CN115905876A

CN115905876A - Model processing method, device and equipment

Info

Publication number: CN115905876A
Application number: CN202211725485.1A
Authority: CN
Inventors: 刘芳卿
Original assignee: Alipay Hangzhou Information Technology Co Ltd
Current assignee: Alipay Hangzhou Information Technology Co Ltd
Priority date: 2022-12-30
Filing date: 2022-12-30
Publication date: 2023-04-04

Abstract

The embodiment of the specification discloses a model processing method, a device and equipment, wherein the method comprises the following steps: performing model training on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, wherein the label information is obtained by labeling sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data; generating corresponding restored label information for the first sample data and second sample data respectively based on the trained label restoration model, wherein the second sample data is sample data which does not contain corresponding labeled label information or is abnormal in labeled label information; and performing model training on a target model applied to the target service based on the first sample data, the second sample data and the generated corresponding reduction label information to obtain a trained target model.

Description

Model processing method, device and equipment

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to a method, an apparatus, and a device for processing a model.

Background

With the expansion of computing resources, people are eagerly required to analyze more data to model and approach the actual development law so as to reduce unknown risks, however, under the condition that the labeling cost of sample data is higher and higher, enough sample data cannot be labeled in a short time generally, and meanwhile, many sample data cannot be labeled with non "0" or "1" through limited cognition, so that how to utilize the part of unlabeled or unlabeled sample data is very important. Taking the example that the data cannot be labeled with "1" or "0" through limited cognition, a decision boundary can still be learned in a direct labeling manner, however, the data cannot be well generalized to unknown data sets, and even the effect on the data sets is more influenced. Therefore, it is necessary to provide a technical solution that can restore the relatively real tag information of the sample data, so as to better utilize the large-scale sample data with the fuzzy tag information and improve the generalization of the model.

Disclosure of Invention

The purpose of the embodiments of the present specification is to provide a technical solution that can restore the relatively real tag information of sample data, so as to better utilize large-scale sample data with fuzzy tag information and improve the generalization of a model.

In order to implement the above technical solution, the embodiments of the present specification are implemented as follows:

the method for processing the model provided by the embodiment of the specification comprises the following steps: model training is carried out on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, wherein the label information is obtained by carrying out label processing on the sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data. And generating corresponding reduction label information for the first sample data and second sample data respectively based on the trained label reduction model, wherein the second sample data is sample data which does not contain corresponding labeling label information or is abnormal in labeling label information. And performing model training on a target model applied to the target service based on the first sample data, the second sample data and the generated corresponding reduction label information to obtain a trained target model.

The processing device of a model that this specification embodiment provided, the device includes: the first training module performs model training on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, wherein the label information is label information obtained by labeling sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data. And the label generation module is used for respectively generating corresponding reduced label information for the first sample data and second sample data based on the trained label reduction model, wherein the second sample data is sample data which does not contain corresponding labeled label information or is abnormal in labeled label information. And the second training module performs model training on the target model applied to the target service based on the first sample data, the second sample data and the generated corresponding reduction label information to obtain the trained target model.

The embodiment of the present specification provides a model processing device, where the model processing device includes: a processor; and a memory arranged to store computer executable instructions that, when executed, cause the processor to: model training is carried out on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, wherein the label information is obtained by carrying out label processing on the sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data. And generating corresponding reduced label information for the first sample data and second sample data respectively based on the trained label reduction model, wherein the second sample data is sample data which does not contain corresponding labeled label information or is abnormal in labeled label information. And performing model training on a target model applied to the target service based on the first sample data, the second sample data and the generated corresponding reduction label information to obtain a trained target model.

The present specification also provides a storage medium for storing computer executable instructions, which when executed by a processor implement the following procedures: model training is carried out on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, wherein the label information is obtained by carrying out label processing on the sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data. And generating corresponding reduced label information for the first sample data and second sample data respectively based on the trained label reduction model, wherein the second sample data is sample data which does not contain corresponding labeled label information or is abnormal in labeled label information. And performing model training on a target model applied to the target service based on the first sample data, the second sample data and the generated corresponding reduction label information to obtain a trained target model.

Drawings

In order to more clearly illustrate the embodiments of the present specification or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, it is obvious that the drawings in the following description are only some embodiments described in the present specification, and for those skilled in the art, other drawings can be obtained according to the drawings without any creative effort.

FIG. 1 illustrates an embodiment of a model processing method of the present disclosure;

FIG. 2 is a process flow diagram of another model embodiment of the present disclosure;

FIG. 3 is a flowchart of another embodiment of a model processing method;

FIG. 4 is a block diagram of an embodiment of a model processing apparatus according to the present disclosure;

FIG. 5 is an embodiment of a model processing device according to the present disclosure.

Detailed Description

The embodiment of the specification provides a model processing method, a model processing device and model processing equipment.

In order to make those skilled in the art better understand the technical solutions in the present specification, the technical solutions in the embodiments of the present specification will be clearly and completely described below with reference to the drawings in the embodiments of the present specification, and it is obvious that the described embodiments are only a part of the embodiments of the present specification, and not all of the embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present specification without making any creative effort shall fall within the protection scope of the present specification.

Example one

As shown in fig. 1, an execution subject of the method may be a terminal device or a server, where the terminal device may be a mobile terminal device such as a mobile phone and a tablet computer, a computer device such as a notebook computer or a desktop computer, or an IoT device (specifically, a smart watch and a vehicle-mounted device), where the server may be an independent server, a server cluster formed by a plurality of servers, and the server may be a backend server such as a financial service or an internet shopping service, or a backend server of an application. The method may specifically comprise the steps of:

in step S102, model training is performed on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, where the label information is obtained by performing label processing on the sample data in advance, and the label reduction model is used to generate corresponding label information for the sample data.

The first sample data may be sample data used for training the label reduction model, the first sample data may be related data obtained from a designated service, and the designated service may be any service, for example, the designated service may be a payment service, a transfer service, an instant messaging service, or the like, which may be specifically set according to an actual situation, and this is not limited in the embodiments of the present specification. The first sample data may be one or a plurality of sample data, and the first sample data may be positive sample data or may be negative sample data. The related data may be service data related to a specific service, and may be specifically determined according to the specific service, for example, the specific service is a payment service, the first sample data may include payment time, place, payment amount, account information of a payer, account information of a recipient, and the like, and in addition, the first sample data may include, in addition to the text data as described above, images, audio data, and the like, for example, images for performing biometric identification, specifically, facial images, fingerprint images, palm print images, iris images, and the like, which may be specifically set according to actual situations, and this is not limited in this specification. The label information may be label information obtained by labeling the sample data in advance, that is, the label information may be label information labeled in advance by a manual method or the like, for example, the specified service is a payment service, the first sample data includes payment time, place, payment amount, account information of a payer, account information of a receiver, a facial image, and the like, and the label information may be label information with a fraud risk or the like, and may be specifically set according to an actual situation, which is not limited in this specification. The label reduction model may be constructed by a plurality of different algorithms, for example, the label reduction model may be constructed by a neural network model, or the label reduction model may be constructed by other algorithms than the neural network model, which may be specifically set according to an actual situation, and this is not limited in this specification.

In implementation, with the expansion of computing resources, people are eager to analyze more data to model and approach the actual development law so as to reduce the unknown risk, however, under the condition that the labeling cost of sample data is higher and higher, enough sample data cannot be labeled in a short time generally, and meanwhile, many sample data cannot be labeled with non "0" or "1" through limited cognition, so that it is important how to utilize the part of unlabeled or unlabeled sample data. Taking the example that the data cannot be labeled with "1" or "0" through limited cognition, a decision boundary can still be learned in a direct labeling manner, however, the data cannot be well generalized to unknown data sets, and even the effect on the data sets is more influenced. If the sample data which is not marked or can not be marked is not used, and only the sample data with the specific marking information is used, the sample data distribution is incomplete, and the finally obtained model lacks generalization; if the unlabelled or unlabelled sample data is used as a single type of data, the distance between the unlabelled or unlabelled sample data and other data is pulled, so that the generalization of the model is influenced; if the sample data which is not labeled or cannot be labeled is endowed with the agent label information and updated, but only a certain degree of freedom is given to the class data in the model training process, but the degree is shallow, and the income is less, so that a technical scheme that the real label information of the sample data can be restored, the large-scale sample data with the fuzzy label information is better utilized, and the generalization of the model is improved is needed to be provided. The embodiments of the present disclosure provide an implementable technical solution, which may specifically include the following.

When a user triggers execution of a specified service, relevant data generated in the process of executing the specified service by the user can be acquired. Alternatively, in the process of performing model training on a certain model, a certain amount of data is acquired from related data (or from a specified historical database) recorded in advance in a specified service as first sample data, and the first sample data is labeled to obtain labeled label information corresponding to the first sample data. Then, a corresponding algorithm, for example, an algorithm corresponding to the neural network model, may be set in advance, and an initial label reduction model may be constructed through the algorithm, and a corresponding convergence condition may be set in advance for the label reduction model. The label reduction model can be subjected to model training by using the first sample data and the label information corresponding to the first sample data, whether the label reduction model is converged is judged according to the convergence condition, if not, the label information corresponding to the first sample data and the first sample data can be continuously obtained, the label reduction model can be subjected to model training by using the label information corresponding to the first sample data and the first sample data until the label reduction model is converged, and finally, the trained label reduction model can be obtained.

In step S104, corresponding restore tag information is generated for the first sample data and the second sample data respectively based on the trained tag restore model, and the second sample data is sample data that does not include corresponding label tagging information or has abnormal label tagging information.

The abnormal labeling label information may be that the labeling label information is obviously wrong or the labeling label information is empty, and the like, and may be specifically set according to an actual situation, which is not limited in the embodiments of the present specification. The second sample data may include one sample data or a plurality of sample data, and in this embodiment, in order to fully train the target model and the tag reduction model, the second sample data may include a plurality of sample data.

In implementation, after the trained label reduction model is obtained in the above manner, the trained label reduction model already has a preliminary label information reduction capability, and therefore, the first sample data may be input into the trained label reduction model, so that the trained label reduction model regenerates corresponding label information (i.e., the reduction label information corresponding to the first sample data) for the first sample data, and in addition, second sample data that does not include corresponding labeling label information or is abnormal in labeling label information may also be obtained and input into the trained label reduction model, so that the trained label reduction model generates corresponding label information (i.e., the reduction label information corresponding to the second sample data) for the second sample data.

In step S106, model training is performed on the target model applied to the target service based on the first sample data and the second sample data and the generated corresponding reduction label information, so as to obtain a trained target model.

The target service may be any service, for example, the target service may be a payment service, a transfer service, an instant messaging service, or the like, and may be specifically set according to an actual situation, which is not limited in this description embodiment. The target model may be any model, and may be specifically set according to a target service, for example, if the target service is a payment service, the target model may be a risk prevention and control model, and may be used to detect whether a fraud risk exists in a process of executing the payment service, or if the target service is an online shopping service, the target model may be an information recommendation model or an information retrieval model, and may be used to recommend or retrieve specified information to a user of the online shopping service, and the target model may be constructed by a plurality of different algorithms, for example, the target model may be constructed by a classification algorithm, or the target model may be constructed by a neural network model, or the target model may be constructed by a genetic algorithm, or the target model may be constructed by an ant colony algorithm, and the target model may be specifically set according to an actual situation, which is not limited in the embodiments of the present specification.

In implementation, in order to make the effect of the label reduction model better and train the model in the specific service by better utilizing large-scale sample data with fuzzy label information, so that the generalization of the model is higher, specifically, a corresponding algorithm, for example, an algorithm corresponding to a neural network model, may be preset, and an initial target model of the target service may be constructed by the algorithm, and a corresponding convergence condition may be preset for the target model. The first sample data and the generated corresponding reduction label information can be used, the second sample data and the generated corresponding reduction label information can be used for carrying out model training on the target model, whether the target model is converged is judged according to the convergence condition, if the target model is not converged, the first sample data and the generated corresponding reduction label information and the second sample data and the generated corresponding reduction label information can be continuously obtained, the first sample data and the generated corresponding reduction label information and the second sample data and the generated corresponding reduction label information can be used for carrying out model training on the target model continuously until the target model is converged, and finally, the trained target model can be obtained.

It should be noted that the above processing is only one processing procedure, in practical applications, the tag reduction model and the target model may not achieve a better effect through the above one processing, and may also need to be executed one or more times in a circulating manner, so that the tag reduction model and the target model may achieve a better effect, that is, the processing in steps S102 to S106 may be executed first, if the tag reduction model or the target model does not achieve a better effect, the processing in steps S102 to S106 may be executed again, and the above circulating execution is performed until the tag reduction model and the target model achieve a better effect, and at this time, the finally trained tag reduction model and target model may be obtained.

The embodiment of the present specification provides a method for processing a model, performing model training on a tag reduction model through tagging information corresponding to first sample data and first sample data to obtain a trained tag reduction model, generating corresponding reduction tag information for the first sample data and second sample data respectively based on the trained tag reduction model, where the second sample data is sample data that does not include corresponding tagging information or is abnormal in tagging information, and finally performing model training on a target model corresponding to a target service based on the first sample data, the second sample data, and the generated corresponding reduction tag information to obtain a trained target model, so that, by alternately optimizing the sample data (namely, the second sample data) of the fuzzy label and the labeled sample data (namely, the first sample data), a better label reduction model and a target model are learned, more fuzzy label sample data are used, the model is given more freedom by a network self-updating mode instead of a parameter label giving mode, the model is allowed to learn better data distribution step by the alternate optimization iteration mode, the corresponding label reduction model can be multiplexed into other similar data scenes to help to give the partial real label information of the fuzzy label sample data, so that the sample data with the fuzzy label information on a large scale is better utilized, and the generalization of the model is improved.

Example two

As shown in fig. 2, an execution subject of the method may be a terminal device or a server, where the terminal device may be a mobile terminal device such as a mobile phone and a tablet computer, a computer device such as a notebook computer or a desktop computer, or an IoT device (specifically, a smart watch and a vehicle-mounted device), where the server may be an independent server, a server cluster formed by multiple servers, and the like, and the server may be a backend server of a financial service or an online shopping service, or a backend server of an application program, and the like. The method may specifically comprise the steps of:

in step S202, model training is performed on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, where the label information is label information obtained by performing label processing on sample data in advance, and the label reduction model is used to generate corresponding label information for the sample data.

The label reduction model may be constructed based on a convolutional neural network model in which the number of multiple layers of perceptrons or network layers is less than a preset number threshold, and the preset number threshold may be set according to an actual situation, specifically 10 or 5. The last network layer of the tag restore model may be a Sigmoid layer.

In implementation, the same data sample set may be extracted from the entire data distribution set, where the same data sample set may include positive sample data, negative sample data, tagging label information corresponding to the positive sample data, tagging label information corresponding to the negative sample data, and the like, and may further include sample data that is not tagged or cannot be tagged (which may also be referred to as fuzzy sample data), where the positive sample data and the negative sample data may constitute first sample data, and the tagging label information corresponding to the positive sample data and the tagging label information corresponding to the negative sample data may constitute tagging label information corresponding to the first sample data. The label reduction model may be subjected to model training based on the first sample data and the label labeling information corresponding to the first sample data, so as to obtain a trained label reduction model, which may specifically refer to the foregoing related contents, and is not described herein again. In this embodiment, the label reduction model needs to be trained through the processing in step S202, so that the label reduction model has a preliminary label information reduction capability, and although the label information corresponding to the fuzzy sample data is unknown, the first sample data is provided, and a preliminary label reduction model can be trained first by optimizing the label information of the part.

It should be noted that the tag reduction model may be somewhat simplified, for example, the tag reduction model may be constructed based on a conventional multilayer perceptron or a shallow convolutional neural network model, and may be specifically set according to actual conditions. The last network layer of the tag reduction model may be a Sigmoid layer, i.e., a layer of tags

G(x；θ)＝(p ₀ ,p ₁ )

Wherein p is ₀ +p ₁ ＝1，p ₀ ＝P(y＝0|x)，p ₁ = P (y = 1|x), x represents sample data (e.g., first sample data), y is label information (e.g., label information), P represents a conditional probability, and θ represents a model parameter of the label reduction model.

In step S204, corresponding restore tag information is generated for the first sample data and the second sample data respectively based on the trained tag restore model, and the second sample data is sample data that does not include corresponding label tagging information or has abnormal label tagging information.

In step S206, a loss function of the target model is constructed according to a preset gradient descent algorithm.

In practical application, the target model for reasoning in learning data distribution can be as complex as possible according to a data modality, for example, the target model can be constructed through a multilayer and more complex convolutional neural network, and the like, and can be specifically set according to actual conditions. The target model may set a corresponding function according to the target service, for example, if the target service is a payment service, the target model may be used as a model for risk prevention and control, for example, the target model may be used to detect whether a risk such as fraud and illegal financial activity exists in a process of executing the payment service, and may be specifically set according to an actual situation, which is not limited in the embodiment of the present specification. The gradient descent algorithm may be a first-order optimization algorithm, which is also commonly referred to as a steepest descent method, and to find a local minimum value of a function using the gradient descent algorithm, an iterative search must be performed to a distance point of a specified step size corresponding to a reverse direction of a gradient (or an approximate gradient) on the function at a current point, and if the iterative search is performed to a positive direction of the gradient, a local maximum value point of the function is approached, and the process is referred to as a gradient ascent method, and is referred to as a gradient descent method in reverse. In practical application, the gradient descent algorithm in step S206 may specifically be a random gradient descent algorithm, and the random gradient descent algorithm may be to randomly extract only one sample data for gradient calculation, and since only the gradient of one sample data is calculated in each gradient descent iteration, the calculation time is much shorter than that of a gradient descent manner corresponding to a small batch of sample data, although the loss function obtained in each iteration is not in the global optimal direction, the overall direction is in the global optimal solution, and the final result is often near the global optimal solution. The loss function may include a plurality of functions, such as a square loss function, an exponential loss function, and the like, and may be specifically set according to actual situations, which is not limited in this specification. In practical applications, the loss function may be as a cross-entropy loss function, i.e. a cross-entropy loss function

L＝-1/N(m ₀ log(p ₀ )+m ₁ log(p ₁ ))

Wherein L represents a cross entropy loss function, N represents the number of sample data, m ₀ And m ₁ Indicating the restore tag information.

In step S208, model training is performed on the target model applied to the target service based on the first sample data and the second sample data, the generated corresponding reduction label information, and the constructed loss function, so as to obtain a trained target model.

In implementation, the first sample data and the second sample data may be input into the target model, and combined into corresponding reduction label information generated by the first sample data and the second sample data, and the constructed loss function, and the loss function is minimized or maximized, so that model training and parameter optimization are performed on the target model, and finally, the trained target model may be obtained.

The processing of the above steps S202 to S208 is only a single processing procedure, and in practical applications, the processing of the above steps S202 to S208 is only executed once, and it may not be possible to achieve a good effect on both the label reduction model and the target model at all, so the label reduction model and the target model may be optimized in a circularly alternating execution manner, which may be specifically referred to the processing of the following steps S210 to S214.

In step S210, model training is performed on the label reduction model based on the first sample data and the labeled label information corresponding to the first sample data, and the second sample data and the generated corresponding reduction label information, so as to obtain a trained label reduction model, where the reduction label information is label information generated based on the label reduction model.

In implementation, in order to make the effect of the label reduction model better, because the label reduction model is preliminarily trained in the manner described above, so that the label reduction model has a certain reduction capability of the label information, part or all of the second sample data may be selected from the second sample data in which the reduced label information has been generated, then, the label information corresponding to the first sample data and the first sample data may be used, and the second sample data and the generated corresponding reduced label information may be used to further train the trained label reduction model, specifically, the trained label reduction model may be obtained, each parameter in the label reduction model may be kept unchanged, and the current label reduction model may be used as an initial label reduction model, and the label information corresponding to the first sample data and the second sample data and the generated corresponding reduced label information may be used to perform model training on the current label reduction model, so as to obtain the trained label reduction model.

The label reduction model is further trained in the above manner, at this time, the effect of the trained label reduction model may be better than that of the label reduction model trained in step S202, but whether the conditions for stopping training and iteration can be met needs to be determined according to specific situations.

In step S212, corresponding restore tag information is generated for the first sample data and the second sample data respectively based on the trained tag restore model, where the second sample data is sample data that does not include corresponding tag information or has tag information abnormality.

The specific processing of step S212 can refer to the related contents, and is not described herein again.

It should be noted that, since the effect of the currently trained label reduction model may be better than that of the label reduction model trained in step S202, the accuracy of the label information generated by the label reduction model may also be improved to some extent. In addition, the second sample data at this time may be completely the same as the second sample data described above, and in practical applications, a certain amount of new sample data may be collected from a specified database as the second sample data, or the new sample data and the previous second sample data may be combined to form new second sample data, which may be specifically set according to practical situations, and this is not limited in the embodiments of this specification.

In step S214, model training is performed on the target model applied to the target service based on the first sample data and the second sample data, the generated corresponding reduction label information, and the constructed loss function, so as to obtain a trained target model.

For a specific processing procedure of the step S214, reference may be made to the related contents, which are not described herein again.

It should be noted that, the accuracy of the label information generated by the label reduction model is improved to a certain extent, so that the model parameters in the trained target model can be kept unchanged, the current target model is used as an initial target model, then, the first sample data and the second sample data with improved accuracy of the reduction label information are used for continuing model training on the current target model, the corresponding reduction label information generated by combining the first sample data and the second sample data and the constructed loss function (specifically, the cross entropy loss function) are combined, and the loss function is minimized or maximized, so that model training and parameter optimization are performed on the target model, and finally, the trained target model can be obtained.

The processing in steps S202 to S214 may still not enable the label reduction model and the target model to achieve a better effect, and therefore, the label reduction model and the target model may be optimized by continuing to use a training alternative execution mode, that is, the processing in steps S210 to S214 may be executed again until the label reduction model and the target model achieve a better effect, so that a final label reduction model and a final target model may be obtained, and the obtained label reduction model and the target model have a better generalization. At this time, the final label reduction model and the target model may be deployed and applied, for example, the final label reduction model may be deployed in some scenarios or services trained by the model, and the final target model may be deployed in the target service. Therefore, by alternately optimizing the fuzzy label sample data and the labeled sample data, a better label reduction model and a target model are learned, more fuzzy label sample data are used, the model is given more freedom degree by a network self-updating mode instead of a parameter label giving mode, the model is allowed to learn better data distribution step by an alternate optimization iteration mode, and the corresponding label reduction model can be more reused in other similar data scenes to help give partial real label information of the fuzzy label sample data.

EXAMPLE III

As shown in fig. 3, an execution main body of the method may be a terminal device or a server, where the terminal device may be a mobile terminal device such as a mobile phone and a tablet computer, or may also be a computer device such as a notebook computer or a desktop computer, or may also be an IoT device (specifically, a smart watch, a vehicle-mounted device, and the like), where the server may be an independent server, or a server cluster formed by multiple servers, and the server may be a background server of a financial service. The first sample data may be a first facial image sample, the second sample data may be a second facial image sample, and the target model may be a facial recognition model. The method may specifically comprise the steps of:

in step S302, model training is performed on the label reduction model based on the first facial image sample and the label labeling information corresponding to the first facial image sample, so as to obtain a trained label reduction model.

The label reduction model can be constructed based on a convolutional neural network model with the number of multilayer perceptrons or network layers smaller than a preset number threshold. The last network layer of the tag restore model may be a Sigmoid layer.

In step S304, corresponding restored label information is generated for the first face image sample and the second face image sample respectively based on the trained label restoration model, and the second face image sample is a face image sample that does not include corresponding annotation label information or has abnormal annotation label information.

In step S306, a loss function of the face recognition model is constructed according to a preset gradient descent algorithm.

In step S308, model training is performed on the face recognition model based on the first face image sample and the second face image sample, the generated corresponding reduction label information, and the constructed loss function, so as to obtain a trained face recognition model.

The processing of steps S302 to S308 is only a single processing procedure, and in practical applications, the processing of steps S302 to S308 is executed only once, and it may not be possible to achieve a good effect on both the label reduction model and the face recognition model, so the label reduction model and the face recognition model may be optimized in a manner of circularly and alternately executing, and specifically, the processing of steps S310 to S314 may be referred to below.

In step S310, model training is performed on the label reduction model based on the first face image sample, the label labeling information corresponding to the first face image sample, the second face image sample, and the generated corresponding reduction label information, so as to obtain a trained label reduction model.

In step S312, corresponding restoration label information is generated for the first face image sample and the second face image sample respectively based on the trained label restoration model, and the second face image sample is a face image sample that does not include corresponding annotation label information or has abnormal annotation label information.

In step S314, model training is performed on the face recognition model based on the first face image sample and the second face image sample, the generated corresponding restoration label information, and the constructed loss function, so as to obtain a trained face recognition model.

The processing in steps S302 to S314 may still not enable the label reduction model and the face recognition model to achieve a good effect, and therefore, the label reduction model and the face recognition model may be optimized by continuing to alternately perform training, that is, the processing in steps S310 to S314 may be performed again until the label reduction model and the face recognition model achieve a good effect, so that a final label reduction model and a final face recognition model may be obtained, and the obtained label reduction model and the obtained face recognition model have a good generalization. At this time, the final label reduction model and the face recognition model can be deployed and applied. Therefore, by alternately optimizing the fuzzy label sample data and the labeled sample data, a better label reduction model and a face recognition model are learned, more fuzzy label sample data are used, the model is given more freedom degree by a network self-updating mode instead of a parameter label giving mode, the model is allowed to learn better data distribution step by an alternate optimization iteration mode, and the corresponding label reduction model can be more reused in other similar data scenes to help give partial real label information of the fuzzy label sample data.

The embodiment of the specification provides a model processing method, which includes performing model training on a label reduction model through label labeling information corresponding to first sample data and first sample data to obtain a trained label reduction model, respectively generating corresponding reduction label information for the first sample data and second sample data based on the trained label reduction model, wherein the second sample data is sample data which does not contain corresponding label labeling information or is abnormal in label labeling information, and finally performing model training on a target model for a target service based on the first sample data, the second sample data and the generated corresponding reduction label information to obtain a trained target model, by alternately optimizing the sample data of the fuzzy label (namely, the second sample data) and the labeled sample data (namely, the first sample data), a better label reduction model and a target model are learned, more fuzzy label sample data are used, the model is given more freedom by a network self-updating mode instead of a parameter label giving mode, the model is allowed to learn better data distribution step by an alternate optimization iteration mode, and the corresponding label reduction model can be further multiplexed into other similar data scenes to help give partial real label information of the fuzzy label sample data, so that the sample data with the fuzzy label information on a large scale is better utilized, and the generalization of the model is improved.

Example four

Based on the same idea, the method for processing the model provided in the embodiment of the present specification further provides a device for processing the model, as shown in fig. 4.

The model processing device comprises: a first training module 401, a label generation module 402, and a second training module 403, wherein:

the first training module 401 performs model training on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, where the label information is label information obtained by performing label processing on sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data;

a tag generation module 402, configured to generate corresponding restored tag information for the first sample data and second sample data respectively based on the trained tag restoration model, where the second sample data is sample data that does not include corresponding labeled tag information or is labeled tag information abnormal;

the second training module 403 performs model training on the target model applied to the target service based on the first sample data, the second sample data, and the generated corresponding reduction label information, to obtain a trained target model.

In this embodiment of the present specification, the first training module 401 performs model training on a label reduction model based on first sample data and label labeling information corresponding to the first sample data, and second sample data and corresponding generated reduction label information to obtain a trained label reduction model, where the reduction label information is label information generated based on the label reduction model.

In an embodiment of the present specification, the tag reduction model is constructed based on a convolutional neural network model in which the number of multiple layers of perceptrons or network layers is smaller than a preset number threshold, the target model is constructed based on the neural network model, and the last network layer of the tag reduction model and the target model is a Sigmoid layer.

In this embodiment, the second training module 403 includes:

the loss function construction unit is used for constructing a loss function of the target model according to a preset random gradient descent algorithm;

and the second training unit is used for performing model training on the target model which is applied to the target service and corresponds to the generated corresponding reduction label information and the constructed loss function based on the first sample data and the second sample data to obtain the trained target model.

In the embodiment of the present specification, the loss function of the target model is a cross entropy loss function.

In the embodiments of the present specification, the gradient descent algorithm includes a random gradient descent algorithm.

In an embodiment of the present specification, the target model is a model for risk prevention and control of a preset risk, where the preset risk includes one or more of a fraud risk and an illegal financial activity.

The embodiment of the specification provides a processing device of a model, model training is performed on a label reduction model through label information corresponding to first sample data and first sample data to obtain a trained label reduction model, corresponding reduction label information is generated for the first sample data and second sample data respectively based on the trained label reduction model, the second sample data is sample data which does not contain corresponding label information or is abnormal in label information, finally, model training is performed on a target model used for a target service based on the first sample data and the second sample data and the generated corresponding reduction label information to obtain the trained target model, therefore, the model can be given more freedom by alternately optimizing fuzzy label sample data (namely, the second sample data) and the labeled sample data (namely, the first sample data), not only better label reduction model and the target model are learned, but also more fuzzy label sample data are used, and fuzzy label information is given through a network self-updating mode instead of giving parameter labels, the model gradually better data distribution by alternately optimizing iteration mode, the corresponding reduction model can be multiplexed into other fuzzy label information of similar data, and the fuzzy label information of the real sample data is provided, so as to help the large-scale fuzzy label information of the sample data, and the sample data.

EXAMPLE five

Based on the same idea, the apparatus for processing a model provided in the embodiments of the present specification further provides a device for processing a model, as shown in fig. 5.

The processing device of the model may be disposed in the terminal device or the server provided in the above embodiments.

The processing devices of the model may vary significantly depending on configuration or performance, and may include one or more processors 501 and memory 502, where the memory 502 may have one or more stored applications or data stored therein. Memory 502 may be, among other things, transient or persistent storage. The application program stored in memory 502 may include one or more modules (not shown), each of which may include a series of computer-executable instructions in a processing device for the model. Still further, the processor 501 may be arranged in communication with the memory 502, executing a series of computer executable instructions in the memory 502 on the processing device of the model. The processing apparatus of the model may also include one or more power supplies 503, one or more wired or wireless network interfaces 504, one or more input-output interfaces 505, one or more keyboards 506.

In particular, in this embodiment, the processing device of the model includes a memory, and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for the processing device of the model, and the one or more programs configured to be executed by the one or more processors include computer-executable instructions for:

performing model training on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, wherein the label information is obtained by labeling sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data;

generating corresponding restored label information for the first sample data and second sample data respectively based on the trained label restoration model, wherein the second sample data is sample data which does not contain corresponding labeled label information or is abnormal in labeled label information;

and performing model training on a target model applied to the target service based on the first sample data, the second sample data and the generated corresponding reduction label information to obtain a trained target model.

In an embodiment of this specification, the performing model training on the label reduction model based on the first sample data and the labeled label information corresponding to the first sample data to obtain a trained label reduction model includes:

and performing model training on the label reduction model based on the first sample data and the label information corresponding to the first sample data, the second sample data and the generated corresponding reduction label information to obtain a trained label reduction model, wherein the reduction label information is label information generated based on the label reduction model.

In an embodiment of this specification, the performing, based on the first sample data, the second sample data, and the generated corresponding reduction label information, model training on the target model applied to the target service to obtain a trained target model includes:

constructing a loss function of the target model according to a preset random gradient descent algorithm;

and performing model training on the target model applied to the target service on the basis of the first sample data, the second sample data, the generated corresponding reduction label information and the constructed loss function to obtain the trained target model.

In an embodiment of the present specification, the target model is a model for risk prevention and control on a preset risk, where the preset risk includes one or more of a fraud risk and an illegal financial activity.

The embodiment of the present specification provides a processing device of a model, performing model training on a tag reduction model through tagging information corresponding to first sample data and first sample data to obtain a trained tag reduction model, generating corresponding reduction tag information for the first sample data and second sample data respectively based on the trained tag reduction model, where the second sample data is sample data not containing corresponding tagging information or having abnormal tagging information, and finally performing model training on a target model used for a target service based on the first sample data and the second sample data and the generated corresponding reduction tag information to obtain a trained target model, so that a greater degree of freedom is given to the model by alternately optimizing fuzzy tag sample data (i.e., the second sample data) and the tagged sample data (i.e., the first sample data), not only learning a better tag reduction model and the target model, but also using more fuzzy tag sample data, and using fuzzy tag information in a manner of network self-updating instead of giving parameter tagging, allowing the model to learn better data distribution step by step, the corresponding reduction model can be reused in fuzzy tag information of other similar data, and the fuzzy tag information of the model can be provided by using more generalized information, and the fuzzy tag information of the real sample data can be provided by the generalized model.

Example six

Further, based on the methods shown in fig. 1 to fig. 3, one or more embodiments of the present specification further provide a storage medium for storing computer-executable instruction information, in a specific embodiment, the storage medium may be a usb disk, an optical disk, a hard disk, or the like, and when the storage medium stores the computer-executable instruction information, the storage medium implements the following processes:

generating corresponding reduced label information for the first sample data and second sample data respectively based on the trained label reduction model, wherein the second sample data is sample data which does not contain corresponding labeled label information or is abnormal in labeled label information;

and performing model training on a target model applied to the target service on the basis of the first sample data, the second sample data, the generated corresponding reduction label information and the constructed loss function to obtain a trained target model.

The embodiment of the present specification provides a storage medium, a label reduction model is subjected to model training through label labeling information corresponding to first sample data and first sample data to obtain a trained label reduction model, corresponding reduction label information is generated for the first sample data and second sample data respectively based on the trained label reduction model, the second sample data is sample data not containing corresponding label labeling information or having abnormal label labeling information, finally, model training is performed on a target model used for a target service based on the first sample data and the second sample data and the generated corresponding reduction label information to obtain a trained target model, so that the better label reduction model and the target model can be learned through alternately optimizing fuzzy label sample data (i.e., the second sample data) and labeled target model (i.e., the first sample data), more fuzzy label samples are used, the model is given more freedom degrees through a network self-updating mode instead of giving parameter labels, the alternately optimizing iterative mode also allows the model to learn data distribution better and the corresponding label reduction model can be reused in fuzzy sample data of other similar label reduction models to provide better fuzzy label information, and the sample data of the fuzzy label reduction model can be provided with better generalization information.

The foregoing description has been directed to specific embodiments of this disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.

In the 90's of the 20 th century, improvements to a technology could clearly distinguish between improvements in hardware (e.g., improvements to circuit structures such as diodes, transistors, switches, etc.) and improvements in software (improvements to process flow). However, as technology advances, many of today's process flow improvements have been seen as direct improvements in hardware circuit architecture. Designers almost always obtain the corresponding hardware circuit structure by programming an improved method flow into the hardware circuit. Thus, it cannot be said that an improvement in the process flow cannot be realized by hardware physical modules. For example, a Programmable Logic Device (PLD) (e.g., a Field Programmable Gate Array (FPGA)) is an integrated circuit whose Logic functions are determined by a user programming the Device. A digital system is "integrated" on a PLD by the designer's own programming without requiring the chip manufacturer to design and fabricate application-specific integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing an Integrated Circuit chip, such Programming is often implemented by "logic compiler" software, which is similar to a software compiler used in program development and writing, but the original code before compiling is also written by a specific Programming Language, which is called Hardware Description Language (HDL), and HDL is not only one but many, such as ABEL (Advanced Boolean Expression Language), AHDL (alternate Hardware Description Language), traffic, CUPL (core universal Programming Language), HDCal, jhddl (Java Hardware Description Language), lava, lola, HDL, PALASM, rhyd (Hardware Description Language), and vhigh-Language (Hardware Description Language), which is currently used in most popular applications. It will also be apparent to those skilled in the art that hardware circuitry that implements the logical method flows can be readily obtained by merely slightly programming the method flows into an integrated circuit using the hardware description languages described above.

The controller may be implemented in any suitable manner, for example, the controller may take the form of, for example, a microprocessor or processor and a computer readable medium that stores computer readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, an Application Specific Integrated Circuit (ASIC), a programmable logic controller, and embedded microcontrollers, examples of which include, but are not limited to, the following microcontrollers: ARC625D, atmel AT91SAM, microchip PIC18F26K20, and Silicone Labs C8051F320, the memory controller may also be implemented as part of the control logic for the memory. Those skilled in the art will also appreciate that, in addition to implementing the controller as pure computer readable program code, the same functionality can be implemented by logically programming method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Such a controller may thus be regarded as a hardware component and the means for performing the various functions included therein may also be regarded as structures within the hardware component. Or even means for performing the functions may be regarded as being both a software module for performing the method and a structure within a hardware component.

The systems, apparatuses, modules or units described in the above embodiments may be specifically implemented by a computer chip or an entity, or implemented by a product with certain functions. One typical implementation device is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

For convenience of description, the above devices are described as being divided into various units by function, respectively. Of course, the functionality of the various elements may be implemented in the same one or more pieces of software and/or hardware in implementing one or more embodiments of the present description.

As will be appreciated by one skilled in the art, embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, one or more embodiments of the present description may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of the present description may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and so forth) having computer-usable program code embodied therein.

Embodiments of the present description are described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the description. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable fraud case serial-parallel apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable fraud case serial-parallel apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable fraud case to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable fraud case serial-parallel apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include forms of volatile memory in a computer readable medium, random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.

Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Disks (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium, which can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising a … …" does not exclude the presence of another identical element in a process, method, article, or apparatus that comprises the element.

As will be appreciated by one skilled in the art, embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, one or more embodiments of the present description may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of the present description may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

One or more embodiments of the specification may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. One or more embodiments of the specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the system embodiment, since it is substantially similar to the method embodiment, the description is simple, and for the relevant points, reference may be made to the partial description of the method embodiment.

The above description is only an example of the present disclosure, and is not intended to limit the present disclosure. Various modifications and alterations to this description will become apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification should be included in the scope of the claims of the present specification.

Claims

1. A method of processing a model, the method comprising:

performing model training on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, wherein the label information is label information obtained by labeling the sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data;

2. The method of claim 1, wherein performing model training on the label reduction model based on the first sample data and the labeled label information corresponding to the first sample data to obtain a trained label reduction model comprises:

model training is carried out on the label reduction model based on the first sample data and the labeling label information corresponding to the first sample data, the second sample data and the generated corresponding reduction label information to obtain a trained label reduction model, and the reduction label information is label information generated based on the label reduction model.

3. The method of claim 2, wherein the tag reduction model is constructed based on a convolutional neural network model in which the number of multilayer perceptrons or network layers is less than a preset number threshold, the target model is constructed based on a neural network model, and the last network layer of the tag reduction model and the target model is a Sigmoid layer.

4. The method according to claim 2, wherein performing model training on a target model applied to a target service based on the first sample data and the second sample data and the generated corresponding reduction label information to obtain a trained target model, includes:

5. The method of claim 4, the loss function of the target model being a cross-entropy loss function.

6. The method of claim 4, the gradient descent algorithm comprising a random gradient descent algorithm.

7. The method of any one of claims 1-6, wherein the target model is a model for risk prevention and control of a predetermined risk, the predetermined risk comprising one or more of a risk of fraud, an illegal financial activity.

8. An apparatus for processing a model, the apparatus comprising:

the first training module is used for carrying out model training on the label reduction model based on the first sample data and label information corresponding to the first sample data to obtain a trained label reduction model, wherein the label information is label information obtained by carrying out label processing on sample data in advance, and the label reduction model is used for generating corresponding label information for the sample data;

the label generation module is used for respectively generating corresponding reduced label information for the first sample data and second sample data based on the trained label reduction model, wherein the second sample data is sample data which does not contain corresponding labeled label information or is abnormal in labeled label information;

and the second training module performs model training on the target model applied to the target service based on the first sample data, the second sample data and the generated corresponding reduction label information to obtain the trained target model.

9. A model processing apparatus, the model processing apparatus comprising:

a processor; and

a memory arranged to store computer executable instructions that, when executed, cause the processor to:

10. A storage medium for storing computer-executable instructions, which when executed by a processor implement the following: