CN117689074A

CN117689074A - User complaint prediction method, device, equipment and medium

Info

Publication number: CN117689074A
Application number: CN202311716588.6A
Authority: CN
Inventors: 孟庆鲁; 常海涛; 修志超
Original assignee: China United Network Communications Group Co Ltd
Current assignee: China United Network Communications Group Co Ltd
Priority date: 2023-12-13
Filing date: 2023-12-13
Publication date: 2024-03-12

Abstract

The application provides a user complaint prediction method, device, equipment and medium, which are characterized in that feature data of a user to be predicted, comprising sub-feature data representing different dimensions, are obtained, and each group of sub-feature data is sequentially input into a first model to perform hidden layer feature extraction so as to obtain a plurality of initial extraction features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and inputting the matrix into a third model for hidden layer feature extraction to obtain predicted extracted features; and inputting the prediction extraction features into a fourth model for complaint prediction to obtain a complaint prediction result of the user to be predicted. The scheme can effectively predict potential complaint users.

Description

User complaint prediction method, device, equipment and medium

Technical Field

The present disclosure relates to the field of data processing, and in particular, to a method, an apparatus, a device, and a medium for predicting user complaints.

Background

In the process of using the network of the operator, the internet surfing user may complain when the network use experience or the service of staff is poor, and excessive complaints may cause the public praise of the operator and the user satisfaction to be reduced. How to deal with complaints in time and to make corresponding improvements is particularly important for customer service work.

Currently, prediction for potential complaint users is mainly based on manual judgment, and staff performs analysis and prediction by user data based on single dimensions, such as signal coverage conditions around the users or historical complaint conditions of the users. However, existing solutions have low accuracy in analysis predictions for potentially complaining users and require significant time and labor costs. Thus, how to effectively predict potential complaint users is a problem that currently needs to be addressed.

Disclosure of Invention

The application provides a user complaint prediction method, device, equipment and medium, which are used for effectively predicting potential complaint users.

In one aspect, the present application provides a method for predicting complaints of a user, including: acquiring characteristic data of a user to be predicted; the feature data comprises a plurality of groups of sub-feature data representing different dimensions, and each group of sub-feature data comprises a plurality of data assignments representing different attributes; inputting each group of sub-feature data into a first model in sequence to extract hidden layer features, and obtaining a plurality of initial extracted features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; wherein, at least one hidden layer up-scaling and hidden layer down-scaling are performed on each group of sub-feature data in the first model; converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and outputting extracted matrix data; the matrix is subjected to multiple convolution layer feature extraction in the second model, and at least two convolution kernels with different sizes exist in the convolution layer feature extraction; expanding the extraction matrix data into second extraction feature data, and inputting the second extraction feature data into a third model for extracting hidden layer features to obtain prediction extraction features; inputting the prediction extraction features into a fourth model to perform complaint prediction, obtaining a first prediction probability, and determining a complaint prediction result of the user to be predicted according to the first prediction probability; wherein the fourth model comprises an output layer comprising 1 neuron and a probability function.

In one possible implementation manner, inputting each group of sub-feature data into the first model in turn to perform hidden layer feature extraction, to obtain a plurality of initial extracted features, including: sequentially judging whether the total data assignment amount of each group of sub-feature data is smaller than a preset threshold value; if the total data assignment amount of the group of sub-feature data is smaller than a preset threshold value, performing hidden layer lifting and nonlinear conversion and hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features; if the total data assignment amount of the group of sub-feature data is not smaller than a preset threshold value, carrying out multiple hidden layer lifting and nonlinear conversion and multiple hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features.

In one possible implementation manner, the method for obtaining the extracted matrix data includes the steps of converting the first extracted feature data into a matrix, and inputting the matrix into a second model for convolutional layer feature extraction, where the method includes: converting the first extracted feature data into a maximum matrix according to the feature quantity of the first extracted feature data; performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of n×n on the maximum matrix in a first convolution layer of a second model, and performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m×m to obtain a first extraction matrix; wherein m and n are positive integers, and m is greater than n; performing convolutional layer feature extraction with a convolutional kernel of n multiplied by n on the maximum matrix in a second convolutional layer of the second model to obtain a second extraction matrix; performing matrix multiplication on the first extraction matrix and the second extraction matrix to obtain a third extraction matrix; performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m×m on the maximum matrix in a third convolution layer of the second model, and performing n×n convolution layer feature extraction and nonlinear conversion with the convolution kernel to obtain a fourth extraction matrix; performing matrix multiplication on the fourth extraction matrix and the second extraction matrix to obtain a fifth extraction matrix; performing matrix splicing on the third extraction matrix and the fifth extraction matrix to obtain a sixth extraction matrix; and performing convolutional kernel n multiplied by 2 convolutional layer feature extraction and nonlinear conversion on the sixth extraction matrix in a fourth convolutional layer of the second model to obtain a seventh extraction matrix, and taking the seventh extraction matrix as extraction matrix data output by the second model.

In one possible implementation, the third model is used to: and carrying out hidden layer ascending and nonlinear conversion and hidden layer descending and nonlinear conversion on the second extracted feature data to obtain the predicted extracted feature.

In one possible implementation, the method further includes: acquiring historical characteristic data of a plurality of users, wherein the historical characteristic data comprises characteristic data and complaint true values of the plurality of users; inputting the characteristic data in the historical characteristic data of each user into the first model, the second model, the third model and the fourth model in sequence, and outputting a second prediction probability; judging whether the second prediction probability is valid or not according to the complaint true value in the historical characteristic data of the user; if the second prediction probability is valid, finishing training of the historical feature data of the user; and if the second prediction probability is invalid, adjusting the weight value of the convolution kernel in the second model, returning to execute the step of inputting the characteristic data in the historical characteristic data of the user into the first model, the second model, the third model and the fourth model, and outputting the second prediction probability until the training of the historical characteristic data of the user is finished.

In one possible implementation, the method further includes: according to the complaint mode, carrying out complaint assignment on the historical complaints of the user; if the historical complaint times of the user are one time, assigning the complaint value as a complaint true value; if the historical complaint times of the user are larger than one time, calculating a plurality of corresponding complaint assignments and complaint times to obtain complaint true values.

In another aspect, the present application provides a user complaint prediction apparatus, including: the acquisition module is used for acquiring the characteristic data of the user to be predicted; the feature data comprises a plurality of groups of sub-feature data representing different dimensions, and each group of sub-feature data comprises a plurality of data assignments representing different attributes; the extraction module is used for sequentially inputting each group of sub-feature data into the first model to extract hidden layer features and obtain a plurality of initial extraction features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; wherein, at least one hidden layer up-scaling and hidden layer down-scaling are performed on each group of sub-feature data in the first model; the adjusting module is used for converting the first extracted characteristic data into a matrix, inputting the matrix into the second model for convolutional layer characteristic extraction, and outputting extracted matrix data; the matrix is subjected to multiple convolution layer feature extraction in the second model, and at least two convolution kernels with different sizes exist in the convolution layer feature extraction; the prediction module is used for expanding the extraction matrix data into second extraction feature data and inputting the second extraction feature data into a third model for extracting hidden layer features to obtain prediction extraction features; inputting the prediction extraction features into a fourth model to perform complaint prediction, obtaining a first prediction probability, and determining a complaint prediction result of the user to be predicted according to the first prediction probability; wherein the fourth model comprises an output layer comprising 1 neuron and a probability function.

In one possible implementation, the extracting module is configured to: sequentially judging whether the total data assignment amount of each group of sub-feature data is smaller than a preset threshold value; if the total data assignment amount of the group of sub-feature data is smaller than a preset threshold value, performing hidden layer lifting and nonlinear conversion and hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features; if the total data assignment amount of the group of sub-feature data is not smaller than a preset threshold value, carrying out multiple hidden layer lifting and nonlinear conversion and multiple hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features.

In another aspect, the present application provides an electronic device, including: a processor, a memory communicatively coupled to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method as described above.

In another aspect, the present application provides a computer-readable storage medium having stored therein computer-executable instructions that, when executed by a processor, are configured to implement a method as above.

In the user complaint prediction method, device, equipment and medium, feature data of a user to be predicted, which comprises sub-feature data representing different dimensions, are acquired, and each group of sub-feature data is sequentially input into a first model to perform hidden layer feature extraction so as to acquire a plurality of initial extracted features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and inputting the matrix into a third model for hidden layer feature extraction to obtain predicted extracted features; and inputting the prediction extraction features into a fourth model for complaint prediction to obtain a complaint prediction result of the user to be predicted. The scheme can effectively predict potential complaint users.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description, serve to explain the principles of the application.

Fig. 1 is a schematic flow chart illustrating a user complaint prediction method according to the first embodiment;

a schematic diagram of extracting initial features of a first model provided in the first embodiment is exemplarily shown in fig. 2;

FIG. 3 is a flow chart schematically showing another method for predicting complaints of users according to the first embodiment;

FIG. 4 is a flow chart schematically showing another method for predicting complaints of users according to the first embodiment;

a second model convolutional layer feature extraction pictorial intent provided in this embodiment one is illustrated in fig. 5;

FIG. 6 is a flow chart schematically showing another method for predicting complaints of users according to the first embodiment;

FIG. 7 is a flow chart schematically showing another method for predicting complaints of users according to the first embodiment;

fig. 8 is a schematic diagram schematically showing a configuration of a user complaint predicting apparatus according to the second embodiment;

fig. 9 is a schematic diagram schematically showing the structure of an electronic device according to the third embodiment.

Specific embodiments thereof have been shown by way of example in the drawings and will herein be described in more detail. These drawings and the written description are not intended to limit the scope of the inventive concepts in any way, but to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments.

Detailed Description

Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The implementations described in the following exemplary examples are not representative of all implementations consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present application as detailed in the accompanying claims.

It should be noted that the brief description of the terms in the present application is only for convenience in understanding the embodiments described below, and is not intended to limit the embodiments of the present application. Unless otherwise indicated, these terms should be construed in their ordinary and customary meaning. The terms first, second and the like in the description and in the claims and in the above-described figures are used for distinguishing between similar or similar objects or entities and not necessarily for describing a particular sequential or chronological order, unless otherwise indicated. It is to be understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the application are, for example, capable of operation in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprise" and "have," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a product or apparatus that comprises a list of elements is not necessarily limited to those elements expressly listed, but may include other elements not expressly listed or inherent to such product or apparatus. The term "module" as used in this application refers to any known or later developed hardware, software, firmware, artificial intelligence, fuzzy logic, or combination of hardware and/or software code that is capable of performing the function associated with that element.

Network operators are increasingly focusing on soft competing aspects such as marketing strategies, client public praise and satisfaction in order to promote their own competitiveness in commercial competition. When the mobile terminal user uses the network, the user can conduct corresponding complaints when the internet surfing speed and the network coverage condition or the service of the staff of the network operator is worse. One of the existing complaint treatment schemes is a passive scheme, namely, after customer service personnel receive complaints, the customer service personnel determine the interpretation caliber of the reply according to the content of the complaints and correspondingly adjust the interpretation caliber. The complaint processing efficiency of the passive scheme is low, and advanced prediction cannot be realized. Another manual prediction scheme is to analyze and predict user data of a single dimension of experienced staff, such as signal coverage around the user or historical complaints of the user. The prediction result of the manual prediction scheme has low accuracy, and a great deal of time cost and labor cost are required to be consumed.

The technical content provided by the application aims to solve the technical problems of the related technology. In the user complaint prediction method, device, equipment and medium, feature data of a user to be predicted, which comprises sub-feature data representing different dimensions, are acquired, and each group of sub-feature data is sequentially input into a first model to perform hidden layer feature extraction so as to acquire a plurality of initial extracted features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and inputting the matrix into a third model for hidden layer feature extraction to obtain predicted extracted features; and inputting the prediction extraction features into a fourth model for complaint prediction to obtain a complaint prediction result of the user to be predicted. The scheme can effectively predict potential complaint users.

The technical scheme of the present application and the technical scheme of the present application are described in detail below with specific examples. The following embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments. In the description of the present application, the terms are to be construed broadly in the art, unless explicitly stated or defined otherwise. Embodiments of the present application will be described below with reference to the accompanying drawings.

Example 1

A flowchart of a method for predicting a user complaint is shown in fig. 1, where the execution subject of this example may be a user complaint prediction device, as shown in fig. 1, and the method includes:

step 101, obtaining characteristic data of a user to be predicted; the feature data comprises a plurality of groups of sub-feature data representing different dimensions, and each group of sub-feature data comprises a plurality of data assignments representing different attributes;

102, sequentially inputting each group of sub-feature data into a first model to extract hidden layer features, and obtaining a plurality of initial extracted features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; wherein, at least one hidden layer up-scaling and hidden layer down-scaling are performed on each group of sub-feature data in the first model;

Step 103, converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and outputting extracted matrix data; the matrix is subjected to multiple convolution layer feature extraction in the second model, and at least two convolution kernels with different sizes exist in the convolution layer feature extraction;

104, expanding the extracted matrix data into second extracted feature data, and inputting the second extracted feature data into a third model for extracting hidden layer features to obtain predicted extracted features; inputting the prediction extraction features into a fourth model to perform complaint prediction, obtaining a first prediction probability, and determining a complaint prediction result of the user to be predicted according to the first prediction probability; wherein the fourth model comprises an output layer comprising 1 neuron and a probability function.

In practical application, the execution subject of the method may be a user complaint prediction device, and various implementation manners are available, for example, the method may be implemented by a computer program, for example, application software, etc.; alternatively, the computer program may be implemented as a medium storing a related computer program, for example, a usb disk, a cloud disk, or the like; still alternatively, it may be implemented by a physical device, e.g., a chip or the like, in which the relevant computer program is integrated or installed.

In this example, the feature data of the user to be predicted may be obtained, specifically, by obtaining the feature data from a server storing the feature data of the user to be predicted or a storage medium locally connected to the user complaint device. Wherein the feature data comprises a plurality of sets of sub-feature data representing different dimensions, each set of sub-feature data comprising a plurality of data assignments representing different attributes. In practical application, based on the requirements of technicians, multiple groups of sub-feature data representing different dimensions can be set; for example, a set of sub-feature data representing a user network situation may be set: network data, a set of sub-feature data that can be set to represent user portrayal situations: user data, a set of sub-feature data that may be set to represent a user's historical complaint conditions: historical data, a set of sub-feature data that may be set to represent user environmental conditions: environmental data, a set of sub-feature data that may be set to represent other situations of the user: and (5) reference data. Each group of sub-feature data comprises a plurality of data assignments representing different attributes, and correspondingly, each group of sub-feature data is endowed with the plurality of data assignments representing the different attributes according to the situation of a user. Table 1 exemplarily shows a data assignment situation of feature data of a user, and as shown in table 1, corresponding numerical values are assigned according to the number of outage times of a base station where the user resides under the attribute of the base station outage situation corresponding to a sub-feature data set of network data. It should be noted that, according to the predicted requirement, the attribute may be correspondingly added, modified or deleted, and the assigned value may be modified.

Table 1 data assignment of user's characteristic data

Wherein, CDN resource: content delivery network resources (Content Delivery Network, CDN for short). The basic idea is to avoid the bottleneck and link on the internet which possibly affects the data transmission speed and stability as much as possible, so that the content is transmitted faster and more stably. VoLTE conversation: the Long Term Evolution Voice bearer (VoLTE for short) is a high-speed wireless communication standard for mobile phones and data terminals.

After multiple groups of sub-feature data are obtained, each group of sub-feature data is sequentially input into a first model to extract hidden layer features, and multiple initial extraction features are obtained. Specifically, after the dimension of each group of sub-feature data is increased once in the first model, the dimension-increased data is subjected to one-time nonlinear conversion by using an activation function, the dimension of the non-linearly converted data is reduced once in the hidden layer, and the dimension-increased data is subjected to one-time nonlinear conversion by using the activation function so as to obtain the corresponding initial extraction feature. In practical application, the number of neurons of the hidden layer is generally larger than the total value of the assigned data of the group of sub-feature data when the dimension of the hidden layer is increased, and the number of neurons of the hidden layer is generally smaller than or equal to the total value of the assigned data of the group of sub-feature data when the dimension of the hidden layer is reduced. The input multiple groups of sub-feature data can be abstracted to another dimension space through the feature extraction of the hidden layer, so that more abstract features of the sub-feature data are displayed, and linear division is better performed. Based on the data of table 1, for example, the initial extracted features L as follows can be obtained by inputting the network data into the first model:

L＝H(f6(H(f12(F1))))

Where F1 represents a plurality of data assignments of the network data, F12 represents 12 neurons of the hidden layer when performing hidden layer dimension-up, H () represents a Hardtanh activation function, and F6 represents the number of 6 neurons of the hidden layer when performing hidden layer dimension-down. In practical application, the Hardtank activation function can be replaced by any activation function so as to achieve the effect of nonlinear conversion of data. Fig. 2 is an exemplary schematic diagram showing initial feature extraction of the first model, and as shown in fig. 2, after network data is input into the first model, corresponding initial extracted features L are obtained through nonlinear transformation of hidden layer rising and Hardtanh activation functions of 12 neurons and through nonlinear transformation of hidden layer falling and Hardtanh activation functions of 6 neurons.

The feature integrity and the independence of the same group of sub-feature data can be improved by carrying out hidden layer feature extraction on the sub-feature data in a grouping mode. After the plurality of initial extraction features are obtained, feature stitching is performed on the plurality of initial extraction features, and the plurality of initial extraction features can be fused to obtain first extraction feature data. In practical application, feature stitching can be performed on a plurality of initial extracted features through a Concat () method.

After the first extracted feature data is obtained, converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and outputting extracted matrix data; the matrix is subjected to multiple convolution layer feature extraction in the second model, and at least two convolution kernels with different sizes exist in the convolution layer feature extraction. Features of the first extracted features after convolution can be obtained through feature extraction of the convolution layers, and features of different levels can be extracted through the convolution layers with different sizes of convolution kernels.

After the extraction matrix data are obtained, the extraction matrix data are unfolded to second extraction feature data, and then the second extraction feature data are input into a third model for hidden layer feature extraction, so that predicted extraction features are obtained. The convolved features can be subjected to feature abstraction again through a third model. Inputting the prediction extraction features into a fourth model to perform complaint prediction, obtaining a first prediction probability, and determining a complaint prediction result of the user to be predicted according to the first prediction probability; wherein the fourth model comprises an output layer comprising 1 neuron and a probability function. Specifically, after feature extraction is performed through an output layer including 1 neuron, probability calculation is performed by using a probability function. In practical application, a Sigmoid function can be adopted to carry out probability calculation, and nonlinear functions with other output intervals ranging from 0 to 1 can also be adopted. In practical application, for the output first prediction probability, a user exceeding 0.65 is defined as a user with a high possibility of complaint, and the user needs to pay attention to the first prediction probability.

In the user complaint prediction method of the example, feature data of a user to be predicted including sub-feature data representing different dimensions is acquired, and each group of sub-feature data is sequentially input into a first model to perform hidden layer feature extraction so as to acquire a plurality of initial extraction features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and inputting the matrix into a third model for hidden layer feature extraction to obtain predicted extracted features; and inputting the prediction extraction features into a fourth model for complaint prediction to obtain a complaint prediction result of the user to be predicted. The scheme can effectively predict potential complaint users.

In order to improve the effect of extracting the characteristics of the sub-characteristic data in the hidden layer, the number scale of data assignment in each group of sub-characteristic data needs to be considered. As an example, fig. 3 illustrates a flowchart of a method for predicting complaints of a user, and step 102 includes, on the basis of any of the examples:

step 201, judging whether the total data assignment amount of each group of sub-feature data is smaller than a preset threshold value in sequence; if the total data assignment amount of the group of sub-feature data is smaller than a preset threshold value, performing hidden layer lifting and nonlinear conversion and hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features;

Step 202, if the total data assignment amount of the group of sub-feature data is not smaller than a preset threshold, performing multiple hidden layer lifting and nonlinear conversion and multiple hidden layer lowering and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features.

In this example, the number of times of hidden layer increase and hidden layer dimension decrease of the sub-feature data is determined according to the scale of the total data assignment amount of each group of sub-feature data. In practical application, the preset threshold value can be set to be 10, and if the total data assignment amount in the sub-feature data is smaller than 10, only one hidden layer lifting and hidden layer dimension reduction are carried out; if the total data assignment amount in the sub-feature data is not less than 10, performing primary hidden layer dimension increasing and decreasing, and then performing primary hidden layer dimension increasing and decreasing. Optionally, according to the prediction requirement, multiple thresholds may be set, and different hidden layer rising and hidden layer dimension reduction times are set in different threshold intervals. According to the method provided by the example, the sub-feature data with larger data assignment total size is subjected to multiple hidden layer lifting and hidden layer dimension reduction, noise and useful information can be separated, and the distinguishing degree of a sample is improved, so that the feature is better extracted and selected.

Considering that the characteristic of the convolution layer is lost along with the change of the convolution depth when the characteristic extraction of the convolution layer is carried out in the second model, the data after the multi-layer convolution needs to be processed. As an example, fig. 4 illustrates a flowchart of a method for predicting complaints of a user, and step 103 includes, on the basis of any of the examples:

step 301, converting the first extracted feature data into a maximum matrix according to the feature quantity of the first extracted feature data;

step 302, performing convolutional layer feature extraction and nonlinear conversion with a convolutional kernel of n×n on a maximum matrix in a first convolutional layer of a second model, and performing convolutional layer feature extraction and nonlinear conversion with a convolutional kernel of m×m to obtain a first extraction matrix; wherein m and n are positive integers, and m is greater than n;

step 303, performing convolutional layer feature extraction with a convolutional kernel of n×n on the maximum matrix in a second convolutional layer of the second model to obtain a second extraction matrix; performing matrix multiplication on the first extraction matrix and the second extraction matrix to obtain a third extraction matrix;

step 304, performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m×m on the maximum matrix in a third convolution layer of the second model, and performing n×n convolution layer feature extraction and nonlinear conversion with the convolution kernel to obtain a fourth extraction matrix; performing matrix multiplication on the fourth extraction matrix and the second extraction matrix to obtain a fifth extraction matrix;

Step 305, performing matrix splicing on the third extraction matrix and the fifth extraction matrix to obtain a sixth extraction matrix; and performing convolutional kernel n multiplied by 2 convolutional layer feature extraction and nonlinear conversion on the sixth extraction matrix in a fourth convolutional layer of the second model to obtain a seventh extraction matrix, and taking the seventh extraction matrix as extraction matrix data output by the second model.

In this example, the first extracted feature data is converted into a maximum matrix according to the feature number of the first extracted feature data. In particular, to facilitate subsequent convolution calculations and reduce feature loss, the first extracted feature data may be converted into a maximum matrix that is less than its number. For example, if the number of features of the first extracted feature data is 30, it is converted into a maximum matrix of 5×5. In practical application, the first extracted feature data can be subjected to dimension reduction and nonlinear conversion through a hidden layer containing 25 neurons, so that the first extracted feature data with 25 feature numbers is obtained and converted into a maximum matrix of 5×5.

After the maximum matrix is obtained, performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of n multiplied by n on the maximum matrix in a first convolution layer of a second model, and performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m multiplied by m to obtain a first extraction matrix; wherein m and n are positive integers, and m is greater than n. In practical application, for a maximum matrix of 5×5, n can be set to 1, m is 3, and correspondingly, the first extraction matrix G1 is:

Wherein G is the maximum matrix,a convolution layer representing a convolution kernel size of 1 x 1 +.>Representing a convolution layer with a convolution kernel size of 3 x 3, H () represents the Hardtanh activation function. It should be noted that, when the maximum matrix of 5×5 is convolved with 1×01, it is still 5×5; and a matrix of 5×5 is convolved by 3×3, a matrix reduction output of 3×3 occurs. Therefore, to avoid the reduction of matrix space for each convolution operation, it is necessary to fill (padding) around the 5×5 matrix after the 1×1 convolution so that the output after the 3×3 convolution is still 5×5. Wherein the value of the ambient filling is typically 0.

Performing convolutional layer feature extraction with a convolutional kernel of n multiplied by n on the maximum matrix in a second convolutional layer of the second model to obtain a second extraction matrix; and multiplying the first extraction matrix and the second extraction matrix by a matrix to obtain a third extraction matrix. Performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m×m on the maximum matrix in a third convolution layer of the second model, and performing n×n convolution layer feature extraction and nonlinear conversion with the convolution kernel to obtain a fourth extraction matrix; and multiplying the fourth extraction matrix and the second extraction matrix by a matrix to obtain a fifth extraction matrix. The second extraction matrix represents shallow convolution of the maximum matrix, and feature supplementation can be performed on the first extraction matrix and the fourth extraction matrix after complex convolution for a plurality of times through weighting of the second extraction matrix.

And performing matrix splicing on the third extraction matrix and the fifth extraction matrix to obtain a sixth extraction matrix. Specifically, the matrix stitching is performed in a manner of overlap stitching, for example, the third extraction matrix of 5×5 and the fifth extraction matrix of 5×5 are the sixth extraction matrix of 5×5×2 after overlap stitching. And performing convolutional kernel n multiplied by 2 convolutional layer feature extraction and nonlinear transformation on the sixth extraction matrix in a fourth convolutional layer of the second model to obtain a seventh extraction matrix, and taking the seventh extraction matrix as extraction matrix data output by the second model. The convolution process described above may be illustrated by the second model convolution layer feature extraction diagram as shown in fig. 5, and after converting the first extracted feature data into the maximum matrix, a seventh extraction matrix may be obtained through convolution multiple times, and the seventh extraction matrix may be used as the extraction matrix data output by the second model as shown in fig. 5.

According to the method provided by the example, subsequent convolution calculation can be facilitated and feature loss can be reduced by converting the first extraction feature data into the maximum matrix, matrixes with different convolution degrees can be obtained through the first extraction matrix and the fourth extraction matrix, feature supplementation can be performed on the first extraction matrix and the fourth extraction matrix through the second extraction matrix, the third extraction matrix and the fifth extraction matrix can be spliced through the sixth extraction matrix, and the sixth extraction matrix can be fused into a single-layer matrix through the seventh extraction matrix.

As one example, the third model is for: and carrying out hidden layer ascending and nonlinear conversion and hidden layer descending and nonlinear conversion on the second extracted feature data to obtain the predicted extracted feature.

In practical application, the prediction extraction characteristics can be obtained through a fully connected network. Specifically, the second extracted feature data is fully connected through a third model based on the feature quantity in the second extracted feature data. For example, if the second extracted feature data is 25, the second extracted feature data is subjected to hidden layer up-conversion and nonlinear conversion including x neurons, and to hidden layer down-conversion and nonlinear conversion including 25 neurons, to obtain the predicted extracted feature. Wherein x is an integer greater than 25 and less than 50 based on the maximum neuron restriction in the ascending dimension. The method provided by the example can perform feature extraction on the convolved second extracted feature data through the third model.

In order to improve the authenticity and accuracy of the projection prediction model, training and fitting are needed to be carried out on the model based on the historical characteristic data of the user, and model parameters are determined after the training and fitting are passed. As an example, fig. 6 illustrates a flowchart of a user complaint prediction method, where on the basis of any example, the method further includes:

Step 401, acquiring historical feature data of a plurality of users, wherein the historical feature data comprises feature data and complaint true values of the plurality of users;

step 402, feature data in the historical feature data of each user is sequentially input into a first model, a second model, a third model and a fourth model, and a second prediction probability is output;

step 403, judging whether the second prediction probability is valid according to the complaint true value in the historical feature data of the user; if the second prediction probability is valid, finishing training of the historical feature data of the user;

and step 404, if the second prediction probability is invalid, adjusting the weight value of the convolution kernel in the second model, and returning to execute the step of inputting the characteristic data in the historical characteristic data of the user into the first model, the second model, the third model and the fourth model, and outputting the second prediction probability until the training of the historical characteristic data of the user is finished.

In this example, historical feature data for a plurality of users is obtained, wherein the historical feature data includes feature data and complaint truth values for the plurality of users. In practical application, the complaint true value is the training direction of the model. And sequentially inputting the characteristic data in the historical characteristic data of each user into the first model, the second model, the third model and the fourth model, and outputting the second prediction probability. In particular, the weight values of the plurality of convolution kernels in the second model prior to non-passing through the user's historical feature data input may be random. For example, in a random case, the weight of the convolution kernel of 1×1 may be 2, and the weight of the convolution kernel of 3×3 may be any 9 values. Judging whether the second prediction probability is valid or not according to the complaint true value in the historical characteristic data of the user; and if the second prediction probability is valid, finishing training of the historical characteristic data of the user. And if the second prediction probability is invalid, adjusting the weight value of the convolution kernel in the second model, returning to execute the step of inputting the characteristic data in the historical characteristic data of the user into the first model, the second model, the third model and the fourth model, and outputting the second prediction probability until the training of the historical characteristic data of the user is finished. And when the second prediction probability is invalid, adjusting a weight value of a convolution kernel set in advance in the second prediction probability, and training the historical characteristic data of the user again. The method of the embodiment provides a training method for the second model based on the historical characteristic data of the user, and the reliability and accuracy of model prediction can be improved by training and fitting the model based on the historical characteristic data of the user.

Considering that there may be many times of historical complaints of a user, it is necessary to determine a complaint value in the historical feature data of the user from the many times of complaints. As an example, fig. 7 illustrates a flowchart of a method for predicting complaints of a user, where on the basis of any example, the method further includes:

step 501, according to the complaint mode, carrying out complaint assignment on the historical complaints of the user;

step 502, if the historical complaint times of the user are one time, assigning the complaint as a complaint true value; if the historical complaint times of the user are larger than one time, calculating a plurality of corresponding complaint assignments and complaint times to obtain complaint true values.

In this example, the user's historical complaints are assigned according to the complaint pattern. For example, complaint modes can be classified into department complaint, group customer service complaint and common complaint, and the three complaints are assigned with values of 1, 0.7 and 0.5, respectively. If the historical complaint times of the user are one time, assigning the complaint value as a complaint true value; if the historical complaint times of the user are larger than one time, calculating a plurality of corresponding complaint assignments and complaint times to obtain complaint true values. For example, when the history complaint of the user has complaint of the letter department 1 times, group customer service complaint 1 times and common complaint 2 times, respectively, the complaint value may be (1×1+0.7×1+0.5×2)/4=0.675. The method provided by the example integrates the complaint times and different complaint types to determine the complaint true value of the user, and improves the authenticity of the complaint true value.

According to the user complaint prediction method provided by the embodiment, the feature data of the user to be predicted, which comprises the sub-feature data representing different dimensions, are obtained, and each group of sub-feature data is sequentially input into a first model to perform hidden layer feature extraction so as to obtain a plurality of initial extraction features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and inputting the matrix into a third model for hidden layer feature extraction to obtain predicted extracted features; and inputting the prediction extraction features into a fourth model for complaint prediction to obtain a complaint prediction result of the user to be predicted. The scheme can effectively predict potential complaint users.

Example two

Fig. 8 schematically illustrates a structural diagram of a user complaint prediction device according to a second embodiment of the present application, where, as shown in fig. 8, the device includes:

an acquisition module 21, configured to acquire feature data of a user to be predicted; the feature data comprises a plurality of groups of sub-feature data representing different dimensions, and each group of sub-feature data comprises a plurality of data assignments representing different attributes;

The extracting module 22 is configured to sequentially input each group of sub-feature data into the first model to perform hidden layer feature extraction, so as to obtain a plurality of initial extracted features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; wherein, at least one hidden layer up-scaling and hidden layer down-scaling are performed on each group of sub-feature data in the first model;

the adjustment module 23 is configured to convert the first extracted feature data into a matrix, input the matrix into the second model, perform convolutional layer feature extraction, and output extracted matrix data; the matrix is subjected to multiple convolution layer feature extraction in the second model, and at least two convolution kernels with different sizes exist in the convolution layer feature extraction;

the prediction module 24 is configured to expand the extracted matrix data into second extracted feature data, and input the second extracted feature data into a third model to perform hidden layer feature extraction, so as to obtain predicted extracted features; inputting the prediction extraction features into a fourth model to perform complaint prediction, obtaining a first prediction probability, and determining a complaint prediction result of the user to be predicted according to the first prediction probability; wherein the fourth model comprises an output layer comprising 1 neuron and a probability function.

L＝H(f6(H(f12(F1))))

where F1 represents a plurality of data assignments of the network data, F12 represents 12 neurons of the hidden layer when performing hidden layer dimension-up, H () represents a Hardtanh activation function, and F6 represents the number of 6 neurons of the hidden layer when performing hidden layer dimension-down. In practical application, the Hardtank activation function can be replaced by any activation function so as to achieve the effect of nonlinear conversion of data.

In the user complaint prediction device of the present example, feature data of a user to be predicted including sub-feature data representing different dimensions is acquired, and each group of sub-feature data is sequentially input into a first model to perform hidden layer feature extraction to obtain a plurality of initial extracted features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and inputting the matrix into a third model for hidden layer feature extraction to obtain predicted extracted features; and inputting the prediction extraction features into a fourth model for complaint prediction to obtain a complaint prediction result of the user to be predicted. The scheme can effectively predict potential complaint users.

In one example, the extraction module 22 is specifically configured to:

sequentially judging whether the total data assignment amount of each group of sub-feature data is smaller than a preset threshold value; if the total data assignment amount of the group of sub-feature data is smaller than a preset threshold value, performing hidden layer lifting and nonlinear conversion and hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features; if the total data assignment amount of the group of sub-feature data is not smaller than a preset threshold value, carrying out multiple hidden layer lifting and nonlinear conversion and multiple hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features.

In this example, the number of times of hidden layer increase and hidden layer dimension decrease of the sub-feature data is determined according to the scale of the total data assignment amount of each group of sub-feature data. In practical application, the preset threshold value can be set to be 10, and if the total data assignment amount in the sub-feature data is smaller than 10, only one hidden layer lifting and hidden layer dimension reduction are carried out; if the total data assignment amount in the sub-feature data is not less than 10, performing primary hidden layer dimension increasing and decreasing, and then performing primary hidden layer dimension increasing and decreasing. Optionally, according to the prediction requirement, multiple thresholds may be set, and different hidden layer rising and hidden layer dimension reduction times are set in different threshold intervals. The device provided by the example carries out multiple hidden layer lifting and hidden layer dimension reduction on the sub-feature data with larger data assignment total size, can separate noise and useful information, and improves the distinguishing degree of samples so as to better extract and select the features.

In one example, the adjustment module 23 is specifically configured to:

converting the first extracted feature data into a maximum matrix according to the feature quantity of the first extracted feature data; performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of n×n on the maximum matrix in a first convolution layer of a second model, and performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m×m to obtain a first extraction matrix; wherein m and n are positive integers, and m is greater than n; performing convolutional layer feature extraction with a convolutional kernel of n multiplied by n on the maximum matrix in a second convolutional layer of the second model to obtain a second extraction matrix; performing matrix multiplication on the first extraction matrix and the second extraction matrix to obtain a third extraction matrix; performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m×m on the maximum matrix in a third convolution layer of the second model, and performing n×n convolution layer feature extraction and nonlinear conversion with the convolution kernel to obtain a fourth extraction matrix; performing matrix multiplication on the fourth extraction matrix and the second extraction matrix to obtain a fifth extraction matrix; performing matrix splicing on the third extraction matrix and the fifth extraction matrix to obtain a sixth extraction matrix; and performing convolutional kernel n multiplied by 2 convolutional layer feature extraction and nonlinear conversion on the sixth extraction matrix in a fourth convolutional layer of the second model to obtain a seventh extraction matrix, and taking the seventh extraction matrix as extraction matrix data output by the second model.

In this example, the first extracted feature data is converted into a maximum matrix according to the number of the first extracted feature data. In particular, to facilitate subsequent convolution calculations and reduce feature loss, the first extracted feature data may be converted into a maximum matrix that is less than its number. For example, if the number of features of the first extracted feature data is 30, it is converted into a maximum matrix of 5×5. In practical application, the first extracted feature data can be subjected to dimension reduction and nonlinear conversion through a hidden layer containing 25 neurons, so that the first extracted feature data with 25 feature numbers is obtained and converted into a maximum matrix of 5×5.

wherein G is the maximum matrix,a convolution layer representing a convolution kernel size of 1 x 1 +. >Representing a convolution layer with a convolution kernel size of 3 x 3, H () represents the Hardtanh activation function. It should be noted that, when the maximum matrix of 5×5 is convolved by 1×1, it is still 5×5; and a matrix of 5×5 is convolved by 3×3, a matrix reduction output of 3×3 occurs. Thus, the first and second substrates are bonded together,to avoid shrinking the matrix space for each convolution operation, the surrounding of the 1×1 convolved 5×5 matrix needs to be padded (padding) so that the output after the 3×3 convolution is still 5×5. Wherein the value of the ambient filling is typically 0.

And performing matrix splicing on the third extraction matrix and the fifth extraction matrix to obtain a sixth extraction matrix. Specifically, the matrix stitching is performed in a manner of overlap stitching, for example, the third extraction matrix of 5×5 and the fifth extraction matrix of 5×5 are the sixth extraction matrix of 5×5×2 after overlap stitching. And performing convolutional kernel n multiplied by 2 convolutional layer feature extraction and nonlinear transformation on the sixth extraction matrix in a fourth convolutional layer of the second model to obtain a seventh extraction matrix, and taking the seventh extraction matrix as extraction matrix data output by the second model. According to the device provided by the example, subsequent convolution calculation can be facilitated and feature loss can be reduced by converting the first extraction feature data into the maximum matrix, matrices with different convolution degrees can be obtained through the first extraction matrix and the fourth extraction matrix, feature supplementation can be performed on the first extraction matrix and the fourth extraction matrix through the second extraction matrix, the third extraction matrix and the fifth extraction matrix can be spliced through the sixth extraction matrix, and the sixth extraction matrix can be fused into a single-layer matrix through the seventh extraction matrix.

In one example, prediction module 24 is specifically configured to:

and carrying out hidden layer lifting and nonlinear conversion on the second extracted feature data through a third model, and carrying out hidden layer reducing and nonlinear conversion to obtain predicted extracted features.

In practical application, the prediction extraction characteristics can be obtained through a fully connected network. Specifically, the second extracted feature data is fully connected through a third model based on the feature quantity in the second extracted feature data. For example, if the second extracted feature data is 25, the second extracted feature data is subjected to hidden layer up-conversion and nonlinear conversion including x neurons, and to hidden layer down-conversion and nonlinear conversion including 25 neurons, to obtain the predicted extracted feature. Wherein x is an integer greater than 25 and less than 50 based on the maximum neuron restriction in the ascending dimension. The device provided by the example can perform feature extraction on the convolved second extracted feature data through the third model.

In one example, the acquisition module 21 is further configured to:

acquiring historical characteristic data of a plurality of users, wherein the historical characteristic data comprises characteristic data and complaint true values of the plurality of users; inputting the characteristic data in the historical characteristic data of each user into the first model, the second model, the third model and the fourth model in sequence, and outputting a second prediction probability; judging whether the second prediction probability is valid or not according to the complaint true value in the historical characteristic data of the user; if the second prediction probability is valid, finishing training of the historical feature data of the user; and if the second prediction probability is invalid, adjusting the weight value of the convolution kernel in the second model, returning to execute the step of inputting the characteristic data in the historical characteristic data of the user into the first model, the second model, the third model and the fourth model, and outputting the second prediction probability until the training of the historical characteristic data of the user is finished.

In this example, historical feature data for a plurality of users is obtained, wherein the historical feature data includes feature data and complaint truth values for the plurality of users. In practical application, the complaint true value is the training direction of the model. And sequentially inputting the characteristic data in the historical characteristic data of each user into the first model, the second model, the third model and the fourth model, and outputting the second prediction probability. In particular, the weight values of the plurality of convolution kernels in the second model prior to non-passing through the user's historical feature data input may be random. For example, in a random case, the weight of the convolution kernel of 1×1 may be 2, and the weight of the convolution kernel of 3×3 may be any 9 values. Judging whether the second prediction probability is valid or not according to the complaint true value in the historical characteristic data of the user; and if the second prediction probability is valid, finishing training of the historical characteristic data of the user. And if the second prediction probability is invalid, adjusting the weight value of the convolution kernel in the second model, returning to execute the step of inputting the characteristic data in the historical characteristic data of the user into the first model, the second model, the third model and the fourth model, and outputting the second prediction probability until the training of the historical characteristic data of the user is finished. And when the second prediction probability is invalid, adjusting a weight value of a convolution kernel set in advance in the second prediction probability, and training the historical characteristic data of the user again. The device of the embodiment provides a training method for the second model based on the historical characteristic data of the user, and the reliability and accuracy of model prediction can be improved by training and fitting the model based on the historical characteristic data of the user.

In one example, the acquisition module 21 is further configured to:

according to the complaint mode, carrying out complaint assignment on the historical complaints of the user; if the historical complaint times of the user are one time, assigning the complaint value as a complaint true value; if the historical complaint times of the user are larger than one time, calculating a plurality of corresponding complaint assignments and complaint times to obtain complaint true values.

In this example, the user's historical complaints are assigned according to the complaint pattern. For example, complaint modes can be classified into department complaint, group customer service complaint and common complaint, and the three complaints are assigned with values of 1, 0.7 and 0.5, respectively. If the historical complaint times of the user are one time, assigning the complaint value as a complaint true value; if the historical complaint times of the user are larger than one time, calculating a plurality of corresponding complaint assignments and complaint times to obtain complaint true values. For example, when the history complaint of the user has complaint of the letter department 1 times, group customer service complaint 1 times and common complaint 2 times, respectively, the complaint value may be (1×1+0.7×1+0.5×2)/4=0.675. The device provided by the example integrates the complaint times and different complaint types to determine the complaint true value of the user, and improves the authenticity of the complaint true value.

According to the user complaint prediction device provided by the embodiment, the feature data of the user to be predicted, which comprises the sub-feature data representing different dimensions, are obtained, and each group of sub-feature data is sequentially input into the first model to perform hidden layer feature extraction so as to obtain a plurality of initial extraction features; performing feature stitching on a plurality of initial extracted features to obtain first extracted feature data; converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and inputting the matrix into a third model for hidden layer feature extraction to obtain predicted extracted features; and inputting the prediction extraction features into a fourth model for complaint prediction to obtain a complaint prediction result of the user to be predicted. The scheme can effectively predict potential complaint users.

Example III

Fig. 9 is a schematic structural diagram of an electronic device according to a third embodiment of the present application, where the electronic device includes:

a processor 291, the electronic device further comprising a memory 292; a communication interface (Communication Interface) 293 and bus 294 may also be included. The processor 291, the memory 292, and the communication interface 293 may communicate with each other via the bus 294. Communication interface 293 may be used for information transfer. The processor 291 may invoke logic instructions in the memory 292 to perform the methods of the examples described above.

Further, the logic instructions in memory 292 described above may be implemented in the form of software functional units and stored in a computer-readable storage medium when sold or used as a stand-alone product.

The memory 292 is a computer readable storage medium, and may be used to store a software program, a computer executable program, and program instructions/modules corresponding to the methods in the embodiments of the present application. The processor 291 executes functional applications and data processing by running software programs, instructions and modules stored in the memory 292, i.e., implements the methods in the method examples described above.

Memory 292 may include a storage program area that may store an operating system, at least one application program required for functionality, and a storage data area; the storage data area may store data created according to the use of the terminal device, etc. Further, memory 292 may include high-speed random access memory, and may also include non-volatile memory.

Embodiments of the present application also provide a computer-readable storage medium having stored therein computer-executable instructions that, when executed by a processor, are configured to implement the method of any of the embodiments.

Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the application following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice within the art to which the application pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the application being indicated by the following claims.

It is to be understood that the present application is not limited to the precise arrangements and instrumentalities shown in the drawings, which have been described above, and that various modifications and changes may be effected without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims

1. A method of predicting complaints of a user, comprising:

acquiring characteristic data of a user to be predicted; wherein the feature data comprises a plurality of sets of sub-feature data representing different dimensions, each set of sub-feature data comprising a plurality of data assignments representing different attributes;

inputting each group of sub-feature data into a first model in sequence to extract hidden layer features, and obtaining a plurality of initial extracted features; performing feature stitching on the plurality of initial extracted features to obtain first extracted feature data; wherein, at least one hidden layer up-scaling and hidden layer down-scaling are performed on each group of sub-feature data in the first model;

Converting the first extracted feature data into a matrix, inputting the matrix into a second model for convolutional layer feature extraction, and outputting extracted matrix data; the matrix is subjected to multiple convolution layer feature extraction in the second model, and at least two convolution kernels with different sizes exist in the convolution layer feature extraction;

expanding the extraction matrix data into second extraction feature data, and inputting the second extraction feature data into a third model for extracting hidden layer features to obtain prediction extraction features; inputting the prediction extraction features into a fourth model to perform complaint prediction, obtaining a first prediction probability, and determining a complaint prediction result of the user to be predicted according to the first prediction probability; wherein the fourth model comprises an output layer comprising 1 neuron and a probability function.

2. The method of claim 1, wherein sequentially inputting each set of sub-feature data into the first model for hidden layer feature extraction to obtain a plurality of initial extracted features, comprises:

sequentially judging whether the total data assignment amount of each group of sub-feature data is smaller than a preset threshold value; if the total data assignment amount of the group of sub-feature data is smaller than a preset threshold value, performing hidden layer lifting and nonlinear conversion and hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features;

If the total data assignment amount of the group of sub-feature data is not smaller than a preset threshold value, carrying out multiple hidden layer lifting and nonlinear conversion and multiple hidden layer reducing and nonlinear conversion on the group of sub-feature data to obtain corresponding initial extraction features.

3. The method of claim 1, wherein converting the first extracted feature data into a matrix and inputting the matrix into a second model for convolutional layer feature extraction to obtain extracted matrix data, comprising:

converting the first extracted feature data into a maximum matrix according to the feature quantity of the first extracted feature data;

performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of n×n on the maximum matrix in a first convolution layer of the second model, and performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m×m to obtain a first extraction matrix; wherein m and n are positive integers, and m is greater than n;

performing convolutional layer feature extraction with a convolutional kernel of n multiplied by n on the maximum matrix in a second convolutional layer of the second model to obtain a second extraction matrix; performing matrix multiplication on the first extraction matrix and the second extraction matrix to obtain a third extraction matrix;

Performing convolution layer feature extraction and nonlinear conversion with a convolution kernel of m×m on the maximum matrix in a third convolution layer of the second model, and performing n×n convolution layer feature extraction and nonlinear conversion with the convolution kernel to obtain a fourth extraction matrix; performing matrix multiplication on the fourth extraction matrix and the second extraction matrix to obtain a fifth extraction matrix;

performing matrix splicing on the third extraction matrix and the fifth extraction matrix to obtain a sixth extraction matrix; and performing convolutional kernel n multiplied by 2 convolutional layer feature extraction and nonlinear conversion on the sixth extraction matrix in a fourth convolutional layer of the second model to obtain a seventh extraction matrix, and taking the seventh extraction matrix as extraction matrix data output by the second model.

4. The method of claim 1, wherein the third model is for:

and carrying out hidden layer ascending and nonlinear conversion and hidden layer descending and nonlinear conversion on the second extracted feature data to obtain predicted extracted features.

5. The method according to any one of claims 1 to 4, further comprising:

acquiring historical characteristic data of a plurality of users, wherein the historical characteristic data comprises characteristic data and complaint truth values of the plurality of users;

Inputting the characteristic data in the historical characteristic data of each user into the first model, the second model, the third model and the fourth model in sequence, and outputting a second prediction probability;

judging whether the second prediction probability is valid or not according to a complaint true value in the historical feature data of the user; if the second prediction probability is valid, finishing training of the historical feature data of the user;

and if the second prediction probability is invalid, adjusting a weight value of a convolution kernel in the second model, and returning to execute the step of inputting the characteristic data in the historical characteristic data of the user into the first model, the second model, the third model and the fourth model to output the second prediction probability until training of the historical characteristic data of the user is finished.

6. The method of claim 5, wherein the method further comprises:

according to the complaint mode, carrying out complaint assignment on the historical complaints of the user;

if the historical complaint times of the user are one time, assigning the complaint value as the complaint true value; if the historical complaint times of the user are larger than one time, calculating a plurality of corresponding complaint assignments and complaint times to obtain the complaint true values.

7. A user complaint predicting apparatus, comprising:

the acquisition module is used for acquiring the characteristic data of the user to be predicted; wherein the feature data comprises a plurality of sets of sub-feature data representing different dimensions, each set of sub-feature data comprising a plurality of data assignments representing different attributes;

the extraction module is used for sequentially inputting each group of sub-feature data into the first model to extract hidden layer features and obtain a plurality of initial extraction features; performing feature stitching on the plurality of initial extracted features to obtain first extracted feature data; wherein, at least one hidden layer up-scaling and hidden layer down-scaling are performed on each group of sub-feature data in the first model;

the adjustment module is used for converting the first extracted characteristic data into a matrix, inputting the matrix into a second model for convolutional layer characteristic extraction, and outputting extracted matrix data; the matrix is subjected to multiple convolution layer feature extraction in the second model, and at least two convolution kernels with different sizes exist in the convolution layer feature extraction;

the prediction module is used for expanding the extraction matrix data into second extraction feature data and inputting the second extraction feature data into a third model for extracting hidden layer features to obtain prediction extraction features; inputting the prediction extraction features into a fourth model to perform complaint prediction, obtaining a first prediction probability, and determining a complaint prediction result of the user to be predicted according to the first prediction probability; wherein the fourth model comprises an output layer comprising 1 neuron and a probability function.

8. The apparatus of claim 7, wherein the extraction module is configured to:

9. An electronic device, comprising: a processor, and a memory communicatively coupled to the processor; the memory stores computer-executable instructions; the processor executes computer-executable instructions stored in the memory to implement the method of any one of claims 1 to 6.

10. A computer readable storage medium having stored therein computer executable instructions which when executed by a processor are adapted to carry out the method of any one of claims 1 to 6.