer.

[0233] Embodiment 4:

[0234] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the complaint refund prediction method in the above-mentioned embodiment 1 is implemented.

[0235] 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). Computer-readable storage media include, but are 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 disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer.

[0236] In summary, the complaint refund prediction method, device and computer-readable storage medium provided in the embodiments of the present invention first obtain the user's historical behavior data; then obtain the user's historical behavior characteristics and future behavior characteristics based on the user's historical behavior data; then fuse the user's historical behavior characteristics and future behavior characteristics to obtain the user's data characteristics; finally, input the user's data characteristics into a preset complaint refund prediction model, and predict the user's complaint refund amount through the complaint refund prediction model to obtain a prediction result. The present invention generates the user's historical behavior characteristics and future behavior characteristics through the user's historical behavior data, and based on the user's data characteristics obtained by fusing the user's historical behavior characteristics and future behavior characteristics, uses the complaint refund prediction model to predict the amount of refunds generated by user complaints, thereby manually intervening in the predicted refund users before user complaints lead to refunds, reducing company losses, improving user perception, and solving the problem that the existing complaint prediction methods can only predict whether users will complain, but cannot predict whether user complaints will result in refunds.

[0237] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting complaint refunds. It is characterized in that The method comprises: Obtain user's historical behavior data; Obtaining historical behavior characteristics and future behavior characteristics of the user according to the historical behavior data of the user; The historical behavior characteristics and future behavior characteristics of the user are integrated to obtain the data characteristics of the user; The data features of the user are input into a preset complaint refund prediction model, and the complaint refund amount of the user is predicted by the complaint refund prediction model to obtain a prediction result.

2. The method according to claim 1, It is characterized in that The obtaining of the historical behavior characteristics and future behavior characteristics of the user according to the historical behavior data of the user specifically includes: Extracting features from the user's historical behavior data to obtain basic features, contact features, preference features, and perception features of the user's history; Obtaining historical behavior characteristics of the user according to basic characteristics, contact characteristics, preference characteristics and perception characteristics of the user history; The user's historical behavior characteristics are input into a preset generative adversarial network model to obtain the user's future behavior characteristics.

3. The method according to claim 2, It is characterized in that The user's historical behavior data includes user basic data, user contact data, user preference data and user perception data. The feature extraction of the user's historical behavior data to obtain the basic features, contact features, preference features and perception features of the user's history specifically includes: Calculating basic features of the user history based on features of each user's basic information in the user basic data; Calculating the contact characteristics of the user history according to the characteristics of each user contact information in the user contact data; Calculating the preference characteristics of the user history according to the characteristics of each user preference information in the user preference data; The perception feature of the user history is calculated according to the feature of each user perception information in the user perception data.

4. The method according to claim 3, It is characterized in that The calculation formula of the basic characteristics of the user history is: Among them, A is the basic feature of user history, a m is the feature of the mth user’s basic information in the user’s basic data, w m is the weight corresponding to the mth user basic information in the user basic data, and M is the number of user basic information; The calculation formula of the contact characteristics of the user history is: Among them, B is the user's historical contact characteristics, b j is the feature of the jth user contact information in the user contact data, w j is the weight corresponding to the jth user contact information in the user contact data, and J is the number of user contact information; The calculation formula of the user's historical preference characteristics is: Among them, C is the user's historical preference feature, c k is the feature of the kth user preference information in the user preference data, w k is the weight corresponding to the kth user preference information in the user preference data, and K is the number of user preference information; The calculation formula of the perception feature of the user history is: Where D is the perceived feature of the user's history, d l is the feature of the lth user perception information in the user perception data, w l is the weight corresponding to the lth user perception information in the user perception data, and L is the number of user perception information.

5. The method according to claim 2, It is characterized in that The generative adversarial network model includes a generator and a discriminator, wherein the generator is used to generate future behavior features of a user, and the discriminator is used to discriminate the feature similarity between the user's historical behavior features and the user's future behavior features.

6. The method according to claim 1, It is characterized in that The fusing of the historical behavior characteristics and future behavior characteristics of the user to obtain the data characteristics of the user specifically includes: The data characteristics of the user are obtained by the following formula: Among them, U is the user's data characteristics, H is the user's historical behavior characteristics, and W h is the weight corresponding to the user's historical behavior characteristics, F is the user's future behavior characteristics, and W f The weight corresponding to the user's future behavior characteristics.

7. The method according to claim 2, It is characterized in that The complaint refund prediction model includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a fully connected layer and a SoftMax layer connected in sequence, wherein the fully connected layer is used to fully connect the features processed by the second pooling layer with the contact features and perception features of the user's history.

8. A complaint refund prediction device, It is characterized in that include: The acquisition module is used to obtain the user's historical behavior data; An obtaining module, connected to the acquisition module, for obtaining the historical behavior characteristics and future behavior characteristics of the user according to the historical behavior data of the user; A fusion module, connected to the obtaining module, for fusing the historical behavior characteristics and future behavior characteristics of the user to obtain the data characteristics of the user; A refund prediction module is connected to the fusion module and is used to input the user's data features into a preset complaint refund prediction model, predict the user's complaint refund amount through the complaint refund prediction model, and obtain a prediction result.

9. A complaint refund prediction device, It is characterized in that It 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 complaint refund prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the complaint refund prediction method according to any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • User complaint prediction method and device and storage medium

    CN116702982A

  • Household broadband potential complaint user prediction method and prediction system

    CN116976853A

  • Complaint prediction processing method and device

    CN117077857A