WO2021042556A1

WO2021042556A1 - Classification model training method, apparatus and device, and computer-readable storage medium

Info

Publication number: WO2021042556A1
Application number: PCT/CN2019/118247
Authority: WO
Inventors: 金戈; 徐亮
Original assignee: 平安科技（深圳）有限公司
Priority date: 2019-09-03
Filing date: 2019-11-14
Publication date: 2021-03-11
Also published as: CN110705592B; CN110705592A

Abstract

The present application relates to the technical field of artificial intelligence. Disclosed are a classification model training method, apparatus and device, and a computer-readable storage medium. The classification model training method comprises: acquiring sample data; obtaining, on the basis of a feature extraction algorithm, features corresponding to the sample data, wherein the features of the sample data comprise discrete features and continuous features; converting the discrete features into continuous features; inputting the continuous features into an autoencoder to obtain implicit features; constructing an initial classification model on the basis of labeled sample data and the implicit features, and performing label prediction on unlabeled sample data on the basis of the initial classification model and a preset expectation step algorithm; optimizing the initial classification model according to a prediction result combined with a preset maximization step algorithm; and when it is detected that the preset expectation step algorithm starts to converge, confirming that the training of the initial classification model is completed, and saving the trained initial classification model. By means of the present application, the generalization capability of a classification model is improved.

Description

Classification model training method, device, equipment and computer readable storage medium

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on September 3, 2019, the application number is 201910826406.8, and the invention title is "Classification Model Training Method, Apparatus, Equipment, and Computer-readable Storage Medium", and its entire contents Incorporated in the application by reference.

Technical field

This application relates to the field of artificial intelligence technology, in particular to classification model training methods, devices, equipment, and computer-readable storage media.

Background technique

In many data classification applications, such as text classification, image classification, and mining of special customer groups, a large number of samples are required for classification model training. Among them, labeled samples are usually difficult to obtain automatically, and generally require manual labeling. Therefore, the number of labeled samples in the training samples is usually small, and most of them are unlabeled samples. The inventor realizes that in the process of training the classification model, due to the existence of a large number of unlabeled samples, the model may be over-fitted or the accuracy rate is not high.

Summary of the invention

The main purpose of this application is to provide a classification model training method, device, equipment, and computer-readable storage medium, aiming to solve the technical problem of overfitting or low accuracy of the existing classification model.

In order to achieve the above objective, the first aspect of the present application provides a classification model training method. The classification model training method includes the following steps: obtaining sample data, wherein the sample data includes labeled sample data and unlabeled sample data Processing the sample data based on a feature extraction algorithm to obtain features corresponding to the sample data, wherein the features of the sample data include discrete features and continuous features, the continuous features are in numerical form, and the discrete features are in non-numerical form; The discrete feature is processed based on the feature conversion method, and the discrete feature is converted into a continuous feature; the continuous feature and the continuous feature obtained by the discrete feature conversion are input into an auto-encoding algorithm for dimensionality reduction processing to obtain the The hidden features corresponding to the sample data; an initial classification model is constructed based on the labeled sample data and the hidden features, and the unlabeled sample data is performed based on the initial classification model and a preset desired step algorithm Label prediction; according to the prediction result, combined with the preset maximization step algorithm to optimize the initial classification model; when it is detected that the preset expected step algorithm starts to converge, confirm that the initial classification model training is completed, and save The initial classification model that has been trained.

A second aspect of the present application provides a classification model training device, the classification model training device includes: a data acquisition module for acquiring sample data, wherein the sample data includes labeled sample data and unlabeled sample data; The feature extraction module is used to process the sample data based on a feature extraction algorithm to obtain features corresponding to the sample data, where the features of the sample data include discrete features and continuous features, and the continuous features are in numerical form, and discrete features In a non-numerical form; a feature conversion module, used to process the discrete features based on a feature conversion method, and convert the discrete features into continuous features; a feature dimensionality reduction module, used to convert the continuous features and the discrete features The converted continuous features are input into the self-encoding algorithm for dimensionality reduction processing to obtain the hidden features corresponding to the sample data; the label prediction module is used to construct an initial classification model based on the labeled sample data and the hidden features , And perform label prediction on the unlabeled sample data based on the initial classification model and the preset expected step algorithm; the model optimization module is used to perform the label prediction on the initial classification based on the prediction result and the preset maximization step algorithm The model is optimized; the model saving module is used to confirm that the training of the initial classification model is completed when it is detected that the preset desired step algorithm starts to converge, and to save the completed initial classification model.

A third aspect of the present application provides a classification model training device, including: a memory and at least one processor, the memory stores instructions, the memory and the at least one processor are interconnected through a wire; the at least one processor The device invokes the instructions in the memory, so that the classification model training device executes the method described in the first aspect.

The fourth aspect of the present application provides a computer-readable storage medium, the computer-readable storage medium stores computer instructions, and when the computer instructions run on a computer, the computer executes the above-mentioned first aspect method.

The classification model training method, device, equipment, and computer-readable storage medium provided in this application first obtain labeled sample data and unlabeled sample data, and obtain discrete and continuous features corresponding to the sample data based on a feature extraction algorithm; Process discrete features, convert them into continuous features, and input all continuous features into the self-encoding algorithm for dimensionality reduction to obtain the hidden features corresponding to the sample data; build an initial classification model based on the labeled sample data and the hidden features , Perform label prediction on unlabeled sample data through the initial classification model and the preset expected step algorithm; according to the prediction result, combined with the preset maximization step algorithm to optimize the initial classification model, when the preset expected step algorithm starts When converging, confirm that the initial classification model training is completed, and save the completed initial classification model. The classification model training method proposed in this application realizes the effective dimensionality reduction of features through a self-encoding algorithm, and combines the maximum expected value algorithm to improve the generalization ability of the classification model by using unlabeled sample data.

Description of the drawings

FIG. 1 is a schematic diagram of the structure of the classification model training device of the hardware operating environment involved in the embodiment of the application;

2 is a schematic flowchart of an embodiment of the classification model training method in this application;

3 is a schematic diagram of functional modules of an embodiment of the classification model training device in this application;

4 is a schematic diagram of functional units of a feature conversion module in an embodiment of the classification model training device in this application;

FIG. 5 is a schematic diagram of functional units of a label prediction module in an embodiment of the classification model training device in this application;

6 is a schematic diagram of functional units of a model optimization module in an embodiment of the classification model training device in this application;

FIG. 7 is a schematic structural diagram of a self-encoding algorithm in an embodiment of the classification model training method in this application.

detailed description

The embodiments of the application provide a classification model training method, device, equipment, and computer-readable storage medium, which are used to achieve effective feature reduction through a self-encoding algorithm, combined with the maximum expected value algorithm, and use unlabeled sample data to improve The generalization ability of the classification model.

In order to enable those skilled in the art to better understand the solution of the present application, the embodiments of the present application will be described below in conjunction with the accompanying drawings in the embodiments of the present application.

As shown in Fig. 1, Fig. 1 is a schematic structural diagram of a classification model training device for a hardware operating environment involved in a solution of an embodiment of the application.

As shown in FIG. 1, the classification model training device may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to implement connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as a magnetic disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

As shown in FIG. 1, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a classification model training program.

In the classification model training device shown in FIG. 1, the network interface 1004 is mainly used to connect to a back-end server and perform data communication with the back-end server; the user interface 1003 is mainly used to connect to a client (user side) and perform data communication with the client; The processor 1001 can be used to call the classification model training program stored in the memory 1005 and execute the operations of the following classification model training methods.

Referring to FIG. 2, FIG. 2 is a schematic flowchart of an embodiment of a classification model training method according to the present application. In this embodiment, the classification model training method includes:

Step S10: Obtain sample data, where the sample data includes labeled sample data and unlabeled sample data.

In this embodiment, first obtain sample data for training the classification model, where the sample data includes a large amount of unlabeled sample data and a small amount of labeled sample data. Taking the population classification model as an example, the label specifically represents the type of population corresponding to the sample data. For example, the label of the sample data can be a high-consumption group; the sample data should also include the personal background information and consumption behavior of the group to be classified. In particular, personal background information may include age, gender, occupation, income, city of residence, and educational background, and consumption behavior may include specific characteristics such as the user's monthly expenditure.

Step S20: Process the sample data based on the feature extraction algorithm to obtain features corresponding to the sample data, where the features of the sample data include discrete features and continuous features, the continuous features are in numerical form, and the discrete features are in non-numerical form.

Further, the processing of sample data is mainly to perform feature extraction on sample data through feature extraction algorithms. Feature extraction algorithms include but are not limited to principal component analysis, independent component analysis, and linear discriminant analysis. In this embodiment, The feature extraction algorithm of sample data is not limited.

In this embodiment, the extracted features include discrete features and continuous features, where the continuous features are in numerical form and the discrete features are in non-numerical form. For example, the income in the sample data is a continuous feature, while the city of residence is a discrete feature.

In step S30, the discrete features are processed based on the feature conversion method, and the discrete features are converted into continuous features.

Further, in order to facilitate the training of the classification model, it is necessary to convert the extracted discrete features into continuous features. In this embodiment, processing discrete features into continuous features includes the following three situations:

1. Discrete features have an order relationship. For example, the discrete feature of "level" can include "first level", "second level" and "third level". Therefore, such discrete features can be directly quantified , Transformed into continuous features;

2. Discrete features have a non-order relationship, and the number of discrete values of discrete features is less than or equal to the preset number, such as the discrete feature of "educational background". The discrete values include college, undergraduate, master, and doctoral degrees. The number of values is limited. Therefore, such discrete features can be processed based on the one-hot encoding method and converted into continuous features;

3. Discrete features have a non-order relationship, and the number of discrete values of discrete features is greater than the preset number, such as the discrete feature of "residential city", if there are many discrete values, you can perform such discrete features Derivative processing transforms the discrete feature of "residential city" into a continuous feature of a higher-level province or city.

Step S40: Input the continuous features obtained by converting the continuous features and the discrete features into the self-encoding algorithm for dimensionality reduction processing to obtain the hidden features corresponding to the sample data.

After the feature extraction of the sample data is completed to obtain continuous features and discrete features, and the discrete features are converted into continuous features, all continuous features are input into the auto-encoding algorithm, so that all continuous features can be performed based on the self-encoding algorithm. Reduce dimensionality to get hidden features.

The self-encoding algorithm is an unsupervised learning method based on the hidden features of neural network learning. The structure of the self-encoding algorithm is symmetrical. As shown in Figure 7, in the self-encoding algorithm, the input is the continuous feature after feature conversion processing. The self-encoding algorithm contains one or more hidden layers, and the output of the intermediate hidden layer is extracted as the implicit dimensionality reduction. Feature output. The specific process is: the trained self-encoding algorithm converts the input continuous features into hidden features through encoding, and then decodes the hidden features to obtain output features similar to the input continuous features, realizing continuous input Feature dimensionality reduction.

In step S50, an initial classification model is constructed based on the labeled sample data and hidden features, and label prediction is performed on the unlabeled sample data based on the initial classification model and a preset desired step algorithm.

Furthermore, on the basis of the hidden features output after dimensionality reduction, a classification model is constructed to realize the semi-supervised learning of the maximum expected value algorithm. Specifically, the maximum expected value algorithm establishes an initial classification model on the basis of labeled sample data. Specifically, the classification model in this embodiment refers to a Gaussian mixture model. The established initial Gaussian mixture model is used to predict the unlabeled data, and the initial Gaussian mixture model is optimized by combining the labeled sample data to obtain the final Gaussian mixture model that can be used for crowd classification.

Specifically, in this embodiment, assuming that the sample data contains k sets of labeled sample data and u sets of unlabeled sample data, the sample data can be expressed as D={(X ₁ , Y ₁ ), (X ₂ , Y ₂ ),..., (X _k , Y _k ), (X _k+1 ), (X _k+2 ),..., (X _k+u )}. Among them, (X ₁ , Y ₁ ), (X ₂ , Y ₂ ),..., (X _k , Y _k ) where X _i represents the sample data, and Y _i represents the label of the i-th group of sample data , The labels of different sample data can be the same or different; (X _k+1 ), (X _k+2 ),..., (X _k+u ) represent sample data without labels.

Further, assuming that the dependent variable in the sample data includes m categories, that is, the label of the sample data includes m categories, it can be known that m≤k. In this embodiment, P(x) can be used to represent _{the probability value of the sample data X j} on the i-th label. The probability distribution of the Gaussian mixture model is shown in the following formula:

Among them, π is the mixing coefficient, x is the eigenvector, μ is the mean vector of x, and Σ is the covariance matrix.

For a sample of tagged data X _i, the probability of the label on the Y _i value of 1, whereas for other types of tags probability value of zero.

In this embodiment, label prediction is performed on the unlabeled sample data according to the initial Gaussian mixture model and the preset desired step algorithm, and the corresponding label is determined.

Step S60: According to the prediction result, the initial classification model is optimized in combination with the preset maximization step algorithm.

After the initial Gaussian mixture model and the preset expected step algorithm determine the label corresponding to the unlabeled sample data, the parameters of the entire initial Gaussian mixture model are further optimized through the preset maximization step algorithm to prevent the initial Gaussian mixture The model is overfitting or the label prediction is inaccurate.

Step S70: When it is detected that the preset desired step algorithm starts to converge, it is confirmed that the training of the initial classification model is completed, and the completed initial classification model is saved.

Repeat the process of label prediction based on the preset desired step algorithm for unlabeled sample data, and the process of optimizing the parameters of the entire initial Gaussian mixture model based on the preset maximum step algorithm until the preset desired step algorithm Start to converge, it can be regarded as the classification model training completed.

Further, in this embodiment, after the training of the classification model is completed, the online prediction of the type of crowd can be performed based on the trained classification model. For the new sample data that needs to be classified and predicted by the type of population, first, the new sample data needs to be preprocessed to obtain the feature information corresponding to the new sample data; and the corresponding feature information is input into the auto-encoding algorithm for dimensionality reduction ; Finally, the dimensionality-reduced features are input into the Gaussian mixture model to realize the classification prediction of the crowd types.

In this embodiment, the labeled sample data and the unlabeled sample data are first acquired, and the discrete and continuous features corresponding to the sample data are acquired based on the feature extraction algorithm; the discrete features are processed, converted into continuous features, and all The continuous features are input into the self-encoding algorithm for dimensionality reduction processing, and the hidden features corresponding to the sample data are obtained; the initial classification model is constructed based on the labeled sample data and the hidden features, and the initial classification model and the preset expected step algorithm Label sample data for label prediction; according to the prediction results, combined with the preset maximization step algorithm to optimize the initial classification model, when the preset expected step algorithm starts to converge, confirm that the initial classification model training is completed, and save the completed training Initial classification model. The classification model training method proposed in this application realizes the effective dimensionality reduction of features through a self-encoding algorithm, and combines the maximum expected value algorithm to improve the generalization ability of the classification model by using unlabeled sample data.

Further, step S50 includes:

_{Step S501: Determine the initial parameters π i} , μ _i and ∑ _{i of the} initial classification model based on the labeled sample data and the hidden features, and construct the initial classification model based on the initial parameters, π _i , μ _i and ∑ _The formula for calculating the initial value of i is as follows:

Among them, ∑ is the covariance matrix, X _j is the sample data, and γ _ij is the posterior probability containing hidden features;

Step S502, in the initial classification model, perform label prediction on the unlabeled sample data through a preset expected step algorithm, and the preset expected step algorithm has the following formula:

Among them, π _i is the mixing coefficient.

In this embodiment, when the dimensionality reduction process is performed on the continuous features through the self-encoding algorithm to obtain the hidden features contained in the sample data, the initial parameters π _i and μ _{i of the Gaussian mixture model are determined based on the labeled sample data and the hidden features} And ∑ _i . Specifically, the calculation formulas for the initial values of the three parameters are as follows:

Among them, Σ is the covariance matrix, X _j is the sample data, and γ _ij is the posterior probability containing hidden features.

By determining the initial parameters of the Gaussian mixture model through the labeled sample data and hidden features, the initial classification model can be constructed. Based on the initial classification model to perform label prediction on the unlabeled sample data, it can be understood that the predicted label at this time may not be correct. Therefore, the initial classification model needs to be optimized through the maximization step algorithm. Specifically, the formula of the maximization step algorithm is as follows:

According to the prediction results, the initial parameters of the initial classification model are updated based on the maximization step algorithm to form a new Gaussian mixture model, and then label prediction is performed on the unlabeled sample data based on the new Gaussian mixture model until the preset desired step algorithm Start to converge, it can be regarded as the completion of model training.

In this embodiment, the hidden features obtained through dimensionality reduction are input into the maximum expected value algorithm, and the classification model is semi-supervised learning combined with labeled and unlabeled sample data to prevent the classification model from over-fitting or under-fitting , To improve the generalization performance of the classification model.

Referring to FIG. 3, FIG. 3 is a schematic diagram of functional modules of an embodiment of a classification model training apparatus according to the present application.

In this embodiment, the classification model training device includes:

The data acquisition module 10 is configured to acquire sample data, where the sample data includes labeled sample data and unlabeled sample data;

The feature extraction module 20 is configured to process the sample data based on a feature extraction algorithm to obtain features corresponding to the sample data, where the features of the sample data include discrete features and continuous features, and the continuous features are in numerical form. Features are in non-numerical form;

The feature conversion module 30 is configured to process the discrete features based on a feature conversion method, and convert the discrete features into continuous features;

The feature dimensionality reduction module 40 is configured to input the continuous feature converted from the continuous feature and the discrete feature into an auto-encoding algorithm for dimensionality reduction processing to obtain the implicit feature corresponding to the sample data;

The label prediction module 50 is configured to construct an initial classification model based on the labeled sample data and the hidden features, and perform label prediction on the unlabeled sample data based on the initial classification model and a preset desired step algorithm ；

The model optimization module 60 is configured to optimize the initial classification model according to the prediction result in combination with a preset maximization step algorithm;

The model saving module 70 is configured to confirm that the training of the initial classification model is completed when it is detected that the preset desired step algorithm starts to converge, and to save the completed initial classification model.

Further, referring to FIG. 4, the feature conversion module 30 includes:

The quantization processing unit 301 is configured to perform quantization processing on the discrete features if the discrete features have an order relationship, and convert the discrete features into continuous features;

The encoding processing unit 302 is configured to: if the discrete features have a non-order relationship, and the number of discrete values of the discrete features is less than or equal to a preset number, perform a one-hot encoding method on the discrete features Processing to convert the discrete features into continuous features;

The derivation processing unit 303 is configured to, if the discrete features have a non-order relationship, and the number of discrete values of the discrete features is greater than a preset number, perform derivation processing on the discrete features to convert the discrete features It is a continuous feature.

Further, referring to FIG. 5, the label prediction module 50 includes:

The model construction unit 501 is configured to determine the initial parameters π _i , μ _i and ∑ _{i of the} initial classification model based on the labeled sample data and the hidden features, and construct the initial classification model based on the initial parameters, π _i , The calculation formulas for the initial values of _{μ i} and ∑ _{i are as follows:}

The label prediction unit 502 is configured to perform label prediction on the unlabeled sample data by using a preset desired step algorithm in the initial classification model, and the preset desired step algorithm has the following formula:

Among them, π _i is the mixing coefficient.

Further, referring to FIG. 6, the model optimization module 60 includes:

The model optimization unit 601 is used to obtain the preset maximization step algorithm formula as follows:

According to the prediction result, the initial parameters of the initial classification model are updated based on the formula.

Further, the feature dimensionality reduction module 40 is specifically used for:

Input the continuous features into the self-encoding algorithm for dimensionality reduction to obtain the initial hidden features;

The initial hidden features are decoded to obtain the hidden features.

This application also provides a classification model training device, including: a memory and at least one processor, the memory stores instructions, the memory and the at least one processor are interconnected through a wire; the at least one processor calls the The instructions in the memory are used to cause the classification model training device to execute the steps in the above classification model training method.

The present application also provides a computer-readable storage medium. The computer-readable storage medium may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores computer instructions, and when the computer instructions are executed on the computer, the computer executes the following steps:

Acquiring sample data, where the sample data includes labeled sample data and unlabeled sample data;

Processing the sample data based on a feature extraction algorithm to obtain features corresponding to the sample data, wherein the features of the sample data include discrete features and continuous features, the continuous features are in a numerical form, and the discrete features are in a non-numerical form;

Processing the discrete features based on a feature conversion method, and converting the discrete features into continuous features;

Input the continuous feature and the continuous feature obtained by the conversion of the discrete feature into an auto-encoding algorithm for dimensionality reduction processing to obtain the hidden feature corresponding to the sample data;

Constructing an initial classification model based on the labeled sample data and the hidden features, and performing label prediction on the unlabeled sample data based on the initial classification model and a preset desired step algorithm;

Optimizing the initial classification model according to the prediction result in combination with a preset maximization step algorithm;

When it is detected that the preset desired step algorithm starts to converge, it is confirmed that the training of the initial classification model is completed, and the completed initial classification model is saved.

As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, not to limit them; those of ordinary skill in the art should understand that they can still modify or modify the technical solutions described in the foregoing embodiments. Some of the technical features are equivalently replaced; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

A classification model training method, the classification model training method includes the following steps:

Acquiring sample data, where the sample data includes labeled sample data and unlabeled sample data;

Processing the sample data based on a feature extraction algorithm to obtain features corresponding to the sample data, wherein the features of the sample data include discrete features and continuous features, the continuous features are in a numerical form, and the discrete features are in a non-numerical form;

Processing the discrete features based on a feature conversion method, and converting the discrete features into continuous features;

Input the continuous feature and the continuous feature obtained by the conversion of the discrete feature into an auto-encoding algorithm for dimensionality reduction processing to obtain the hidden feature corresponding to the sample data;

Constructing an initial classification model based on the labeled sample data and the hidden features, and performing label prediction on the unlabeled sample data based on the initial classification model and a preset desired step algorithm;

Optimizing the initial classification model according to the prediction result in combination with a preset maximization step algorithm;

When it is detected that the preset desired step algorithm starts to converge, it is confirmed that the training of the initial classification model is completed, and the completed initial classification model is saved.
8. The classification model training method according to claim 1, wherein the processing of the discrete features based on the feature conversion method, and converting the discrete features into continuous features comprises:

If the discrete features have an order relationship, perform quantization processing on the discrete features, and convert the discrete features into continuous features;

If the discrete feature has a non-order relationship, and the number of discrete values of the discrete feature is less than or equal to the preset number, the discrete feature is processed based on the one-hot one-hot encoding method, and the discrete Convert features to continuous features;

If the discrete feature has a non-order relationship, and the number of discrete values of the discrete feature is greater than the preset number, then the discrete feature is derivatized to convert the discrete feature into a continuous feature.
The classification model training method according to claim 1, wherein the initial classification model is constructed based on the labeled sample data and the hidden features, and the initial classification model is based on the initial classification model and a preset desired step algorithm for the Labeled sample data for label prediction includes:

The initial parameters π i , μ i and ∑ i of the initial classification model are determined based on the labeled sample data and the hidden features, and the initial classification model is constructed based on the initial parameters. The initial parameters of π i , μ i and ∑ i The value calculation formula is as follows:

Among them, ∑ is the covariance matrix, X j is the sample data, and γ ij is the posterior probability containing hidden features;

In the initial classification model, label prediction is performed on the unlabeled sample data through a preset desired step algorithm, and the formula of the preset desired step algorithm is as follows:

Among them, π i is the mixing coefficient.
The classification model training method according to claim 3, wherein the optimization of the initial classification model based on the prediction result in combination with a preset maximization step algorithm comprises:

The formula for obtaining the preset maximization step algorithm is as follows:

According to the prediction result, the initial parameters of the initial classification model are updated based on the formula.
The classification model training method according to claim 1, wherein the continuous features obtained by converting the continuous features and the discrete features are input into an auto-encoding algorithm for dimensionality reduction processing to obtain the hidden features corresponding to the sample data include:

Input the continuous features into the self-encoding algorithm for dimensionality reduction to obtain the initial hidden features;

The initial hidden features are decoded to obtain the hidden features.
A classification model training device, the classification model training device includes:

A data acquisition module for acquiring sample data, where the sample data includes labeled sample data and unlabeled sample data;

The feature extraction module is used to process the sample data based on a feature extraction algorithm to obtain features corresponding to the sample data, where the features of the sample data include discrete features and continuous features, and the continuous features are in numerical form, and discrete features Is a non-numerical form;

A feature conversion module, configured to process the discrete features based on a feature conversion method, and convert the discrete features into continuous features;

The feature dimensionality reduction module is used to input the continuous feature obtained by the conversion of the continuous feature and the discrete feature into an auto-encoding algorithm for dimensionality reduction processing to obtain the hidden feature corresponding to the sample data;

The label prediction module is configured to construct an initial classification model based on the labeled sample data and the hidden features, and perform label prediction on the unlabeled sample data based on the initial classification model and a preset desired step algorithm;

The model optimization module is used to optimize the initial classification model according to the prediction result in combination with a preset maximization step algorithm;

The model saving module is used to confirm the completion of the training of the initial classification model when it is detected that the preset desired step algorithm starts to converge, and save the completed initial classification model.
7. The classification model training device according to claim 6, wherein the feature conversion module comprises:

A quantization processing unit, configured to perform quantization processing on the discrete features if the discrete features have an order relationship, and convert the discrete features into continuous features;

The encoding processing unit is configured to perform the discrete feature on the discrete feature based on the one-hot encoding method if the discrete feature has a non-order relationship and the number of discrete values of the discrete feature is less than or equal to a preset number Processing, converting the discrete features into continuous features;

A derivation processing unit, configured to, if the discrete features have a non-order relationship, and the number of discrete values of the discrete features is greater than a preset number, perform derivation processing on the discrete features to convert the discrete features into Continuous features.
The classification model training device according to claim 6, wherein the label prediction module comprises:

The model construction unit is used to determine the initial parameters π i , μ i and ∑ i of the initial classification model based on the labeled sample data and the hidden features, and construct the initial classification model based on the initial parameters, π i , μ The calculation formulas for the initial values of i and ∑ i are as follows:

Among them, ∑ is the covariance matrix, X j is the sample data, and γ ij is the posterior probability containing hidden features;

The label prediction unit is configured to perform label prediction on the unlabeled sample data through a preset desired step algorithm in the initial classification model, and the preset desired step algorithm has the following formula:

Among them, π i is the mixing coefficient.
8. The classification model training device according to claim 8, wherein the model optimization module comprises:

The model optimization unit is used to obtain the preset maximization step algorithm formula as follows:

According to the prediction result, the initial parameters of the initial classification model are updated based on the formula.
The classification model training device according to claim 6, wherein the feature conversion module is specifically configured to:

Input the continuous features into the self-encoding algorithm for dimensionality reduction to obtain the initial hidden features;

The initial hidden features are decoded to obtain the hidden features.
A classification model training device includes a memory, a processor, and a computer program stored on the memory and capable of running on the processor, and the processor implements the following steps when the processor executes the computer program:

Acquiring sample data, where the sample data includes labeled sample data and unlabeled sample data;

Processing the sample data based on a feature extraction algorithm to obtain features corresponding to the sample data, wherein the features of the sample data include discrete features and continuous features, the continuous features are in a numerical form, and the discrete features are in a non-numerical form;

Processing the discrete features based on a feature conversion method, and converting the discrete features into continuous features;

Input the continuous feature and the continuous feature obtained by the conversion of the discrete feature into an auto-encoding algorithm for dimensionality reduction processing to obtain the hidden feature corresponding to the sample data;

Constructing an initial classification model based on the labeled sample data and the hidden features, and performing label prediction on the unlabeled sample data based on the initial classification model and a preset desired step algorithm;

Optimizing the initial classification model according to the prediction result in combination with a preset maximization step algorithm;

When it is detected that the preset desired step algorithm starts to converge, it is confirmed that the training of the initial classification model is completed, and the completed initial classification model is saved.
The classification model training device according to claim 11, wherein the processor executes the computer program to implement the feature conversion method to process the discrete features, and when the discrete features are converted into continuous features, the process includes the following step:

If the discrete features have an order relationship, perform quantization processing on the discrete features, and convert the discrete features into continuous features;

If the discrete feature has a non-order relationship, and the number of discrete values of the discrete feature is less than or equal to the preset number, the discrete feature is processed based on the one-hot one-hot encoding method, and the discrete Convert features to continuous features;

If the discrete feature has a non-order relationship, and the number of discrete values of the discrete feature is greater than the preset number, then the discrete feature is derivatized to convert the discrete feature into a continuous feature.
The classification model training device according to claim 11, wherein the processor implements the construction of the initial classification model based on the labeled sample data and the hidden features when the computer program is executed, and is based on the initial classification model When performing label prediction on the unlabeled sample data with the preset expected step algorithm, the following steps are included:

The initial parameters π i , μ i and ∑ i of the initial classification model are determined based on the labeled sample data and the hidden features, and the initial classification model is constructed based on the initial parameters. The initial parameters of π i , μ i and ∑ i The value calculation formula is as follows:

Among them, ∑ is the covariance matrix, X j is the sample data, and γ ij is the posterior probability containing hidden features;

In the initial classification model, label prediction is performed on the unlabeled sample data through a preset desired step algorithm, and the formula of the preset desired step algorithm is as follows:

Among them, π i is the mixing coefficient.
The classification model training device according to claim 13, when the processor executes the computer program to realize the optimization of the initial classification model according to the prediction result in combination with a preset maximization step algorithm, comprising: The following steps:

The formula for obtaining the preset maximization step algorithm is as follows:

According to the prediction result, the initial parameters of the initial classification model are updated based on the formula.
The classification model training device according to claim 11, the device, the processor implements the input of the continuous feature and the continuous feature obtained by the conversion of the discrete feature into an auto-encoding algorithm when the processor executes the computer program. Dimension processing to obtain the hidden features corresponding to the sample data includes the following steps:

Input the continuous features into the self-encoding algorithm for dimensionality reduction to obtain the initial hidden features;

The initial hidden features are decoded to obtain the hidden features.
A computer-readable storage medium stores computer instructions in the computer-readable storage medium, and when the computer instructions are executed on a computer, the computer executes the following steps:

Acquiring sample data, where the sample data includes labeled sample data and unlabeled sample data;

Processing the sample data based on a feature extraction algorithm to obtain features corresponding to the sample data, wherein the features of the sample data include discrete features and continuous features, the continuous features are in a numerical form, and the discrete features are in a non-numerical form;

Processing the discrete features based on a feature conversion method, and converting the discrete features into continuous features;

Input the continuous feature and the continuous feature obtained by the conversion of the discrete feature into an auto-encoding algorithm for dimensionality reduction processing to obtain the hidden feature corresponding to the sample data;

Constructing an initial classification model based on the labeled sample data and the hidden features, and performing label prediction on the unlabeled sample data based on the initial classification model and a preset desired step algorithm;

Optimizing the initial classification model according to the prediction result in combination with a preset maximization step algorithm;

When it is detected that the preset desired step algorithm starts to converge, it is confirmed that the training of the initial classification model is completed, and the completed initial classification model is saved.
The computer-readable storage medium according to claim 16, when the computer instructions run the feature-based conversion method on the computer to process the discrete features and convert the discrete features into continuous features, the computer is caused to execute The following steps:

If the discrete features have an order relationship, perform quantization processing on the discrete features, and convert the discrete features into continuous features;

If the discrete feature has a non-order relationship, and the number of discrete values of the discrete feature is less than or equal to the preset number, the discrete feature is processed based on the one-hot one-hot encoding method, and the discrete Convert features to continuous features;

If the discrete feature has a non-order relationship, and the number of discrete values of the discrete feature is greater than the preset number, then the discrete feature is derivatized to convert the discrete feature into a continuous feature.
The computer-readable storage medium according to claim 16, when the computer instructions are executed on a computer, the initial classification model is constructed based on the labeled sample data and the hidden features, and based on the initial classification model and When the preset expected step algorithm performs label prediction on the unlabeled sample data, the computer executes the following steps:

The initial parameters π i , μ i and ∑ i of the initial classification model are determined based on the labeled sample data and the hidden features, and the initial classification model is constructed based on the initial parameters. The initial parameters of π i , μ i and ∑ i The value calculation formula is as follows:

Among them, ∑ is the covariance matrix, X j is the sample data, and γ ij is the posterior probability containing hidden features;

In the initial classification model, label prediction is performed on the unlabeled sample data through a preset desired step algorithm, and the formula of the preset desired step algorithm is as follows:

Among them, π i is the mixing coefficient.
The computer-readable storage medium according to claim 18, when the computer instruction is executed on the computer according to the prediction result, combined with a preset maximization step algorithm to optimize the initial classification model, so that the computer executes the following step:

The formula for obtaining the preset maximization step algorithm is as follows:

According to the prediction result, the initial parameters of the initial classification model are updated based on the formula.
16. The computer-readable storage medium according to claim 16, when the computer instructions are executed on a computer, the continuous features obtained by converting the continuous features and the discrete features are input into an auto-encoding algorithm for dimensionality reduction processing, When the hidden features corresponding to the sample data are obtained, the computer is caused to execute the following steps:

Input the continuous features into the self-encoding algorithm for dimensionality reduction to obtain the initial hidden features;

The initial hidden features are decoded to obtain the hidden features.