Method and apparatus for classifying user, device, and storage medium

Abstract

The present application belongs to the field of data classification, and discloses a method and apparatus for classifying a user, a device, and a storage medium. The method comprises: acquiring structured data, image data, and text data of a user; inputting the structured data, the image data, and the text data of the user into three models, respectively, to obtain three outputs outputted by the three models; and then, on the basis of the three outputs, obtaining a target probability of the user belonging to each user category within a preset plurality of user categories. Since the target probability is a result obtained by integrating a plurality of different types of data, the influence of fluctuations in any single type of data on the target probability can be reduced, so that the effect of improving the stability of the method for classifying a user can be achieved.

Description

User classification method, device, equipment and storage medium Technical Field

[0001] The present application relates to the field of data classification, and in particular to a user classification method, apparatus, device and storage medium. Background Art

[0002] The user classification method is a method for classifying users based on various parameters of the users.

[0003] In a user classification method, structured data such as the user's height, weight, gender, occupation, and hobbies are first obtained, and then the structured data is input into a trained neural network model. The neural network model then outputs the probability that the user belongs to various categories.

[0004] However, the probability determined by the above user classification method may be greatly affected by some data fluctuations, resulting in poor stability of the classification method.

[0005] Summary of the Invention

[0006] The present application provides a user classification method, apparatus, device, and storage medium. The technical solution is as follows:

[0007] According to a first aspect of the present application, a method for classifying users is provided, the method comprising:

[0008] Obtain user's structured data, image data and text data;

[0009] Inputting the structured data into a first model to obtain a first output of the first model, wherein the first model is configured to output a first output including a probability that the user belongs to each of a plurality of preset user categories based on the structured data;

[0010] inputting the image data into a second model to obtain a second output of the second model, wherein the second model is configured to output a second output including a probability that the user belongs to each of the plurality of preset user categories based on the image data;

[0011] Inputting the text data into a third model to obtain a third output of the third model, wherein the third model is configured to output a third output including a probability that the user belongs to each of the plurality of preset user categories based on the text data;

[0012] Determining a target probability that the user belongs to each of the preset multiple categories of users based on the first output, the second output, and the third output;

[0013] At least one of the first model, the second model and the third model includes a target neural network having small-world characteristics and power-law characteristics.

[0014] Optionally, the target neural network with small-world characteristics and power-law characteristics includes an input layer, multiple hidden layers and an output layer, the multiple hidden layers include at least two target hidden layers, the degree distribution of the nodes of the at least two target hidden layers conforms to the power-law distribution, and the other layers in the target neural network except the at least two target hidden layers conform to the small-world characteristics.

[0015] Optionally, the first model includes a first input layer, a structured data encoder, a first output layer, a fully connected layer and a first normalization module, and the first input layer, the structured data encoder and the first output layer constitute the first neural network with small-world characteristics and power-law characteristics.

[0016] Optionally, the second model includes a visual encoder, a second neural network, and a second normalization module;

[0017] The second neural network is the target neural network having small-world characteristics and power-law characteristics.

[0018] Optionally, the third model includes a text encoder, a third neural network and a third normalization module;

[0019] The third neural network is the target neural network having small-world characteristics and power-law characteristics.

[0020] Optionally, the method further includes:

[0021] Obtaining a variational autoencoder network to be trained, a second model to be trained, and a third model to be trained, wherein the variational autoencoder network includes the first input layer, a structured data encoder to be trained, the first output layer, a structured data decoder, and a decoder output layer; the second model to be trained includes a visual encoder, a second neural network to be trained, and a second normalization module; and the third model to be trained includes a text encoder, a third neural network to be trained, and a third normalization module;

[0022] Acquire a target training sample from the target sample set, where the target training sample includes a structured data sample, an image data sample, and a text data sample of a sample user;

[0023] Inputting the structured data sample into the variational autoencoder network to be trained, inputting the image data sample into the second model to be trained, and inputting the text data sample into the third model to be trained;

[0024] Obtaining a first training output of the variational autoencoder network to be trained, a second training output of the second model to be trained, and a third training output of the third model to be trained;

[0025] Obtaining a comprehensive training output based on the first training output, the second training output, and the third training output, wherein the comprehensive training output includes a target probability that the sample user belongs to each of the preset multiple user categories;

[0026] Adjusting parameters of the second neural network to be trained in the second model based on the training synthesis output, adjusting parameters of the third neural network to be trained in the third model based on the training synthesis output, and adjusting parameters of the structured data encoder to be trained based on the training synthesis output and the output of the decoder output layer;

[0027] When the preset training cutoff condition is not met, executing the step of obtaining a target training sample from the target sample set;

[0028] When a preset training cutoff condition is reached, the structured data encoder to be trained is determined as the structured data encoder, the second neural network to be trained is determined as the second neural network, and the third neural network to be trained is determined as the third neural network.

[0029] Optionally, the method further comprises: obtaining the loss of the variational autoencoder network to be trained and the loss of the training integrated output;

[0030] Determining a loss value for this training based on the loss of the variational autoencoder network to be trained and the loss of the comprehensive output of the training;

[0031] When the loss value of the current training is greater than or equal to the target loss value, it is determined that the training end condition is met, and the target loss value is the average of the loss values ​​of the target number of trainings before the current training.

[0032] Optionally, the method further includes: determining whether the preset training cutoff condition is reached based on a loss function, wherein the loss function includes: L loss =L1+αL2;

[0033] Wherein, α is the loss function adjustment weight, 0<α<1, L1 is the first loss of the variational autoencoder network to be trained, L2 is the loss of the comprehensive output of the training, L lossis the value of the loss function, N is the number of samples in the target sample set, n is the number of features output by the structured data encoder to be trained, X is the input of the structured data encoder to be trained, X' is the output of the structured data decoder, and x i is the i-th eigenvalue of X, x′ i is the i-th eigenvalue of X′; M is the number of user categories of the multi-category user. When the true category of sample j is equal to c, y jc is 1, the target training sample also includes the true category to which the sample user belongs. When the true category of the sample user j is not equal to c, y jc is 0, p jc is the probability that the sample user j belongs to category c in the target probability.

[0034] Optionally, the first model, the second model and the third model all include a target neural network having small-world characteristics and power-law characteristics.

[0035] Optionally, the target neural network included in the first model, the second model and the third model has any one of the small-world properties of Watt-Strogatz small-world property and the Newman-Watt small-world property.

[0036] Optionally, the structured data includes data of multiple categories of the user, the categories including at least age, height, weight, and multiple health parameters;

[0037] The image data includes medical imaging data of the user;

[0038] The text data includes text data in the user's electronic medical record.

[0039] According to another aspect of the present application, a user classification device is provided, the user classification device comprising:

[0040] The acquisition module is used to obtain the user's structured data, image data and text data;

[0041] a first input module, configured to input the structured data into a first model to obtain a first output of the first model, wherein the first model is configured to output a first output including a probability that the user belongs to each of multiple preset user categories based on the structured data;

[0042] a second input module, configured to input the image data into a second model to obtain a second output of the second model, wherein the second model is configured to output, based on the image data, a second output including a probability that the user belongs to each of the plurality of preset user categories;

[0043] a third input module, configured to input the text data into a third model to obtain a third output of the third model, wherein the third model is configured to output the third output including a probability that the user belongs to each of the plurality of preset user categories based on the text data;

[0044] a determination module, configured to determine a target probability that the user belongs to each of the preset multiple categories of users based on the first output, the second output, and the third output;

[0045] At least one of the first model, the second model and the third model includes a target neural network having small-world characteristics and power-law characteristics.

[0046] Optionally, the target neural network with small-world characteristics and power-law characteristics includes an input layer, multiple hidden layers and an output layer, the multiple hidden layers include at least two target hidden layers, the degree distribution of the nodes of the at least two target hidden layers conforms to the power-law distribution, and the other layers in the target neural network except the at least two target hidden layers conform to the small-world characteristics.

[0047] Optionally, the user classification device further includes:

[0048] A model acquisition module, configured to acquire a variational autoencoder network to be trained, a second model to be trained, and a third model to be trained, wherein the variational autoencoder network includes the first input layer, a structured data encoder to be trained, the first output layer, a structured data decoder, and a decoder output layer; the second model includes a visual encoder, a second neural network to be trained, and a second normalization module; and the third model to be trained includes a text encoder, a third neural network to be trained, and a third normalization module;

[0049] A sample acquisition module is used to acquire a target training sample from a target sample set, wherein the target training sample includes a structured data sample, an image data sample, and a text data sample of a sample user;

[0050] A training input module, configured to input the structured data sample into the variational autoencoder network to be trained, input the image data sample into the second model to be trained, and input the text data sample into the third model to be trained;

[0051] An output acquisition module, configured to acquire a first training output of the variational autoencoder network to be trained, a second training output of the second model to be trained, and a third training output of the third model to be trained;

[0052] a comprehensive output module, configured to obtain a comprehensive training output based on the first training output, the second training output, and the third training output, wherein the comprehensive training output includes a target probability that the sample user belongs to each of the plurality of preset user categories;

[0053] an adjustment module, configured to adjust parameters of the second neural network to be trained in the second model based on the training synthesis output, adjust parameters of the third neural network to be trained in the third model based on the training synthesis output, and adjust parameters of the structured data encoder to be trained based on the training synthesis output and an output of the decoder output layer;

[0054] A repeated training module, configured to execute the step of obtaining a target training sample from the target sample set when a preset training cutoff condition is not met;

[0055] A training cutoff module is used to determine the structured data encoder to be trained as the structured data encoder, determine the second neural network to be trained as the second neural network, and determine the third neural network to be trained as the third neural network when a preset training cutoff condition is reached.

[0056] According to another aspect of the present application, a user classification device is provided, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the user classification method as described above.

[0057] According to another aspect of the present application, a computer storage medium is provided, in which at least one instruction, at least one program, a code set or an instruction set is stored. The at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement the user classification method as described above.

[0058] The beneficial effects of the technical solutions provided in the embodiments of the present application include at least:

[0059] By inputting the user's structured data, image data and text data into three models respectively to obtain the three outputs of the three models, the target probability of the user belonging to each of the preset multiple categories of users can be obtained based on these three outputs. Since the target probability is the result obtained by integrating multiple different types of data, the impact of the target probability on fluctuations of a single type of data can be reduced, thereby achieving the effect of improving the stability of the user's classification method.

[0060] In addition, at least one of the first model, the second model and the third model includes a target neural network with small-world characteristics and power-law characteristics, which can improve the data transmission efficiency of the neural network and reduce the number of parameters in the neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0062] FIG1 is a schematic diagram of an implementation environment provided by an embodiment of the present application;

[0063] FIG2 is a flow chart of a method for classifying users shown in an embodiment of the present application;

[0064] FIG3 is a flow chart of a method for classifying users shown in an embodiment of the present application;

[0065] FIG4 is a schematic structural diagram of a model provided in an embodiment of the present application;

[0066] FIG5 is a schematic diagram of the structure of a target neural network in an embodiment of the present application;

[0067] FIG6 is a structural block diagram of a user classification device provided in an embodiment of the present application;

[0068] FIG7 is a structural block diagram of another user classification device provided in an embodiment of the present application.

[0069] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0070] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0071] 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application, which may include a server 11 and multiple terminals 12. The server 11 may establish wired or wireless connections with the multiple terminals 12.

[0072] The terminal 12 can provide the server 11 with various user data, such as structured data, image data, and text data. The terminal 12 can be any electronic product that can interact with the user through one or more methods such as a keyboard, touchpad, touch screen, remote control, voice interaction, or handwriting device, such as a personal computer (PC), mobile phone, smart phone, personal digital assistant (PDA), wearable device, PDA, tablet computer, smart TV, smart speaker, etc.

[0073] The server 11 can obtain various user data from the terminal 12 and determine the category to which the user belongs based on this data. The server 11 can be a standalone server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, or a cloud computing service center.

[0074] FIG2 is a flow chart of a user classification method according to an embodiment of the present application. The user classification method can be applied to the server in the implementation environment shown in FIG1 , and the user classification method may include the following steps:

[0075] Step 201: Acquire the user's structured data, image data, and text data.

[0076] Step 202: Input the structured data into the first model to obtain a first output of the first model, where the first model is configured to output a first output including a probability that the user belongs to each of multiple preset user categories based on the structured data.

[0077] Step 203: Input the image data into the second model to obtain a second output of the second model, where the second model is configured to output a second output including a probability that the user belongs to each of multiple preset user categories based on the image data.

[0078] Step 204: Input the text data into the third model to obtain a third output of the third model, where the third model is configured to output a third output including a probability that the user belongs to each of multiple preset user categories based on the text data.

[0079] Step 205: Determine a target probability that the user belongs to each of the preset multiple user categories based on the first output, the second output, and the third output.

[0080] At least one of the first model, the second model, and the third model includes a target neural network having small-world characteristics and power-law characteristics.

[0081] To sum up, the user classification method provided in the embodiment of the present application inputs the user's structured data, image data and text data into three models respectively to obtain three outputs of the three models. Then, based on these three outputs, the target probability of the user belonging to each of the preset multiple categories of users can be obtained. Since the target probability is the result obtained by integrating multiple different types of data, the influence of the target probability on the fluctuation of a single type of data can be reduced, thereby achieving the effect of improving the stability of the user classification method.

[0082] In addition, at least one of the first model, the second model and the third model includes a target neural network with small-world characteristics and power-law characteristics, which can improve the data transmission efficiency of the neural network and reduce the number of parameters in the neural network.

[0083] FIG3 is a flow chart of a user classification method according to an embodiment of the present application. The user classification method can be applied to the server in the implementation environment shown in FIG1 , and the user classification method may include the following steps:

[0084] Step 301: Obtain a variational autoencoder network to be trained, a second model to be trained, and a third model to be trained.

[0085] When applying the user classification method provided in the embodiment of the present application, the server can first obtain the variational autoencoder network to be trained, the second model to be trained, and the third model to be trained, and train the first model to be trained, the second model to be trained, and the third model to be trained in a variety of ways to obtain the first model, the second model, and the third model. In the embodiment of the present application, these three models can be trained simultaneously.

[0086] Please refer to Figure 4, which is a structural diagram of a model provided in an embodiment of the present application, wherein the variational autoencoder network a includes a first input layer a1, a structured data encoder to be trained a2, a first output layer a3, a structured data decoder a4 and a decoder output layer a5.

[0087] In addition, the neural network structure also includes a fully connected layer a6 and a normalization model a7. The first input layer a1, the structured data encoder a2 to be trained, the first output layer a3, the fully connected layer a6, and the normalization model a7 can constitute the first model a to be trained. The first input layer a1 can be used to receive data from the outside, which is usually raw data or preprocessed data. In the embodiment of the present application, the data can be the user's structured data. The first input layer a1 can convert the received data into a format that can be processed by the neural network.

[0088] The structured data encoder a2 to be trained can be used to perform feature extraction on the data provided by the first input layer a1. The first output layer a3 is used to provide the data output by the structured data encoder a2 to the structured data decoder a4 and the fully connected layer a6. The structured data decoder a4 can decode the data provided by the first output layer a3 and input the decoded data into the decoder output layer a5. In addition, the fully connected layer a6 can process the received data and input it into the normalization module a7. The normalization module a7 can output the probability that the user belongs to each of the preset multiple user categories. For example, the preset multiple user categories can include categories 1, 2, 3, 4, 5, and 6. The normalization module a7 can output 0.1, 0.3, 0.2, 0.7, 0.6, and 0.2, which can be the probabilities that the user belongs to categories 1, 2, 3, 4, 5, and 6, respectively. The actual meaning of these categories 1, 2, 3, 4, 5, and 6 can be determined by the training data. The normalization module a7 may include a normalization function, such as a Sigmoid function.

[0089] The second model b to be trained includes a visual encoder b1, a second neural network to be trained b2, and a second normalization module b3, wherein the visual encoder b1 can include various visual encoders, such as a vision transformer (VIT). The visual encoder b1 can be used to extract features from image data to obtain image features. The second neural network to be trained b2 can process the image features and input the processed data into the second normalization module b3. Similarly, the second normalization module b3 can output the probability that the user belongs to each of the preset multiple categories of users.

[0090] The third model c to be trained includes a text encoder c1, a third neural network c2 to be trained, and a third normalization module c3. The text encoder c1 can include various text encoders, such as a text translator (Transformer). The text encoder c1 can be used to extract features from text data to obtain text features. The second neural network b2 to be trained can process the text features and input the processed data into the third normalization module c3. Similarly, the third normalization module c3 can output the probability that the user belongs to each of the multiple preset user categories.

[0091] In addition, the model shown in Figure 4 can also include a comprehensive determination module D, which is used to determine the target probability of the user belonging to each of the preset multiple categories of users based on the outputs of the first normalization module a7, the second normalization module b3 and the third normalization module c3.

[0092] The model shown in FIG. 4 may be a comprehensive model, and the comprehensive model may be deployed in a server.

[0093] Step 302: Obtain a target training sample from the target sample set. The target training sample includes a structured data sample, an image data sample, and a text data sample.

[0094] The server may obtain a target training sample in the target sample set. For example, the server may obtain a target training sample in the target sample set from the terminal.

[0095] The target sample set may include multiple target training samples, each of which may include a sample user's structured data sample, an image data sample, and a text data sample. Furthermore, each target training sample may include a true probability that the sample user belongs to each of multiple preset user categories. For example, the multiple preset user categories include A, B, and C, and the sample user belongs to categories B and C. The target training sample may include a probability of 0 for the sample user belonging to category A, a probability of 1 for the sample user belonging to category B, and a probability of 1 for the sample user belonging to category C. The true probability of the sample user belonging to each of the multiple preset user categories may be considered the true category to which the sample user belongs.

[0096] Structured data is data in rows, and the attributes of each row of data can be different. One type of structured data can include multiple categories of user data, including at least age, height, weight, and multiple health parameters.

[0097] For example, a user's structured data may include:

[0098] Age: xxx

[0099] Height: xxx

[0100] Weight (or BMI): xxx

[0101] Blood sugar: xxx

[0102] Diastolic blood pressure: xxx

[0103] Of course, in the embodiment provided in the embodiment of the present application, the structured data may also include other categories of user data, such as gender, region, favorite songs, favorite games, and favorite books, etc. The embodiment of the present application does not limit this.

[0104] In an exemplary embodiment, the image data includes medical imaging data of the user, such as computed tomography (CT) image data, B-mode ultrasound image data, and magnetic resonance imaging (MRI) image data.

[0105] Of course, in the embodiments provided in the embodiments of the present application, the image data may also include other categories of image data of the user, for example, it may include a front-facing full-body photo, a back-facing full-body photo, a front-facing half-body photo, and a back-facing half-body photo of the user, etc. The embodiments of the present application do not limit this.

[0106] In an exemplary embodiment, the text data includes text data in the user's electronic medical record. The text data may include a text description of the user's examination results by an examiner, or a text description of some conditions of the user by a physician.

[0107] Of course, in the embodiment provided in the embodiment of the present application, the text data may also include some text descriptions of the user himself, text descriptions of the user by the user's related personnel (such as relatives, friends and colleagues), etc. The embodiment of the present application does not limit this.

[0108] Step 303: Input the structured data sample into the variational autoencoder network to be trained, input the image data sample into the second model to be trained, and input the text data sample into the third model to be trained.

[0109] The server can input the structured data sample into the variational autoencoder network to be trained, input the image data sample into the second model to be trained, and input the text data sample into the third model to be trained. Specifically, the structured data sample can be input into the first input layer a1 of the variational autoencoder network to be trained, the image data sample can be input into the second input layer b1 of the second model to be trained, and the text data sample can be input into the third input layer c1 of the third model to be trained.

[0110] Step 304: Obtain a first training output of the variational autoencoder network to be trained, a second training output of the second model to be trained, and a third training output of the third model to be trained.

[0111] The server may obtain the outputs of the three models corresponding to the target training sample input this time, wherein the first training output, the second training output, and the third training output all include the probability that the user belongs to each of the preset multiple categories of users.

[0112] Step 305: Obtain a comprehensive training output based on the first training output, the second training output, and the third training output. The comprehensive training output includes a target probability that the user belongs to each of the preset multiple user categories.

[0113] Based on the comprehensive determination module D shown in FIG4 , the server can generate a comprehensive training output using the first, second, and third training outputs. The comprehensive training output includes a target probability that the user belongs to each of multiple preset user categories. The target probability can be a probability obtained by combining the first, second, and third training outputs.

[0114] The comprehensive determination module D can determine the target probability in a variety of ways. In an exemplary embodiment, for any one of the preset multiple categories of users, the target probability that the user belongs to that category of users can be equal to the average of the probabilities that the user belongs to that category of users in the first training output, the second training output, and the third training output. The average can be an arithmetic mean or a weighted average. When the average is a weighted average, three weights can be set for the first training output, the second training output, and the third training output, respectively, and the corresponding weighted average can be calculated.

[0115] Exemplarily, the preset multiple categories of users include A, B and C. The first training output may include (0.2, 0.5, 0.7), the second training output may include (0.1, 0.4, 0.9), and the third training output may include (0.4, 0.6, 0.6). In each training output, the first value represents the probability that the user belongs to category A user, the second value represents the probability that the user belongs to category B user, and the third value represents the probability that the user belongs to category C user. In this way, when the arithmetic mean is applied to determine the target probability, the target probability that the user belongs to category A user is (0.2+0.1+0.4) / 3≈0.23, the target probability that the user belongs to category B user is (0.5+0.4+0.6) / 3=0.5, and the target probability that the user belongs to category C user is (0.7+0.9+0.6) / 3≈0.73. Therefore, the target probability output by the server can be expressed as (0.23, 0.5, 0.73). When applying the weighted average to determine the target probability, if the weight of the first training output is 0.9, the weight of the second training output is 1.1, and the weight of the third training output is 0.8, then the target probability that the user belongs to Class A is 0.9*(0.2+0.1+0.4) / 3=0.21, the target probability that the user belongs to Class B is 1.1*(0.5+0.4+0.6) / 3=0.55, and the target probability that the user belongs to Class C is 0.8*(0.7+0.9+0.6) / 3≈0.59. The comprehensive training output output by the server can be expressed as (0.21, 0.55, 0.59).

[0116] In addition, the comprehensive determination module D may also include a neural network model, through which the training comprehensive output may be determined.

[0117] In the method provided in the embodiment of the present application, the preset multiple categories of users may be users who may suffer from multiple categories of diseases, and each category of users may be users who suffer from one category of diseases. For example, the above-mentioned A, B, and C may be users who suffer from three diseases respectively, and the probability that the user belongs to which category of users can be considered as the probability that the user suffers from that category of diseases. For example, category A users may be users with meningioma, category B users may be users with glioma, and category C users may be users with medulloblastoma. On this basis, the server can determine the probability of the user suffering from various diseases through the method provided in the embodiment of the present application, and the server can provide the probability to the user's physician for reference, so that the physician can diagnose the user's disease and conduct subsequent treatment.

[0118] Step 306: Adjust the parameters of the second neural network to be trained in the second model based on the training comprehensive output, adjust the parameters of the third neural network to be trained in the third model based on the training comprehensive output, and adjust the parameters of the structured data encoder to be trained based on the training comprehensive output and the output of the decoder output layer.

[0119] After obtaining the training synthesis output, the server can adjust the parameters of the second neural network to be trained in the second model based on the training synthesis output, adjust the parameters of the third neural network to be trained in the third model based on the training synthesis output, and adjust the parameters of the structured data encoder to be trained based on the training synthesis output and the output of the decoder output layer. The method for adjusting these parameters can be referred to in related technologies and will not be further described in detail in this embodiment of the present application.

[0120] Step 307: Determine whether the preset training end condition is met. If the preset training end condition is not met, execute step 302; if the preset training end condition is met, execute step 308.

[0121] The server can determine whether a preset training deadline condition is reached after this training, and if the preset training deadline condition is not reached, re-acquire a target training sample from the target sample set and perform the next training.

[0122] In an exemplary embodiment, the server may determine whether the training cutoff condition is reached based on a loss function, where the loss function includes: L loss =L1+αL2;

[0123] Among them, α is the loss function adjustment weight, 0<α<1, L1 is the first loss of the variational autoencoder network to be trained, L2 is the loss of the comprehensive output of the training, L loss is the value of the loss function, N is the number of samples in the target sample set, n is the number of features output by the structured data encoder to be trained, X is the input of the structured data encoder to be trained, X' is the output of the structured data decoder, and x i is the i-th eigenvalue of X, x′ i is the i-th eigenvalue of X′; M is the number of user categories for multi-category users. When the true category of sample user j is equal to c, y jc When y is 1, the true category of sample user j is not equal to c jc is 0. For example, the true category of the sample user 1 (j=1) corresponding to a target training sample is 1 (the true category can be included in the target training sample), then y 11 =1,y 12 =0,y 13= 0, the true category of the sample user 2 (j = 2) corresponding to the other target training sample is 3, then y 21 =0,y 22 =0,y 23 =1. p jc is the probability that user j belongs to category c in the target probability. α can be learned automatically. For example, the initial value of α can be 1, and the server can update the value of α during training using algorithms such as gradient descent.

[0124] Optionally, after the server adjusts the parameters in step 306, the method may further include:

[0125] 1. Obtain the loss of the variational autoencoder network to be trained and the loss of the training comprehensive output, and determine the loss value of this training based on the loss of the variational autoencoder network to be trained and the loss of the training comprehensive output.

[0126] The server can obtain the loss L1 of the variational autoencoder network and the loss L2 of the training comprehensive output based on the calculation method of L1 and L2 described above. However, the server can also obtain the loss of the variational autoencoder network to be trained and the loss of the training comprehensive output by other related methods, and the embodiments of the present application are not limited to this.

[0127] 2. When the loss value of this training is greater than or equal to the target loss value, it is determined that the training end condition is met.

[0128] The server can determine whether the loss value of this training is greater than or equal to the target loss value, and when the loss value of this training is greater than or equal to the target loss value, determine that the training cutoff condition is met, and when the loss value of this training is less than the target loss value, determine that the training cutoff condition is not met. The target loss value can be a value determined based on the loss value of the target number of trainings before this training (the target number of trainings can be a preset value, such as 5, 10, etc.). For example, the target loss value is the average of the loss values ​​of the target number of trainings before this training. Of course, the server can also determine the target loss value in other ways, and the embodiments of the present application are not limited to this.

[0129] Of course, the training end condition may also include other conditions, such as a preset number of training times, etc. The server may determine whether the current number of training times reaches the preset number of training times after adjusting the parameters in step 306 each time, and determine that the training end condition is met when the preset number of training times is reached, and determine that the training end condition is not met when the preset number of training times is not reached.

[0130] Alternatively, the server may also determine that if the accuracy of the current training is compared with the accuracy of the t2 trainings (t2 can be a preset value) before the current training (for example, it can be compared with the average of the accuracy of the t2 trainings before the current training), and if it does not increase, it can be determined that the training cutoff condition is met; if it increases, it can be determined that the training cutoff condition is not met. The accuracy of each training can be obtained based on the above-mentioned comprehensive training output and the true probability that the sample user included in the target training sample belongs to each of the preset multiple categories of users. For example, in the target training sample, the true probability of the sample user Xiao A belonging to category 1 of the preset multiple categories of users is 1, the true probability of belonging to category 2 is 0, and the true probability of belonging to category 3 is 0. Then, the closer the probability of the sample user Xiao A belonging to category 1, category 2, and category 3 determined in the comprehensive training output is to the true probability, the higher the accuracy.

[0131] Step 308: Determine the structured data encoder to be trained as the structured data encoder, determine the second neural network to be trained as the second neural network, and determine the third neural network to be trained as the third neural network.

[0132] When a preset training cutoff condition is met, the server can determine the structured data encoder to be trained as the structured data encoder, the second neural network to be trained as the second neural network, and the third neural network to be trained as the third neural network. In this way, the trained structured data encoder, the first input layer, the first output layer, the fully connected layer, and the first normalization module can be combined into a first model, the trained second neural network, the visual encoder, and the second normalization module can be combined into a second model, and the trained third neural network, the visual encoder, and the second normalization module can be combined into a third model. In this way, the training of the three models is completed.

[0133] That is, the first model m1 includes a first input layer a1, a structured data encoder a21, a first output layer a3, a fully connected layer a6, and a first normalization module a7.

[0134] Optionally, the second model m2 includes a visual encoder b1, a second neural network b21 and a second normalization module b3.

[0135] Optionally, the third model m3 includes a text encoder c1, a third neural network c21, and a third normalization module c3.

[0136] In an exemplary embodiment, the first input layer a1, the structured data encoder to be trained a2, and the first output layer a3 can constitute a first neural network, and the structured data decoder can also be a neural network. At least one of the first neural network, the second neural network, the third neural network, and the structured data decoder can be a target neural network with small-world and power-law properties. This can improve the data transmission efficiency of the neural network, reduce the number of parameters in the neural network, and reduce the possibility of overfitting. Optionally, the first neural network, the second neural network, the third neural network, and the structured data decoder can all be target neural networks with small-world and power-law properties, which can further improve the data transmission efficiency of the neural network, reduce the number of parameters in the neural network, and reduce the possibility of overfitting. In this case, the first neural network, the second neural network, the third neural network, and the structured data decoder can have any one of the small-world properties of Watts-Strogatz (WS) and Newman-Watts (NW). That is, for any one of the first neural network, the second neural network, the third neural network and the structured data decoder, they can all have the WS small-world characteristic or the NW small-world characteristic, and the small-world characteristics in different neural networks can be the same or different.

[0137] Please refer to Figure 5, which is a structural diagram of a target neural network in an embodiment of the present application, wherein the target neural network with small-world characteristics and power-law characteristics includes an input layer s1, multiple hidden layers s2 and an output layer s3, the multiple hidden layers s2 include at least two target hidden layers s21, the degree distribution of the nodes of the at least two target hidden layers s21 conforms to the power-law distribution, and the other layers in the target neural network except the at least two target hidden layers s21 conform to the small-world characteristics.

[0138] In the user classification method provided in the embodiment of the present application, the neural network can be processed before the training model to make the neural network have small-world characteristics and power-law characteristics. Exemplarily, before step 301, the embodiment of the present application may also include:

[0139] 1. Obtain the initial variational autoencoder network, the initial second model, and the initial third model.

[0140] The server obtains an initial variational autoencoder network, an initial second model, and an initial third model. The structure of the initial variational autoencoder network is similar to that of the variational autoencoder network to be trained, and may include a first input layer, an initial structured data encoder, a first output layer, an initial structured data decoder, and a decoder output layer. The first input layer, the initial structured data encoder, and the first output layer may constitute a neural network to be processed, and the initial structured data decoder may be a neural network to be processed.

[0141] The structure of the initial second model is similar to that of the second model, and may include a visual encoder, an initial second neural network, and a second normalization module, wherein the initial second neural network may be a neural network to be processed.

[0142] The structure of the initial third model is similar to that of the third model, and may include a text encoder, an initial third neural network, and a third normalization module, wherein the initial third neural network may be a neural network to be processed.

[0143] The above-mentioned neural networks to be processed all include an input layer, multiple fully connected layers and an output layer, wherein the multiple fully connected layers include at least two target fully connected layers.

[0144] 2. Process at least two target fully connected layers in at least one of the neural networks to be processed in the initial variational autoencoder network, the initial second model, and the initial third model so that the degree distribution in the at least two target fully connected layers obeys a power-law distribution.

[0145] The server may process at least two target fully connected layers in at least one of the neural networks to be processed in the initial variational autoencoder network, the initial second model, and the initial third model.

[0146] At least two target fully-connected layers in at least one of the neural networks to be processed, among the initial variational autoencoder network, the initial second model, and the initial third model, are processed so that the degree distributions in the at least two target fully-connected layers obey a power-law distribution. After the degree distributions in the at least two target fully-connected layers obey the power-law distribution, the network formed by the at least two target fully-connected layers is transformed into a scale-free network.

[0147] 3. Processing the other layers except at least two target fully connected layers in the multiple fully connected layers in the processed neural network so that the other layers have small-world characteristics.

[0148] In addition, the server may process other layers of the multiple fully connected layers in the neural network to be processed except for the at least two target fully connected layers (the other layers except for the at least two target fully connected layers may be referred to as the network to be processed) so that the network to be processed has a small-world characteristic. The small-world characteristic may be a WS small-world characteristic or an NW small-world characteristic. When the network to be processed has the WS small-world characteristic, the server may perform random reconnection processing on the network to be processed. The random reconnection processing may include:

[0149] Randomly reconnect each edge in the network to be processed with probability p (probability p is a preset empirical value). Specifically, for each edge, one endpoint of the edge can remain unchanged, while the other endpoint is connected to another node in the network to be processed. Any two different nodes can have at most one edge between them, and no node can have an edge connected to itself. The server can continue this process until the maximum number of reconnections is reached, which can be equal to the product of the number of nodes in the input layer and the number of nodes in the output layer.

[0150] When the network to be processed is made to have the NW small-world characteristic, the server may perform random edge addition processing on the network to be processed. The random edge addition processing may include:

[0151] In the network to be processed, a pair of nodes are randomly selected with probability p (probability p is a preset empirical value) and an edge is added between them. There can be at most one edge between any two different nodes, and each node cannot have an edge connected to itself. This process continues until the maximum number of added edges is reached, which can be equal to the product of the number of nodes in the input layer and the number of nodes in the output layer.

[0152] The initial variational autoencoder network after the above processing 2 and 3 will be transformed into the above variational autoencoder network to be trained, the initial second model after the above processing 2 and 3 will be transformed into the above second model to be trained, and the initial third model after the above processing 2 and 3 will be transformed into the above third model to be trained.

[0153] The above steps are steps for training the model, and the server can determine the user classification based on the trained model.

[0154] Step 309: Acquire the user's structured data, image data, and text data.

[0155] The server may obtain structured data, image data, and text data of the user to be classified. These structured data, image data, and text data may be provided by a terminal connected to the server.

[0156] Step 310: Input the structured data into the first model to obtain a first output of the first model.

[0157] The server may input the structured data into the first model to obtain a first output of the first model, the first output including the probability that the user belongs to each of the multiple pre-set user categories. The structured data is in rows, and each row of data may have different attributes. One type of structured data may include multiple categories of user data, including at least age, height, weight, and multiple health parameters.

[0158] Step 311: Input the image data into the second model to obtain a second output of the second model.

[0159] The server may input the above image data into a second model and obtain a second output, where the second output includes the probability that the user belongs to each of the preset multiple categories of users.

[0160] The image data may include medical imaging data of the user, such as computed tomography image data, B-mode ultrasound image data, and magnetic resonance imaging image data.

[0161] Of course, in the embodiments provided in the embodiments of the present application, the image data may also include other categories of image data of the user, for example, it may include a front-facing full-body photo, a back-facing full-body photo, a front-facing half-body photo, and a back-facing half-body photo of the user, etc. The embodiments of the present application do not limit this.

[0162] Step 312: Input the text data into the third model to obtain a third output of the third model.

[0163] The server may input the above text data into a third model and obtain a third output, where the third output includes the probability that the user belongs to each of the preset multiple categories of users.

[0164] Step 313: Determine a target probability that the user belongs to each of the preset multiple user categories based on the first output, the second output, and the third output.

[0165] After obtaining the probabilities of the user belonging to each of the preset multiple user categories output by the three models, the server can determine the target probability of the user belonging to each of the preset multiple user categories based on these probabilities. The method for determining the target probability can be similar to the method for determining the target probability in step 305. That is, for any one of the preset multiple user categories, the server can determine the target probability of the user belonging to that category as the average of the probabilities of the user belonging to that category in the first output, the second output, and the third output. This average can be an arithmetic mean or a weighted average. If the average is a weighted average, three weights can be set for each of the first output, the second output, and the third output, and the corresponding weighted average can be calculated.

[0166] To sum up, the user classification method provided in the embodiment of the present application inputs the user's structured data, image data and text data into three models respectively to obtain three outputs of the three models. Then, based on these three outputs, the target probability of the user belonging to each of the preset multiple categories of users can be obtained. Since the target probability is the result obtained by integrating multiple different types of data, the influence of the target probability on the fluctuation of a single type of data can be reduced, thereby achieving the effect of improving the stability of the user classification method.

[0167] In addition, at least one of the first model, the second model and the third model includes a target neural network with small-world characteristics and power-law characteristics, which can improve the data transmission efficiency of the neural network and reduce the number of parameters in the neural network.

[0168] The following are embodiments of the apparatus disclosed herein, which can be used to implement the method embodiments disclosed herein. For details not disclosed in the apparatus embodiments disclosed herein, please refer to the method embodiments disclosed herein.

[0169] FIG6 is a structural block diagram of a user classification device provided in an embodiment of the present application. The user classification device 600 can be combined with the server shown in FIG1 . The user classification device 600 includes:

[0170] An acquisition module 610 is used to acquire the user's structured data, image data, and text data;

[0171] A first input module 620 is configured to input structured data into a first model to obtain a first output of the first model, wherein the first model is configured to output a first output including a probability that the user belongs to each of multiple preset user categories based on the structured data;

[0172] A second input module 630 is configured to input the image data into a second model, and the second model is configured to output a second output including a probability that the user belongs to each of a plurality of preset user categories based on the image data;

[0173] A third input module 640 is configured to input text data into a third model, and the third model is configured to output a third output including a probability that the user belongs to each of a plurality of preset user categories based on the text data;

[0174] A determination module 650 is configured to determine a target probability that the user belongs to each of the plurality of preset user categories based on the first output, the second output, and the third output;

[0175] At least one of the first model, the second model, and the third model includes a target neural network having small-world characteristics and power-law characteristics.

[0176] To sum up, the user classification device provided in the embodiment of the present application inputs the user's structured data, image data and text data into three models respectively to obtain three outputs of the three models. Then, based on these three outputs, the target probability of the user belonging to each of the preset multiple categories of users can be obtained. Since the target probability is the result obtained by integrating multiple different types of data, the influence of the target probability on the fluctuation of a single type of data can be reduced, thereby achieving the effect of improving the stability of the user classification method.

[0177] In addition, at least one of the first model, the second model and the third model includes a target neural network with small-world characteristics and power-law characteristics, which can improve the data transmission efficiency of the neural network and reduce the number of parameters in the neural network.

[0178] Optionally, the target neural network with small-world characteristics and power-law characteristics includes an input layer, multiple hidden layers and an output layer, the multiple hidden layers include at least two target hidden layers, the degree distribution of nodes in the at least two target hidden layers conforms to the power-law distribution, and the other layers in the neural network except the at least two target hidden layers conform to the small-world characteristics.

[0179] Alternatively, please refer to FIG. 7 , which is a structural block diagram of another user classification device provided in an embodiment of the present application. The user classification device is adjusted based on the user classification device shown in FIG. 6 , and further includes:

[0180] Model acquisition module 660, used to acquire a variational autoencoder network to be trained, a second model to be trained, and a third model to be trained, wherein the variational autoencoder network includes a first input layer, a structured data encoder to be trained, a first output layer, a structured data decoder, and a decoder output layer; the second model includes a visual encoder, a second neural network to be trained, and a second normalization module; and the third model to be trained includes a text encoder, a third neural network to be trained, and a third normalization module;

[0181] A sample acquisition module 670 is configured to acquire a target training sample from a target sample set, where the target training sample includes a structured data sample, an image data sample, and a text data sample;

[0182] The training input module 680 is used to input structured data samples into the variational autoencoder network to be trained, input image data samples into the second model to be trained, and input text data samples into the third model to be trained.

[0183] The output acquisition module 690 is used to obtain the first training output of the variational autoencoder network to be trained, the second training output of the second model to be trained, and the third training output of the third model to be trained.

[0184] The comprehensive output module 691 is configured to obtain a comprehensive training output based on the first training output, the second training output, and the third training output. The comprehensive training output includes a target probability that the user belongs to each of the preset multiple user categories.

[0185] An adjustment module 692 is configured to adjust parameters of the second neural network to be trained in the second model based on the training synthesis output, adjust parameters of the third neural network to be trained in the third model based on the training synthesis output, and adjust parameters of the structured data encoder to be trained based on the training synthesis output and the output of the decoder output layer.

[0186] The repeated training module 693 is configured to execute the step of acquiring a target training sample from the target sample set when a preset training cutoff condition is not met.

[0187] The training cutoff module 694 is used to determine the structured data encoder to be trained as the structured data encoder, the second neural network to be trained as the second neural network, and the third neural network to be trained as the third neural network when a preset training cutoff condition is reached.

[0188] To sum up, the user classification device provided in the embodiment of the present application inputs the user's structured data, image data and text data into three models respectively to obtain three outputs of the three models. Then, based on these three outputs, the target probability of the user belonging to each of the preset multiple categories of users can be obtained. Since the target probability is the result obtained by integrating multiple different types of data, the influence of the target probability on the fluctuation of a single type of data can be reduced, thereby achieving the effect of improving the stability of the user classification method.

[0189] In addition, at least one of the first model, the second model and the third model includes a target neural network with small-world characteristics and power-law characteristics, which can improve the data transmission efficiency of the neural network and reduce the number of parameters in the neural network.

[0190] In addition, an embodiment of the present application also provides a user classification device, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the user classification method as described above.

[0191] An embodiment of the present application also provides a computer storage medium, in which at least one instruction, at least one program, code set or instruction set is stored, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement the user classification method as described above.

[0192] An embodiment of the present application also provides a server, which is deployed with the comprehensive model shown in Figure 4 of the above embodiment, and then the server can implement the user classification method provided in the above embodiment based on the comprehensive model.

[0193] In this application, the terms "first", "second", and "third" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. The term "plurality" refers to two or more than two, unless otherwise clearly defined.

[0194] It should be noted that the information involved in this application (including but not limited to the user's structured data, image data, and text data, etc.) is authorized by the user or fully authorized by all parties, and the collection, use, and processing of the relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the target sample set and the user's structured data, image data, and text data involved in this application are all obtained with full authorization.

[0195] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0196] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0197] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0198] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A classification method for users, characterized in that, The method includes: Obtaining the structured data, image data, and text data of the user; Inputting the structured data into a first model to obtain a first output of the first model, where the first model is used to output the first output including the probability of the user belonging to each of the preset multiple types of users based on the structured data; Inputting the image data into a second model to obtain a second output of the second model, where the second model is used to output the second output including the probability of the user belonging to each of the preset multiple types of users based on the image data; Inputting the text data into a third model to obtain a third output of the third model, where the third model is used to output the third output including the probability of the user belonging to each of the preset multiple types of users based on the text data; Based on the first output, the second output, and the third output, determining the target probability of the user belonging to each of the preset multiple types of users; Wherein, at least one of the first model, the second model, and the third model includes a target neural network having small-world characteristics and power-law characteristics.

2. The method according to claim 1, wherein The target neural network having small-world characteristics and power-law characteristics includes an input layer, a plurality of hidden layers, and an output layer. The plurality of hidden layers includes at least two target hidden layers, and the degree distribution of the nodes in the at least two target hidden layers conforms to the power-law distribution. Other layers in the target neural network except the at least two target hidden layers conform to the small-world characteristics.

3. The method according to claim 1, wherein The first model includes a first input layer, a structured data encoder, a first output layer, a fully connected layer, and a first normalization module. The first input layer, the structured data encoder, and the first output layer constitute the first neural network having small-world characteristics and power-law characteristics.

4. The method according to claim 3, characterized in that The second model includes a visual encoder, a second neural network, and a second normalization module; The second neural network is the target neural network having small-world characteristics and power-law characteristics.

5. The method according to claim 3, characterized in that, The third model includes a text encoder, a third neural network, and a third normalization module; The third neural network is the target neural network having small-world characteristics and power-law characteristics.

6. The method according to claim 1, wherein The method further includes: Obtaining a variational autoencoder network to be trained, a second model to be trained, and a third model to be trained. The variational autoencoder network includes the first input layer, a structured data encoder to be trained, the first output layer, a structured data decoder, and a decoder output layer. The second model to be trained includes a visual encoder, a second neural network to be trained, and a second normalization module. The third model to be trained includes a text encoder, a third neural network to be trained, and a third normalization module; Obtaining a target training sample from a target sample set, where the target training sample includes a structured data sample, an image data sample, and a text data sample of a sample user; Input the structured data sample into the variational autoencoder network to be trained, input the image data sample into the second model to be trained, and input the text data sample into the third model to be trained; Obtain the first training output output by the variational autoencoder network to be trained, the second training output output by the second model to be trained, and the third training output output by the third model to be trained; Obtain a training comprehensive output based on the first training output, the second training output, and the third training output, where the training comprehensive output includes the target probability that the sample user belongs to each type of user among the preset multiple types of users; Adjust the parameters of the second neural network to be trained in the second model based on the training comprehensive output, adjust the parameters of the third neural network to be trained in the third model based on the training comprehensive output, and adjust the parameters of the structured data encoder to be trained based on the training comprehensive output and the output of the decoder output layer; When the preset training cut-off condition is not reached, execute the step of obtaining a target training sample in the target sample set; When the preset training cut-off condition is reached, determine the structured data encoder to be trained as the structured data encoder, determine the second neural network to be trained as the second neural network, and determine the third neural network to be trained as the third neural network.

7. The method according to claim 6, wherein The method further includes: Obtain the loss of the variational autoencoder network to be trained and the loss of the training comprehensive output; Determine the loss value of this training based on the loss of the variational autoencoder network to be trained and the loss of the training comprehensive output; When the loss value of this training is greater than or equal to the target loss value, determine that the training cut-off condition is reached, where the target loss value is the average value of the loss values of the target number of trainings before this training.

8. The method according to claim 1, wherein The first model, the second model, and the third model all include a target neural network with small-world characteristics and power-law characteristics.

9. The method according to claim 8, wherein The target neural network included in the first model, the second model, and the third model has any one of the small-world characteristics of Watts-Strogatz small-world characteristics and Newman-Watts small-world characteristics.

10. The method according to any one of claims 1 to 8, characterized in that, The structured data includes various types of data of the user, and the types at least include age, height, weight, and various health parameters; The image data includes the medical image data of the user; The text data includes the text data in the user's electronic medical record.

11. A classification device for users, characterized in that, The classification device for the user includes: An acquisition module, configured to acquire the structured data, image data, and text data of the user; A first input module, configured to input the structured data into the first model to obtain a first output of the first model, where the first model is configured to output a first output including the probability that the user belongs to each type of user among the preset multiple types of users based on the structured data; A second input module, configured to input the image data into a second model to obtain a second output of the second model, where the second model is configured to output the second output including the probability that the user belongs to each type of user in the preset multiple types of users based on the image data; A third input module, configured to input the text data into a third model to obtain a third output of the third model, where the third model is configured to output the third output including the probability that the user belongs to each type of user in the preset multiple types of users based on the text data; A determination module, configured to determine, based on the first output, the second output, and the third output, the target probability that the user belongs to each type of user in the preset multiple types of users; Wherein, at least one of the first model, the second model, and the third model includes a target neural network having small-world characteristics and power-law characteristics.

12. The classification device for a user according to claim 11, characterized in that, The target neural network having small-world characteristics and power-law characteristics includes an input layer, a plurality of hidden layers, and an output layer. The plurality of hidden layers include at least two target hidden layers. The degree distribution of the nodes in the at least two target hidden layers conforms to a power-law distribution, and other layers in the target neural network except the at least two target hidden layers conform to small-world characteristics.

13. The classification device for a user according to claim 11, characterized in that, The classification device for the user further includes: A model acquisition module, configured to acquire a variational autoencoder network to be trained, a second model to be trained, and a third model to be trained. The variational autoencoder network includes the first input layer, a structured data encoder to be trained, the first output layer, a structured data decoder, and a decoder output layer. The second model includes a visual encoder, a second neural network to be trained, and a second normalization module. The third model to be trained includes a text encoder, a third neural network to be trained, and a third normalization module; A sample acquisition module, configured to acquire a target training sample from a target sample set, where the target training sample includes a structured data sample, an image data sample, and a text data sample; A training input module, configured to input the structured data sample into the variational autoencoder network to be trained, input the image data sample into the second model to be trained, and input the text data sample into the third model to be trained; An output acquisition module, configured to acquire a first training output output by the variational autoencoder network to be trained, a second training output output by the second model to be trained, and a third training output output by the third model to be trained; A comprehensive output module, configured to obtain a training comprehensive output based on the first training output, the second training output, and the third training output, where the training comprehensive output includes the target probability that the user belongs to each type of user in the preset multiple types of users; An adjustment module, configured to adjust the second neural network to be trained in the second model based on the training comprehensive output The parameters of the neural network, adjust the parameters of the third neural network to be trained in the third model based on the training comprehensive output, and adjust the parameters of the structured data encoder to be trained based on the training comprehensive output and the output of the decoder output layer; A repeated training module, configured to execute the step of obtaining a target training sample in the target sample set when the preset training cut-off condition is not reached; A training cut-off module, configured to, when the preset training cut-off condition is reached, determine the structured data encoder to be trained as the structured data encoder, determine the second neural network to be trained as the second neural network, and determine the third neural network to be trained as the third neural network.

14. A classification device for users, characterized in that, The classification device of the user includes a processor and a memory, and at least one instruction, at least one program, a code set or an instruction set is stored in the memory, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the classification method of the user according to any one of claims 1 to 10.

15. A non-volatile computer storage medium, characterized in that, At least one instruction, at least one program, a code set or an instruction set is stored in the computer storage medium, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement the classification method of the user according to any one of claims 1 to 10.

Images