CN110196908A - Data classification method, device, computer installation and storage medium - Google Patents

Abstract

The invention provides a data classification method, device, computer device and storage medium. The method includes: obtaining a data set to be marked; marking the data set through a labeling function to obtain an initial label of the data set; calculating the pairwise correlation of the identification function according to the initial label, and according to the The pairwise correlation constructs the generation model of the labeling function; estimates the probability label of the data set according to the generation model; trains the discriminant model according to the probability label to obtain the trained discriminant model; The data is input into the trained discriminant model to obtain the category of the data to be classified. The invention improves the labeling efficiency and accuracy of the training data, and the discriminant model can be quickly trained by using the training data, and fast and accurate data classification is realized by using the discriminant model.

Description

Data classification method, device, computer device and storage medium

technical field

The invention relates to the technical field of machine learning, in particular to a data classification method, device, computer device and computer storage medium.

Background technique

With the rapid development of artificial intelligence, machine learning technology (especially deep learning technology) has been applied in various industries. At this point, training data labeling has gradually become the biggest bottleneck in the widespread deployment of machine learning systems.

The existing traditional manual labeling methods are time-consuming, labor-intensive and costly, and existing data enhancement methods such as semi-supervised learning, active learning, and transfer learning cannot quickly generate large-scale training data.

How to formulate a suitable solution, reduce the workload of manually labeling training data, and improve the efficiency of labeling training data is a technical problem that relevant technical personnel need to solve at present.

Contents of the invention

In view of the above, it is necessary to propose a data classification method, device, computer device and computer storage medium, which can improve the labeling efficiency of training data and classify data quickly and accurately.

The first aspect of the present application provides a data classification method applied to a machine learning system, the method comprising:

Obtain the dataset to be labeled {x _i |i=1, 2, ..., m};

The data set is marked by the labeling function λ _j , j=1, 2, ..., n, and the initial label Λ _{i, j} = λ _j ( _xi ), i=1, 2 of the data set is obtained ,...,m,j=1,2,...,n;

calculating the pairwise correlation of the identification function according to the initial label, and constructing a generative model of the labeling function according to the pairwise correlation;

Estimating a probability label of the data set according to the generative model;

training the discriminant model of the machine learning system according to the probability label to obtain a trained discriminant model;

Inputting the data to be classified into the trained discriminant model to obtain the category of the data to be classified.

In another possible implementation, the generation model is:

Where Λ represents the initial label matrix formed by the initial label, Y represents the real label matrix, Z _w is a normalization constant, φ _i (Λ, y _i ), i=1, 2, ..., m is the is the pairwise correlation of the labeling function of each data in the data set, w is an undetermined parameter of the generation model, w∈R ^2n+|C| .

In another possible implementation, the pairwise correlation is:

Among them, II{Λ _{i, j} =Λ _{i, k} } indicates the value when the condition in brackets {} is satisfied or not.

In another possible implementation, the data set to be labeled is an image set, and the data to be classified is an image to be classified; or

The data set to be labeled is a text set, and the data to be classified is a text to be classified; or

The data set to be marked is a speech set, and the data to be classified is a speech to be classified.

In another possible implementation manner, the inputting the data to be classified into the trained discriminant model, and obtaining the category of the data to be classified includes:

Inputting the image to be classified into the trained discriminant model to obtain the user, object or face attributes corresponding to the image to be classified;

Inputting the text to be classified into the trained discriminant model to obtain the emotional tendency, subject matter or technical field corresponding to the text to be classified;

Input the speech to be classified into the trained discriminant model to obtain the user, age group or emotion corresponding to the speech to be classified.

In another possible implementation manner, the training the discriminant model of the machine learning system according to the probability label includes:

The discriminative model is trained on the probabilistic labels by minimizing the noise-aware variant of the loss function of the discriminative model.

In another possible implementation manner, before labeling the data set with a labeling function λ _j , j=1, 2, ..., n, the method further includes:

filling missing values in said data set; and/or

Correct for outliers in the dataset.

A second aspect of the present application provides a data classification device applied to a machine learning system, the device comprising:

An acquisition module, used to acquire the data set {x _i | i=1, 2, ..., m} to be labeled;

A labeling module, configured to label the data set with a labeling function λ _j , j=1, 2, ..., n, to obtain the initial label Λ _{i, j} = λ _j ( _xi ) of the data set, i=1, 2, ..., m, j = 1, 2, ..., n;

A building module for calculating the pairwise correlation of the identification function according to the initial label, and constructing a generative model of the labeling function according to the pairwise correlation;

An estimation module, configured to estimate the probability label of the data set according to the generation model;

A training module, configured to train the discriminant model of the machine learning system according to the probability label to obtain the trained discriminant model;

A classification module, configured to input the data to be classified into the trained discriminant model to obtain the category of the data to be classified.

In another possible implementation, the generation model is:

Where Λ represents the initial label matrix formed by the initial label, Y represents the real label matrix, Z _w is a normalization constant, φ _i (Λ, y _i ), i=1, 2, ..., m is the is the pairwise correlation of the labeling function of each data in the data set, w is an undetermined parameter of the generation model, w∈R ^2n+|C| .

In another possible implementation, the pairwise correlation is:

Among them, II{Λ _{i, j} =Λ _{i, k} } indicates the value when the condition in brackets {} is satisfied or not.

In another possible implementation, the data set to be labeled is an image set, and the data to be classified is an image to be classified; or

The data set to be labeled is a text set, and the data to be classified is a text to be classified; or

The data set to be marked is a speech set, and the data to be classified is a speech to be classified.

In another possible implementation manner, the inputting the data to be classified into the trained discriminant model, and obtaining the category of the data to be classified includes:

Inputting the image to be classified into the trained discriminant model to obtain the user, object or face attributes corresponding to the image to be classified;

Inputting the text to be classified into the trained discriminant model to obtain the emotional tendency, subject matter or technical field corresponding to the text to be classified;

Input the speech to be classified into the trained discriminant model to obtain the user, age group or emotion corresponding to the speech to be classified.

In another possible implementation manner, the training the discriminant model of the machine learning system according to the probability label includes:

The discriminative model is trained on the probabilistic labels by minimizing the noise-aware variant of the loss function of the discriminative model.

In another possible implementation manner, the device further includes:

A preprocessing module is used for filling missing values in the data set, and/or correcting outliers in the data set.

A third aspect of the present application provides a computer device, the computer device includes a processor, and the processor is configured to implement the data classification method when executing a computer program stored in a memory.

A fourth aspect of the present application provides a computer storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the data classification method is implemented.

The present invention obtains the data set {x _i | i=1, 2, ..., m} to be marked; the data set is marked by the marking function λ _j , j = 1, 2, ..., n, Obtain the initial label Λ _{i, j} = λ _j ( _xi ), i=1, 2, ..., m, j = 1, 2, ..., n of the data set; calculate according to the initial label The pairwise correlation of the identification function, constructing the generation model of the label function according to the pairwise correlation; predicting the probability label of the data set according to the generation model; The discriminant model of the learning system is trained to obtain the trained discriminant model; the data to be classified is input into the trained discriminant model to obtain the category of the data to be classified. The invention can quickly generate the training data required by the discriminant model of the machine learning system, solve the technical problems of difficulty in obtaining manual labeling training data, long labeling time, and unguaranteed accuracy, reduce the workload of manual labeling training data, and improve The labeling efficiency and accuracy of the training data, using the training data, can quickly train the discriminant model, and use the discriminant model to realize fast and accurate data classification.

Description of drawings

Fig. 1 is a flowchart of a data classification method provided by an embodiment of the present invention.

Fig. 2 is a structural diagram of a data classification device provided by an embodiment of the present invention.

Fig. 3 is a schematic diagram of a computer device provided by an embodiment of the present invention.

Detailed ways

In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

Many specific details are set forth in the following description to facilitate a full understanding of the present invention, and the described embodiments are only some of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field of the invention. The terms used herein in the description of the present invention are for the purpose of describing specific embodiments only, and are not intended to limit the present invention.

Preferably, the data classification method of the present invention is applied in one or more computer devices. The computer device is a device that can automatically perform numerical calculations and/or information processing according to preset or stored instructions, and its hardware includes but not limited to microprocessors, Application Specific Integrated Circuits (ASICs), Programmable gate array (Field-Programmable Gate Array, FPGA), digital processor (Digital Signal Processor, DSP), embedded devices, etc.

The computer device may be a computing device such as a desktop computer, a notebook, a palmtop computer, or a cloud server. The computer device can perform human-computer interaction with the user through keyboard, mouse, remote controller, touch panel or voice control equipment.

Embodiment one

FIG. 1 is a flowchart of a data classification method provided by Embodiment 1 of the present invention. The data classification method is applied to a computer device.

The data classification method of the present invention is applied to a machine learning system for generating training data, using the training data to train a discriminant model of the machine learning system, and using the trained discriminant model to classify the data to be classified. The data classification method can quickly generate the training data required by the discriminant model of the machine learning system, solve the technical problems that manual labeling training data is difficult to obtain, the labeling time is long, and the accuracy rate cannot be guaranteed, and the work of manual labeling training data can be reduced Quantity, improve the labeling efficiency and accuracy of the training data, use the training data to quickly train the discriminant model, and use the discriminant model to achieve fast and accurate data classification.

As shown in Figure 1, the data classification methods include:

Step 101, obtain the data set {x _i |i=1, 2, . . . , m} to be labeled.

The data set to be labeled includes multiple data that need to be labeled. After each data in the dataset is labeled, the label of the data is obtained, which is used to train the discriminant model.

In an application scenario where the data classification method is applied to image classification, the data set to be marked may be an image set. For example, the dataset to be labeled includes images of different users, and the user corresponding to each image needs to be labeled. As another example, the data set to be labeled includes images of different objects (such as pens, balls, and books), and the objects corresponding to each image need to be labeled. For another example, the data set to be labeled includes images of different face attributes (such as race, gender, age, and expression), and the face attributes corresponding to each image need to be labeled.

In an application scenario where the data classification method is applied to text classification, the data set to be labeled may also be a text set. For example, the data set to be labeled includes texts with different emotional tendencies, and the corresponding emotional tendencies of each text should be marked. As another example, the data set to be marked includes texts of different subjects, and the subject matter corresponding to each text needs to be marked. As another example, the data set to be labeled includes texts in different technical fields (such as physics, chemistry, and machinery), and the technical field corresponding to each text needs to be marked.

In an application scenario where the data classification method is applied to speech classification, the data set to be marked may also be a speech set. For example, the data set to be labeled includes voices of multiple different users, and the user corresponding to each voice needs to be labeled. As another example, the data set to be marked includes voices of users of multiple different age groups, and the age group corresponding to each voice needs to be marked. For another example, the data set to be marked includes multiple voices with different emotions, and the emotion corresponding to each voice needs to be marked.

Data can be collected in real time, and the data collected in real time constitutes a data set to be marked. For example, images of people can be captured in real time, and each of the images of people captured in real time can be used as the data set to be labeled.

Alternatively, the dataset to be labeled can be obtained from a preset data source. For example, data can be obtained from a preset database to form a dataset to be labeled, and the preset database can pre-store a large amount of data, such as images.

Alternatively, a data set to be labeled input by a user may be received. For example, multiple images input by the user may be received, and the multiple images input by the user may be used as the data set to be labeled.

Step 102, label the data set { _xi |i=1, 2, ..., m} through the labeling function λ _j , j=1, 2, ..., n, and obtain the data set { x _i |i=1, 2, ..., m}'s initial label Λ _{i, j} = λ _j ( _xi ), where i = 1, 2, ..., m, j = 1, 2,. . . ., n.

The labeling function is a function that represents the mapping relationship between data and labels. The labeling function receives data and outputs the label of the data. The annotation function is a black-box function, which can be expressed as λ: Where λ represents the labeling function, X represents the data, and Y represents the initial label corresponding to X, Indicates that the callout function abstains.

Compared with manually annotating training data, the labeling function allows to utilize various weakly supervised source information (such as heuristic information, external knowledge base, etc.) to generate the initial labels. For example, there are two people A and B in the image, and it is necessary to mark the relationship between the two people. It is known that A is Ding’s father and B is Ding’s mother. According to the heuristic information “A is C’s father, B is C’s mother- > A and B are husband and wife”, then get the labeling result (that is, the initial label) that A and B are husband and wife.

Labeling functions do not require accuracy. That is, the initial labels obtained from the labeling function are unreliable. Unreliability can include incorrect labeling, multiple labeling, insufficient labeling, partial labeling, etc.

Multiple labeling functions can be defined in advance as required, for example, six labeling functions are defined.

Different labeling functions are allowed to conflict with the labeling results of the same data. For example, labeling function 1 labels the data as husband and wife, and labeling function 2 labels the data as siblings.

Step 103, calculate the pairwise correlation of the identification function λ _j , j=1, 2, . . . , n according to the initial label, and construct the labeling function λ _j , j= Generative models for 1, 2, ..., n.

The pairwise correlation of labeling functions λ _j , j=1, 2, . . . , n refers to the dependency relationship between two labeling functions. This method models the statistical dependencies between labeled functions to improve predictive performance. For example, if two labeling functions express similar heuristic information, it is possible to include this dependency in the generative model and avoid the "double counting" problem. This pairwise correlation is the most common, so the set C of label function pairs (j, k) is chosen to be modeled as correlated.

In this embodiment, the pairwise correlation of labeling function λ _j , j=1, 2, . . . , n can be expressed as:

Among them, Λ is the initial label matrix formed by the initial label, Λ _{i, j} = λ _j ( _xi ); Y is the real label matrix, Y = (y ₁ , y ₂ ,..., y _m ); C is a collection of label function pairs (i, j). II{Λ _{i, j} = Λ _{i, k} } indicates the value when the conditions in brackets {} are true or not. In this embodiment, the value is 1 when the condition in brackets {} is satisfied, and the value is 0 when the condition in brackets {} is not satisfied.

A generative model of the labeling function λ _j , j=1, 2, . . . , n can be constructed according to the pairwise correlation.

The core operation of the method is to model and integrate the noise signals provided by a set of labeling functions, modeling each labeling function as a noise "voter" that generates errors relative to other labeling functions.

For each data x _i in the data set {x _i |i=1, 2, ..., m}, the initial label vector is Λ _i = (Λ _{i, 1} , Λ _{i, 2,} , ..., Λi _,n ). In this embodiment, the generative model constructed according to the pairwise correlation is:

where Z _w is a normalization constant, φ _i (Λ, y _i ) is the pairwise correlation of the labeling function λ _j , j=1, 2, . . . , n for each data in the data set , w is an undetermined parameter of the generative model, w∈R ^2n+|C| .

Step 104, estimate the probability label of the data set { _xi |i=1, 2, ..., m} according to the generative model

In this example, the generated model is:

To learn the model without access to the true labels, the negative log marginal probability can be minimized given the observed initial label matrix Λ:

This objective can be optimized by interleaving stochastic gradient descent steps with Gibbs sampling steps, resulting in estimates of the parameters w of the generative model

the estimated value After the determination, the generation model is determined, and the initial label matrix Λ is input into the generation model to obtain the probability labels of the data set { _xi |i=1, 2, ..., m}

Steps 103-104 are to use the generative model to denoise the initial label Λ _{i, j} = λ _j ( _xi ), and obtain the data set { _xi |i=1, 2, ..., m} Probability label. The data set with the probability label is the training data of the machine learning system.

Step 105: Train the discriminant model of the machine learning system according to the probability label to obtain the trained discriminant model.

Training the discriminant model of the machine learning system according to the probability label is to use the data set with the probability label as a training sample to train the discriminant model.

According to the probability label The discriminative model is trained with the ultimate goal of training a discriminative model that exceeds the information expressed by the labeling function. The noise-aware variable can be minimized by minimizing the loss function l(h _θ ( _xi ), y) of the discriminative model (i.e. relative to the expected loss) in the probability label Train the discriminative model h _θ on:

When the discriminant model is trained, parameters of the discriminant model are adjusted to make the noise perception variable obtain a minimum value. The training process can use the RMSprop algorithm. RMSprop is a modified stochastic gradient descent algorithm. The RMSprop algorithm is a well-known technology in the art, and will not be repeated here.

Step 106, input the data to be classified into the trained discriminant model to obtain the category of the data to be classified.

In an application scenario where the data classification method is applied to image classification, the image to be classified is input into the trained discriminant model to obtain the category of the image to be classified. For example, the user corresponding to the image to be classified is obtained. In another example, an object corresponding to the image to be classified is obtained. For another example, the face attribute corresponding to the image to be classified is obtained.

In an application scenario where the data classification method is applied to text classification, the text to be classified is input into the trained discriminant model to obtain the category of the text to be classified. For example, the sentiment orientation (for example, positive sentiment orientation or negative sentiment orientation) of the text to be classified is obtained. In another example, the subject matter corresponding to the text to be classified is obtained. As another example, the technical field of the text to be classified is obtained.

In an application scenario where the method for generating training data is applied to speech classification, the speech to be classified is input into the trained discriminant model to obtain the category of the speech to be classified. For example, the user corresponding to the speech to be classified is obtained. In another example, the age group corresponding to the speech to be classified is obtained. For another example, the emotion of the speech to be classified is obtained.

The data classification method in Embodiment 1 obtains the data set {x _i |i=1, 2, ..., m} to be labeled; through the labeling function λ _j , j = 1, 2, ..., n pairs the The data set is labeled to obtain the initial label Λ _{i, j} = λ _j ( _xi ), i=1, 2,..., m, j=1, 2,..., n of the data set; according to The initial label calculates the pairwise correlation of the identification function, constructs the generation model of the label function according to the pairwise correlation; estimates the probability label of the data set according to the generation model; according to the probability The label trains the discriminant model of the machine learning system to obtain the trained discriminant model; input the data to be classified into the trained discriminant model to obtain the category of the to-be-classified data. This embodiment can quickly generate the training data required by the discriminant model of the machine learning system, solve the technical problems of manual labeling training data acquisition difficulty, long labeling time, and accuracy rate cannot be guaranteed, and reduce the workload of manual labeling training data. The labeling efficiency and accuracy of training data are improved, and a discriminant model can be quickly trained by using the training data, and fast and accurate data classification can be realized by using the discriminant model.

In another embodiment, before labeling the data set { _xi |i=1, 2, ..., m} by the labeling function λ _j , the method may further include: labeling the data set {x _i |i=1, 2, . . . , m} for preprocessing.

Preprocessing the dataset {x _i | _i =1, 2, ..., m} may include filling in missing value.

The K-nearest neighbor algorithm can be used to determine the K data closest to the data with missing values (for example, according to the Euclidean distance to determine the K data closest to the data with missing values), and the weighted average of the values of the K data to estimate the The missing value of the data.

Alternatively, a prediction model can be used to predict missing values. If the missing value is numerical, the missing value can be filled with the mean value, and if the missing value is non-numeric, the missing value can be filled with the mode.

Alternatively, the mean method can be used to replace missing values. Preferably, since the method of using the mean method to replace missing values is based on the assumption of complete random missing, the variance and standard deviation of the data will become smaller. Therefore, the method may also include: replacing the value obtained by the mean value The filling value is multiplied with the preset expansion factor to obtain new data as the final filling value. The preset expansion factor is a preset expansion factor, and the expansion factor is greater than 1.

Other methods can also be used to fill in the missing values. For example, missing values can be filled by regression fitting or interpolation.

Preprocessing the data set { _xi |i=1, 2, ..., m} may also include correcting the data set { _xi |i = 1, 2, ..., m} outliers. Outliers are values that deviate significantly from the rest of the data.

The method of correcting outliers can be the same as the method of filling missing values. For example, the K-nearest neighbor algorithm can be used to determine the K data closest to the data with abnormal values (for example, according to the Euclidean distance to determine the K data closest to the data with abnormal values), and the weighted average of the values of the K data to Estimates the correction value for outliers in the data. Alternatively, a prediction model may be used to predict the correction value. If the outlier is numerical, the outlier may be corrected by the mean value, and if the outlier is non-numerical, the outlier may be corrected by the mode.

Alternatively, the mean method can be used to replace outliers. Preferably, since the method of using the mean value method to replace the outliers is based on the assumption of complete missing at random, the variance and standard deviation of the data will become smaller, therefore, the method may also include: replacing the mean value obtained by The correction value is multiplied by the preset expansion factor to obtain a new value as the final correction value. The preset expansion factor is a preset expansion factor, and the expansion factor is greater than 1.

Other methods can also be used to correct the outliers. For example, outliers can be corrected by a regression fitting method or an interpolation method.

The method of correcting outliers can also be different from the method of filling missing values.

Preprocessing the data set may also include directly discarding data with missing values and data with possible outliers. Directly discarding data with missing values and data with outliers can ensure the cleanliness of the data set.

Embodiment two

Fig. 2 is a structural diagram of a data classification device provided by Embodiment 2 of the present invention. The data classification device 20 is applied to a machine learning system for generating training data, using the training data to train a discriminant model of the machine learning system, and using the trained discriminant model to classify the data to be classified. The data classification device 20 can quickly generate the training data required by the discriminant model of the machine learning system, solve the technical problems that the manual labeling training data is difficult to obtain, the labeling time is long, and the accuracy rate cannot be guaranteed, and the manual labeling training data is reduced. workload, improve the labeling efficiency and accuracy of the training data, use the training data to quickly train the discriminant model, and use the discriminant model to achieve fast and accurate data classification.

As shown in FIG. 2 , the data classification apparatus 20 may include an acquisition module 201 , a labeling module 202 , a construction module 203 , an estimation module 204 , a training module 205 , and a classification module 206 .

The obtaining module 201 is used to obtain the data set { _xi |i=1, 2, . . . , m} to be labeled.

The data set to be labeled includes multiple data that need to be labeled. After each data in the dataset is labeled, the label of the data is obtained, which is used to train the discriminant model.

In an application scenario where the data classification method is applied to image classification, the data set to be marked may be an image set. For example, the dataset to be labeled includes images of different users, and the user corresponding to each image needs to be labeled. As another example, the data set to be labeled includes images of different objects (such as pens, balls, and books), and the objects corresponding to each image need to be labeled. For another example, the dataset to be labeled includes images of different face attributes (such as race, gender, age, and expression), and the face attributes corresponding to each image need to be labeled.

In an application scenario where the data classification method is applied to text classification, the data set to be labeled may also be a text set. For example, the data set to be labeled includes texts with different emotional tendencies, and the corresponding emotional tendencies of each text should be marked. As another example, the data set to be marked includes texts of different subjects, and the subject matter corresponding to each text needs to be marked. As another example, the data set to be labeled includes texts in different technical fields (such as physics, chemistry, and machinery), and the technical field corresponding to each text needs to be marked.

In an application scenario where the data classification method is applied to speech classification, the data set to be marked may also be a speech set. For example, the data set to be labeled includes voices of multiple different users, and the user corresponding to each voice needs to be labeled. As another example, the data set to be marked includes voices of users of multiple different age groups, and the age group corresponding to each voice needs to be marked. For another example, the data set to be marked includes multiple voices with different emotions, and the emotion corresponding to each voice needs to be marked.

Data can be collected in real time, and the data collected in real time constitutes a data set to be marked. For example, images of people can be captured in real time, and each of the images of people captured in real time can be used as the data set to be labeled.

Alternatively, the dataset to be labeled can be obtained from a preset data source. For example, data can be obtained from a preset database to form a dataset to be labeled, and the preset database can pre-store a large amount of data, such as images.

Alternatively, a data set to be labeled input by a user may be received. For example, multiple images input by the user may be received, and the multiple images input by the user may be used as the data set to be labeled.

The labeling module 202 is used to label the data set { _xi |i=1, 2, ..., m} through the labeling function λ _j to obtain the data set { _xi |i=1, 2, ..., m}'s initial label Λ _{i, j} = λ _j ( _xi ), where i=1, 2, ..., m, j = 1, 2, ..., n.

The labeling function is a function that represents the mapping relationship between data and labels. The labeling function receives data and outputs the label of the data. The annotation function is a black-box function, which can be expressed as λ: Where λ represents the labeling function, X represents the data, and Y represents the initial label corresponding to X, Indicates that the callout function abstains.

Compared with manually annotating training data, the labeling function allows to utilize various weakly supervised source information (such as heuristic information, external knowledge base, etc.) to generate the initial labels. For example, there are two people A and B in the image, and it is necessary to mark the relationship between the two people. It is known that A is Ding’s father and B is Ding’s mother. According to the heuristic information “A is C’s father, B is C’s mother- > A and B are husband and wife”, then get the labeling result (that is, the initial label) that A and B are husband and wife.

Labeling functions do not require accuracy. That is, the initial labels obtained from the labeling function are unreliable. Unreliability can include incorrect labeling, multiple labeling, insufficient labeling, partial labeling, etc.

Multiple labeling functions can be defined in advance as required, for example, six labeling functions are defined.

Different labeling functions are allowed to conflict with the labeling results of the same data. For example, labeling function 1 labels the data as husband and wife, and labeling function 2 labels the data as siblings.

A construction module 203, configured to calculate the pairwise correlation of the identification function λ _j , j=1, 2, ..., n according to the initial label, and construct the label function λ _j according to the pairwise correlation , a generative model for j=1, 2, . . . , n.

The pairwise correlation of labeling functions λ _j , j=1, 2, . . . , n refers to the dependency relationship between two labeling functions. This method models the statistical dependencies between labeled functions to improve predictive performance. For example, if two labeling functions express similar heuristic information, it is possible to include this dependency in the generative model and avoid the "double counting" problem. This pairwise correlation is the most common, so the set C of label function pairs (j, k) is chosen to be modeled as correlated.

In this embodiment, the pairwise correlation of labeling function λ _j , j=1, 2, . . . , n can be expressed as:

Among them, Λ is the initial label matrix formed by the initial label, Λ _{i, j} = λ _j ( _xi ); Y is the real label matrix, Y = (y ₁ , y ₂ ,..., y _m ); C is a collection of label function pairs (i, j). II{Λ _{i, j} = Λ _{i, k} } indicates the value when the conditions in brackets {} are true or not. In this embodiment, the value is 1 when the condition in brackets {} is satisfied, and the value is 0 when the condition in brackets {} is not satisfied.

A generative model of the labeling function λ _j , j=1, 2, . . . , n can be constructed according to the pairwise correlation.

The core operation of the method is to model and integrate the noise signals provided by a set of labeling functions, modeling each labeling function as a noise "voter" that generates errors relative to other labeling functions.

For each data x _i in the data set {x _i |i=1, 2, ..., m}, the initial label vector is Λ _i = (Λ _{i, 1} , Λ _{i, 2,} , ..., Λi _,n ). In this embodiment, the generative model constructed according to the pairwise correlation is:

where Z _w is a normalization constant, φ _i (Λ, y _i ) is the pairwise correlation of the labeling function λ _j , j=1, 2, . . . , n for each data in the data set , w is an undetermined parameter of the generative model, w∈R ^2n+|C| .

Estimation module 204, configured to estimate the probability label of the data set { _xi |i=1, 2, ..., m} according to the generative model

In this example, the generated model is:

To learn the model without access to the true labels, the negative log marginal probability can be minimized given the observed initial label matrix Λ:

This objective can be optimized by interleaving stochastic gradient descent steps with Gibbs sampling steps, resulting in estimates of the parameters w of the generative model

the estimated value After the determination, the generation model is determined, and the initial label matrix Λ is input into the generation model to obtain the probability labels of the data set { _xi |i=1, 2, ..., m}

The construction module 203 and the estimation module 204 use the generative model to denoise the initial label Λ _i,j =λ _j ( _xi ), and obtain the data set { _xi |i=1,2,... , the probability label for m}. The data set with the probability label is obtained training data of the machine learning system.

The training module 205 is configured to train the discriminant model of the machine learning system according to the probability label to obtain the trained discriminant model.

Training the discriminant model of the machine learning system according to the probability label is to use the data set with the probability label as a training sample to train the discriminant model.

According to the probability label The discriminative model is trained with the ultimate goal of training a discriminative model that exceeds the information expressed by the labeling function. The noise-aware variable can be minimized by minimizing the loss function l(h _θ ( _xi ), y) of the discriminative model (i.e. relative to the expected loss) in the probability label Train the discriminative model h _θ on:

When the discriminant model is trained, parameters of the discriminant model are adjusted to make the noise perception variable obtain a minimum value. The training process can use the RMSprop algorithm. RMSprop is a modified stochastic gradient descent algorithm. The RMSprop algorithm is a well-known technology in the art, and will not be repeated here.

The classification module 206 is configured to input the data to be classified into the trained discriminant model to obtain the category of the data to be classified.

In an application scenario where the data classification method is applied to image classification, the image to be classified is input into the trained discriminant model to obtain the category of the image to be classified. For example, the user corresponding to the image to be classified is obtained. In another example, an object corresponding to the image to be classified is obtained. For another example, the face attribute corresponding to the image to be classified is obtained.

In an application scenario where the data classification method is applied to text classification, the text to be classified is input into the trained discriminant model to obtain the category of the text to be classified. For example, the sentiment orientation (for example, positive sentiment orientation or negative sentiment orientation) of the text to be classified is obtained. In another example, the subject matter corresponding to the text to be classified is obtained. As another example, the technical field of the text to be classified is obtained.

In an application scenario where the method for generating training data is applied to speech classification, the speech to be classified is input into the trained discriminant model to obtain the category of the speech to be classified. For example, the user corresponding to the speech to be classified is obtained. In another example, the age group corresponding to the speech to be classified is obtained. For another example, the emotion of the speech to be classified is obtained.

The data classification device 20 of the second embodiment obtains the data set {x _i |i=1, 2, ..., m} to be labeled; through the labeling function λ _j , j = 1, 2, ..., n Label the above data set to obtain the initial label Λ _{i, j} = λ _j ( _xi ), i=1, 2, ..., m, j = 1, 2, ..., n of the data set; Calculate the pairwise correlation of the identification function according to the initial label, construct the generation model of the label function according to the pairwise correlation; estimate the probability label of the data set according to the generation model; The probability label trains the discriminant model of the machine learning system to obtain a trained discriminant model; input the data to be classified into the trained discriminant model to obtain the category of the to-be-classified data. This embodiment can quickly generate the training data required by the discriminant model of the machine learning system, solve the technical problems of manual labeling training data acquisition difficulty, long labeling time, and accuracy rate cannot be guaranteed, and reduce the workload of manual labeling training data. The labeling efficiency and accuracy of training data are improved, and a discriminant model can be quickly trained by using the training data, and fast and accurate data classification can be realized by using the discriminant model.

In another embodiment, the data classification apparatus 20 may further include: a preprocessing module, configured to preprocess the data set { _xi |i=1, 2, . . . , m}.

Preprocessing the dataset {x _i | _i =1, 2, ..., m} may include filling in missing value.

The K-nearest neighbor algorithm can be used to determine the K data closest to the data with missing values (for example, according to the Euclidean distance to determine the K data closest to the data with missing values), and the weighted average of the values of the K data to estimate the The missing value of the data.

Alternatively, a prediction model can be used to predict missing values. If the missing value is numerical, the missing value can be filled with the mean value, and if the missing value is non-numeric, the missing value can be filled with the mode.

Alternatively, the mean method can be used to replace missing values. Preferably, since the method of using the mean method to replace missing values is based on the assumption of complete random missing, the variance and standard deviation of the data will become smaller. Therefore, the method may also include: replacing the value obtained by the mean value The filling value is multiplied with the preset expansion factor to obtain new data as the final filling value. The preset expansion factor is a preset expansion factor, and the expansion factor is greater than 1.

Other methods can also be used to fill in the missing values. For example, missing values can be filled by regression fitting or interpolation.

Preprocessing the data set { _xi |i=1, 2, ..., m} may also include correcting the data set { _xi |i = 1, 2, ..., m} outliers. Outliers are values that deviate significantly from the rest of the data.

The method of correcting outliers can be the same as the method of filling missing values. For example, the K-nearest neighbor algorithm can be used to determine the K data closest to the data with abnormal values (for example, according to the Euclidean distance to determine the K data closest to the data with abnormal values), and the weighted average of the values of the K data to Estimates the correction value for outliers in the data. Alternatively, a prediction model may be used to predict the correction value. If the outlier is numerical, the outlier may be corrected by the mean value, and if the outlier is non-numerical, the outlier may be corrected by the mode.

Alternatively, the mean method can be used to replace outliers. Preferably, since the method of using the mean method to replace outliers is based on the assumption of complete missing at random, it will cause the variance and standard deviation of the data to become smaller. Therefore, the method may also include: The correction value is multiplied by the preset expansion factor to obtain a new value as the final correction value. The preset expansion factor is a preset expansion factor, and the expansion factor is greater than 1.

Other methods can also be used to correct the outliers. For example, outliers can be corrected by a regression fitting method or an interpolation method.

The method of correcting outliers can also be different from the method of filling missing values.

Preprocessing the data set may also include directly discarding data with missing values and data with possible outliers. Directly discarding data with missing values and data with outliers can ensure the cleanliness of the data set.

Embodiment three

This embodiment provides a computer storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above embodiment of the data classification method are implemented, such as steps 101-106 shown in FIG. 1:

Step 101, obtain the data set {x _i |i=1, 2, ..., m} to be labeled;

Step 102, label the data set with the labeling function λ _j , j=1, 2, ..., n, and obtain the initial label Λ _{i, j} = λ _j ( _xi ), i= 1, 2, ..., m, j = 1, 2, ..., n;

Step 103, calculating the pairwise correlation of the labeling function according to the initial label, and constructing a generative model of the labeling function according to the pairwise correlation;

Step 104, estimating the probability label of the data set according to the generation model;

Step 105, training the discriminant model of the machine learning system according to the probability label to obtain the trained discriminant model;

Step 106, input the data to be classified into the trained discriminant model to obtain the category of the data to be classified.

Or, when the computer program is executed by the processor, the functions of the modules in the above-mentioned device embodiments are implemented, such as modules 201-206 in FIG. 2:

An acquisition module 201, configured to acquire a dataset to be labeled { _xi | i=1, 2, ..., m};

Annotation module 202, configured to annotate the data set with an annotation function λ _j , j=1, 2, ..., n, to obtain an initial label Λ _{i, j} = λ _j ( _xi ) of the data set , i=1, 2, ..., m, j = 1, 2, ..., n;

A construction module 203, configured to calculate the pairwise correlation of the identification function according to the initial label, and construct a generation model of the labeling function according to the pairwise correlation;

An estimation module 204, configured to estimate the probability label of the data set according to the generation model;

A training module 205, configured to train a discriminant model of the machine learning system according to the probability label to obtain a trained discriminant model;

The classification module 206 is configured to input the data to be classified into the trained discriminant model to obtain the category of the data to be classified.

Embodiment four

FIG. 3 is a schematic diagram of a computer device provided by Embodiment 4 of the present invention. The computer device 30 includes a memory 301 , a processor 302 and a computer program 303 stored in the memory 301 and executable on the processor 302 , such as a data classification program. When the processor 302 executes the computer program 303, the steps in the above-mentioned data classification method embodiment are implemented, for example, steps 101-106 shown in FIG. 1:

Step 101, obtain the data set {x _i |i=1, 2, ..., m} to be labeled;

Step 102, label the data set with the labeling function λ _j , j=1, 2, ..., n, and obtain the initial label Λ _{i, j} = λ _j ( _xi ), i= 1, 2, ..., m, j = 1, 2, ..., n;

Step 103, calculating the pairwise correlation of the labeling function according to the initial label, and constructing a generative model of the labeling function according to the pairwise correlation;

Step 104, estimating the probability label of the data set according to the generation model;

Step 105, training the discriminant model of the machine learning system according to the probability label to obtain the trained discriminant model;

Step 106, input the data to be classified into the trained discriminant model to obtain the category of the data to be classified.

Or, when the computer program is executed by the processor, the functions of the modules in the above-mentioned device embodiments are implemented, such as modules 201-206 in FIG. 2:

An acquisition module 201, configured to acquire a data set to be labeled { _xi |i=1, 2, ..., m};

Annotation module 202, configured to annotate the data set with an annotation function λ _j , j=1, 2, ..., n, to obtain an initial label Λ _{i, j} = λ _j ( _xi ) of the data set , i=1, 2, ..., m, j = 1, 2, ..., n;

A construction module 203, configured to calculate the pairwise correlation of the identification function according to the initial label, and construct a generation model of the labeling function according to the pairwise correlation;

An estimation module 204, configured to estimate the probability label of the data set according to the generation model;

A training module 205, configured to train a discriminant model of the machine learning system according to the probability label to obtain a trained discriminant model;

The classification module 206 is configured to input the data to be classified into the trained discriminant model to obtain the category of the data to be classified.

Exemplarily, the computer program 303 can be divided into one or more modules, and the one or more modules are stored in the memory 301 and executed by the processor 302 to complete the method. The one or more modules may be a series of computer program instruction segments capable of accomplishing specific functions, and the instruction segments are used to describe the execution process of the computer program 303 in the computer device 30 . For example, the computer program 303 can be divided into the acquisition module 201, labeling module 202, construction module 203, estimation module 204, training module 205, and classification module 206 in FIG.

The computer device 30 may be a computing device such as a desktop computer, a notebook, a palmtop computer, or a cloud server. Those skilled in the art can understand that the schematic diagram 3 is only an example of the computer device 30, and does not constitute a limitation to the computer device 30. It may include more or less components than those shown in the figure, or combine certain components, or be different. For example, the computer device 30 may also include input and output devices, network access devices, buses, and the like.

The so-called processor 302 may be a central processing unit (Central Processing Unit, CPU), and may also be other general-purpose processors, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), Off-the-shelf programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor 302 can also be any conventional processor, etc., the processor 302 is the control center of the computer device 30, and uses various interfaces and lines to connect the entire computer device 30. various parts.

The memory 301 can be used to store the computer program 303, and the processor 302 implements the computer device by running or executing the computer program or module stored in the memory 301 and calling the data stored in the memory 301 30 various functions. The memory 301 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playback function, an image playback function, etc.) and the like; the storage data area can store Data created according to use of the computer device 30 (such as audio data, a phone book, etc.) and the like are stored. In addition, the memory 301 may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, flash memory card (Flash Card), at least one magnetic disk storage device, flash memory device, or other volatile solid state storage devices.

If the integrated modules of the computer device 30 are realized in the form of software function modules and sold or used as independent products, they can be stored in a computer storage medium. Based on this understanding, the present invention realizes all or part of the processes in the methods of the above embodiments, and can also be completed by instructing related hardware through a computer program. The computer program can be stored in a computer storage medium, and the computer program is in When executed by the processor, the steps in the above-mentioned various method embodiments can be realized. Wherein, the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM, Read-Only Memory) , random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, computer-readable media Excludes electrical carrier signals and telecommunication signals.

In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or may also be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, each module may exist separately physically, or two or more modules may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware, or in the form of hardware plus software function modules.

The above-mentioned integrated modules realized in the form of software function modules may be stored in a computer storage medium. The above-mentioned software function modules are stored in a storage medium, and include several instructions to make a computer device (which may be a personal computer, server, or network device, etc.) or a processor (processor) execute the methods described in various embodiments of the present invention. partial steps.

It will be apparent to those skilled in the art that the invention is not limited to the details of the above-described exemplary embodiments, but that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Accordingly, the embodiments should be regarded in all points of view as exemplary and not restrictive, the scope of the invention being defined by the appended claims rather than the foregoing description, and it is therefore intended that the scope of the invention be defined by the appended claims rather than by the foregoing description. All changes within the meaning and range of equivalents of the elements are embraced in the present invention. Any reference sign in a claim should not be construed as limiting the claim concerned. Furthermore, it is clear that the word "comprising" does not exclude other modules or steps, and the singular does not exclude the plural. Multiple modules or devices stated in the system claims may also be realized by one module or device through software or hardware. The words first, second, etc. are used to denote names and do not imply any particular order.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent replacements can be made without departing from the spirit and scope of the technical solutions of the present invention.

Claims (10)

1. A data classification method applied to a machine learning system, wherein the method comprises: Obtain the dataset to be labeled {x _i |i=1,2,...,m}; The data set is marked by the labeling function λ _j , j=1,2,...,n, and the initial label Λ _i,j =λ _j ( _xi ), i=1,2,... of the data set is obtained ,m,j=1,2,...,n; calculating the pairwise correlation of the identification function according to the initial label, and constructing a generative model of the labeling function according to the pairwise correlation; Estimating a probability label of the data set according to the generative model; training the discriminant model of the machine learning system according to the probability label to obtain a trained discriminant model; Inputting the data to be classified into the trained discriminant model to obtain the category of the data to be classified. 2. The method according to claim 1, wherein the generated model is: Where Λ represents the initial label matrix formed by the initial label, Y represents the real label matrix, Z _w is a normalization constant, φ _i (Λ, y _i ), i=1, 2,..., m are The pairwise correlation of the labeling function of each data in the set, w is an undetermined parameter of the generation model, w∈R ^2n+|C| . 3. The method of claim 2, wherein the pairwise correlation is: in Indicates the value when the condition in brackets {} is true or not. 4. The method of claim 1, wherein: The data set to be marked is an image set, and the data to be classified is an image to be classified; or The data set to be labeled is a text set, and the data to be classified is a text to be classified; or The data set to be marked is a speech set, and the data to be classified is a speech to be classified. 5. The method according to claim 4, wherein the input of the data to be classified into the trained discriminant model, obtaining the category of the data to be classified comprises: Inputting the image to be classified into the trained discriminant model to obtain the user, object or face attributes corresponding to the image to be classified; Inputting the text to be classified into the trained discriminant model to obtain the emotional tendency, subject matter or technical field corresponding to the text to be classified; Input the speech to be classified into the trained discriminant model to obtain the user, age group or emotion corresponding to the speech to be classified. 6. The method according to claim 1, wherein the training of the discriminant model of the machine learning system according to the probability label comprises: The discriminative model is trained on the probabilistic labels by minimizing the noise-aware variant of the loss function of the discriminative model. 7. The method according to any one of claims 1-6, characterized in that, before the data set is marked by the marking function λ _j , j=1, 2,..., n, the method Also includes: filling missing values in said data set; and/or Correct for outliers in the dataset. 8. A data classification device applied to a machine learning system, characterized in that the device comprises: An acquisition module, used to acquire the data set {x _i |i=1,2,...,m} to be labeled; A labeling module, configured to label the data set with a labeling function λ _j , j=1,2,...,n, to obtain the initial label Λ _i,j =λ _j ( _xi ), i= 1,2,...,m, j=1,2,...,n; A building module for calculating the pairwise correlation of the identification function according to the initial label, and constructing a generative model of the labeling function according to the pairwise correlation; An estimation module, configured to estimate the probability label of the data set according to the generation model; A training module, configured to train the discriminant model of the machine learning system according to the probability label to obtain the trained discriminant model; A classification module, configured to input the data to be classified into the trained discriminant model to obtain the category of the data to be classified. 9. A computer device, characterized in that: the computer device comprises a processor, and the processor is configured to execute a computer program stored in a memory to realize the data classification method according to any one of claims 1-7. 10. A computer storage medium, wherein a computer program is stored on the computer storage medium, wherein when the computer program is executed by a processor, the data classification method according to any one of claims 1-7 is realized.