CN115130536A - Training method of feature extraction model, data processing method, device and equipment - Google Patents

Abstract

The embodiment of the application provides a training method, a data processing method, a device and equipment of a feature extraction model, and relates to the technical fields of artificial intelligence, payment safety, finance and cloud, wherein the training method comprises the following steps: acquiring a training set, and constructing a plurality of positive sample pairs and a plurality of negative sample pairs based on the training set, wherein the positive sample pairs comprise two training samples in the same category, and the negative sample pairs comprise two training samples in different categories; repeatedly executing training operation on the neural network model based on the training set until a preset condition is met, and obtaining a trained feature extraction model; and for each training operation, if the preset condition is not met, determining a plurality of new sample pairs based on the similarity between the feature vectors of the training samples output by the model, and taking the new sample pairs as the sample pairs based on the subsequent training operation. Based on the training method provided by the application, the performance of the feature extraction model can be effectively improved.

Description

Training method, data processing method, device and equipment for feature extraction model

technical field

The present application relates to the fields of artificial intelligence, payment security, finance and cloud technologies. Specifically, the embodiments of the present application relate to a training method, data processing method, apparatus and equipment for a feature extraction model.

Background technique

With the rapid development of artificial intelligence technology, more and more artificial intelligence technologies are applied in various fields. Data processing methods based on artificial intelligence technology have been widely used in more and more scenarios. data processing belongs to one of the very important branches.

For the data processing method based on the neural network model, the feature extraction of the input data of the model is usually essential, and the expression ability of the features extracted by the model is a very key factor that affects the quality of the model processing results. Therefore, by Training methods to obtain high-performance feature extraction models is an important research topic in artificial intelligence technology. In the prior art, in order to improve the performance of the model, a variety of model training methods have been proposed. Although the models trained by some existing training methods can meet the application requirements to a certain extent, the performance of the model still needs to be improved, especially when the model is trained. When applied to specific tasks (such as classification tasks in specific scenarios), the expressiveness of the features extracted by the model still needs to be improved.

SUMMARY OF THE INVENTION

The purpose of the embodiments of the present application is to provide a training method that can effectively improve the performance of a feature extraction model, and a data processing method, apparatus, electronic device, and computer-readable storage medium based on the training method. In order to achieve this purpose, the technical solutions provided in the embodiments of the present application are as follows:

On the one hand, an embodiment of the present application provides a method for training a feature extraction model, the method comprising:

obtaining a training set, the training set includes training samples of multiple categories;

Construct a plurality of sample pairs based on the training set, the plurality of sample pairs include a plurality of positive sample pairs and a plurality of negative sample pairs, wherein the positive sample pairs include two training samples belonging to the same category, the negative sample pairs The sample pair includes two training samples of the different classes;

The training operation is repeatedly performed on the neural network model based on the training set until the preset conditions are met, and the neural network model satisfying the preset conditions is used as the trained feature extraction model; wherein, the preset conditions include the neural network model The corresponding training total loss converges or the number of training times reaches the set number of times, and the training operations include:

Input each training sample in the multiple sample pairs into the neural network model respectively, and obtain the feature vector of each training sample;

Determine the total training loss based on the first similarity between the feature vectors of the training samples in each of the sample pairs;

If the total training loss does not converge and the number of training times does not reach the set number of times, the model parameters of the neural network model are adjusted, and based on the second similarity between the feature vectors of the respective training samples, a plurality of new sample pairs, and use the new multiple sample pairs as multiple sample pairs on which subsequent training operations are based.

In an optional embodiment of the present application, the above-mentioned total training loss represents the degree of difference between pairs of positive samples and the degree of similarity between pairs of negative samples in multiple sample pairs input to the model.

In an optional embodiment of the present application, the second similarity corresponding to the new positive sample pair corresponding to each training sample is smaller than the second similarity corresponding to the pre-updated positive sample pair corresponding to the training sample, and each The second similarity corresponding to the new negative sample pair corresponding to the training sample is greater than the second similarity corresponding to the pre-updated negative sample pair corresponding to the training sample.

Optionally, the above-mentioned feature extraction model is a feature extraction model in a classification model, and the classification model is used to extract the feature vector of the first data to be processed through the feature extraction model, and identify the classification corresponding to the first data to be processed based on the extracted feature vector. Result; wherein, the classification result is one of multiple specified categories, the multiple categories include the multiple specified categories, and a training sample of each specified category is the second data to be processed corresponding to the specified category.

Optionally, the above classification model may specifically identify whether the first data to be processed is data corresponding to the target category based on the extracted feature vector, that is, the above classification result indicates that the first data to be processed is data corresponding to the target category or is a non-target category. Corresponding data, correspondingly, the above-mentioned multiple designated categories include a target category and at least one non-target category.

Optionally, the above-mentioned specified category is a specified object category, the above-mentioned second data to be processed is business data corresponding to a sample object of the specified object category, the above-mentioned first to-be-processed data is business data corresponding to the target object, and the above classification result represents the target. Whether the object is an object of the target class.

On the other hand, an embodiment of the present application provides a training device for a feature extraction model, and the device includes:

a training data acquisition module for acquiring a training set, the training set including training samples of multiple categories;

A training data processing module, configured to construct a plurality of sample pairs based on the training set, the plurality of sample pairs include a plurality of positive sample pairs and a plurality of negative sample pairs, wherein the positive sample pairs include two samples of the same category training samples, the negative sample pair includes two training samples of the different categories;

A model training module, configured to repeatedly perform training operations on the neural network model based on the training set until a preset condition is met, and use the neural network model that satisfies the preset condition as a trained feature extraction model; wherein the above preset The conditions include that the total training loss corresponding to the neural network model converges or the number of training times reaches a set number of times, and the training operations include:

Input each training sample in the multiple sample pairs into the neural network model respectively, and obtain the feature vector of each training sample;

Determine the total training loss based on the first similarity between the feature vectors of the training samples in each of the sample pairs;

If the total training loss does not converge and the number of training times does not reach the set number of times, the model parameters of the neural network model are adjusted, and based on the second similarity between the feature vectors of the respective training samples, a plurality of new sample pairs, and use the new multiple sample pairs as multiple sample pairs on which subsequent training operations are based.

Optionally, the model training module can be used to: for each of the training samples, respectively determine the second similarity between the feature vector of the training sample and the feature vector of each first sample, and assign the corresponding second similarity. The first sample with the lowest degree and the training sample are taken as a new positive sample pair, wherein the first sample is a training sample belonging to the same category as the training sample among the training samples; for each of the training samples training samples, respectively determine the second similarity between the feature vector of the training sample and the feature vector of each second sample, and use the corresponding second sample with the highest second similarity and the training sample as a new negative sample pair , wherein the second sample is a training sample belonging to a different category from the training sample among the training samples.

Optionally, the training data processing module can be used to: take each training sample in the training set as an anchor point, and construct a sample group corresponding to each anchor point, and the sample group corresponding to each anchor point includes the corresponding sample group of the anchor point. A positive sample pair and a negative sample pair, wherein a positive sample pair corresponding to an anchor point includes the anchor point and the positive sample of the anchor point, and a negative sample pair corresponding to an anchor point includes the anchor point and the negative sample of the anchor point. sample;

Correspondingly, when determining the total training loss, the model training module can be used to: for each sample group, according to the first similarity between the feature vectors of the two samples in the positive sample pair of the sample group, and the The first similarity between the feature vectors of the two samples in the negative sample pair determines the training loss corresponding to the sample group; the total training loss is determined according to the training loss corresponding to each of the sample groups;

When determining a plurality of new sample pairs, the model training module can be used to: determine a new sample group corresponding to each anchor point based on the second similarity between the feature vectors of each training sample, The sample pairs in the new sample group corresponding to the anchor point are used as multiple sample pairs in subsequent training operations.

Optionally, the model training module may be used when determining a new sample group corresponding to each of the anchor points: for each of the anchor points, respectively determine the feature vector of the anchor point and the feature vector of each first sample. The second similarity between the two is determined, and the first sample with the lowest second similarity is determined as a new positive sample corresponding to the anchor point, wherein the first sample is the same as the The anchor points belong to the training samples of the same category; for each anchor point, the second similarity between the feature vector of the anchor point and the feature vector of each second sample is determined, and the corresponding second similarity is the highest. The second sample is determined as a new negative sample corresponding to the anchor point, wherein the second sample is a training sample belonging to a different category from the anchor point among the training samples.

Optionally, the training data processing module may be configured to: construct at least one batch data set according to the training set when constructing the sample groups corresponding to the anchor points, wherein each of the batch data sets includes p categories. , and the number of training samples for each category is k, p≥2, k≥3; for each batch of data sets, each training sample in the batch of data sets is used as an anchor point, based on For each training sample in the batch of datasets, construct a sample group corresponding to each anchor point in the batch of datasets;

Correspondingly, the model training module can be used for: repeating training operations on the neural network model based on each of the batch data sets, wherein each training operation is performed based on a sample group corresponding to each anchor point in one of the batch data sets ;

When determining a new sample group corresponding to each of the anchor points, the model training module can be used to: for each anchor point in the batch data set corresponding to the current training operation, divide the anchor point according to the anchor point and the batch data set. The second similarity between the training samples other than the anchor point determines the new sample group corresponding to the anchor point.

Optionally, for each sample group, when determining the training loss corresponding to the sample group, the model training module can be used to:

Determine the first distance between the feature vectors of the two samples in the positive sample pair of the sample group and the second distance between the feature vectors of the two samples in the negative sample pair of the sample group, where the first distance represents the first similarity corresponding to the positive sample pair of the sample group, and the second distance represents the first similarity corresponding to the negative sample pair of the sample group;

determining the difference between the first distance and the second distance; determining the training loss corresponding to the sample group according to the difference, wherein the training loss corresponding to the sample group is positively correlated with the difference.

Optionally, the training loss corresponding to each of the sample groups is determined based on the following expression:

s( ^x )=ln(1+ex )

x=d(a,p)-d(a,n)+β

Among them, s(x) represents the training loss corresponding to the sample group, a, p and n represent the anchor points, positive samples and negative samples in the sample group, respectively, d(a, p) represents the first distance, d(a, n ) represents the second distance, and β represents the preset adjustment threshold.

On the other hand, an embodiment of the present application also provides a data processing method based on a neural network model, the method comprising:

Obtaining data to be processed; inputting the data to be processed into a first feature extraction model, and extracting a first feature vector corresponding to the data to be processed through the first feature extraction model, wherein the first feature extraction model is Obtained by training using the training method provided in any optional embodiment of the present application;

Based on the first feature vector, a classification result corresponding to the data to be processed is determined.

On the other hand, an embodiment of the present application also provides a data processing device based on a neural network model, the device comprising:

The data acquisition module is used to acquire the data to be processed;

a data processing module, configured to input the data to be processed into a first feature extraction model, extract a first feature vector corresponding to the data to be processed through the first feature extraction model, and determine based on the first feature vector The classification result corresponding to the data to be processed, wherein the first feature extraction model is obtained by training using the training method provided in any optional embodiment of the present application.

Optionally, the data processing module may be used to: extract the second feature vector of the data to be processed by using the second feature extraction model; fuse the first feature vector and the second feature vector; based on The fused features determine the classification result corresponding to the data to be processed.

Optionally, the data acquisition module can be used to: acquire the business data of the target business in multiple time periods corresponding to the target object, and the business data corresponding to each time period includes the attribute value of at least one business attribute of the target business; business data corresponding to multiple time periods, construct a business time sequence feature matrix corresponding to the target object, use the business time sequence feature matrix as the data to be processed, and the classification result represents the object type of the target object; wherein, The number of rows of the service time sequence feature matrix is the number of time periods of the plurality of time periods, the number of columns is the number of attributes of the at least one service attribute, and each element value in the service time sequence feature matrix represents one of the The attribute value of a business attribute corresponding to the time period.

Based on the methods provided by the embodiments of the present application, the present application further provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the present application The training method provided by any optional embodiment, or the data processing method provided by any optional embodiment of the present application is implemented.

Embodiments of the present application further provide a computer-readable storage medium, where a computer program is stored in the storage medium, and when the computer program is executed by a processor, the training method provided by any optional embodiment of the present application is implemented, or the computer program is implemented. Apply for the data processing method provided by any optional embodiment.

An embodiment of the present application further provides a computer program product, the computer product includes a computer program, and when the computer program is executed by a processor, implements the training method provided by any optional embodiment of the present application, or implements any optional embodiment of the present application. The data processing method provided by the selected embodiment.

The beneficial effects brought by the technical solutions provided in the embodiments of the present application are as follows:

In the training method provided by the embodiment of the present application, during the training process of the neural network model based on the positive sample pair and the negative sample pair, based on the similarity between the feature vectors of the training samples extracted by the model, the training of each sample pair is continuously performed. Update, that is, determine a new sample pair according to the similarity. In this way, in the process of model training, instead of simply selecting a positive sample pair and a negative sample pair at random, the sample pair is selected according to the similarity between the samples to realize the optimization of the sample pair. , to improve the training effect of the model. Based on the method of the present application, matching sample pairs can be selected according to application requirements in the training process to better meet application requirements. Optionally, the model can learn by selecting sample combinations that are difficult to learn. Effectively improve the model performance of the trained feature extraction model, improve the discrimination of the feature vector output by the model, and speed up the training efficiency of the model. Training the model in this way can better meet the needs of practical applications.

Description of drawings

In order to illustrate the technical solutions in the embodiments of the present application more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments of the present application.

FIG. 1 is a schematic diagram of the architecture of a data processing system to which an embodiment of the application is applicable;

2 is a schematic flowchart of a data processing method in an application scenario provided by an embodiment of the present application;

3 is a schematic flowchart of a training method for a feature extraction model provided by an embodiment of the present application;

4 is a schematic flowchart of a training method of a feature extraction model provided by an embodiment of the present application;

5 is a schematic diagram of a model training principle provided by an embodiment of the present application;

6 is a schematic flowchart of a data processing method provided by an embodiment of the present application;

7 is a schematic structural diagram of a training device provided by an embodiment of the present application;

FIG. 8 is a schematic structural diagram of a data processing apparatus according to an embodiment of the present application;

FIG. 9 is a schematic structural diagram of an electronic device to which an embodiment of the present application is applied.

Detailed ways

Embodiments of the present application are described below with reference to the accompanying drawings in the present application. It should be understood that the embodiments described below in conjunction with the accompanying drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application, and do not limit the technical solutions of the embodiments of the present application.

It will be understood by those skilled in the art that the singular forms "a", "an", "the" and "the" as used herein can include the plural forms as well, unless expressly stated otherwise. It should be further understood that the terms "comprising" and "comprising" used in the embodiments of the present application mean that corresponding features can be implemented as presented features, information, data, steps, operations, elements and/or components, but do not exclude Implementations support other features, information, data, steps, operations, elements, components, and/or combinations thereof, etc., as supported in the art. It will be understood that when we refer to an element as being "connected" or "coupled" to another element, the one element can be directly connected or coupled to the other element, or the one element and the other element may be intervening through intervening elements Establish a connection relationship. Furthermore, "connected" or "coupled" as used herein may include wirelessly connected or wirelessly coupled. The term "and/or" as used herein indicates at least one of the items defined by the term, eg "A and/or B" can be implemented as "A", or as "B", or as "A and B" ". When describing multiple (two or more) items, if the relationship between the multiple items is not clearly defined, the multiple items may refer to one, more or all of the multiple items, for example, The description of "parameter A includes A1, A2, and A3" can be implemented as parameter A includes A1, A2, or A3, or as parameter A includes at least two of the three parameters A1, A2, and A3.

In the current related art, for feature extraction of data, such as feature extraction of time series data, a feature construction scheme based on statistical domain, frequency domain or time domain is usually adopted, or a feature Embedding method based on deep learning is adopted. However, the former requires prior knowledge and expert experience, and has a certain degree of information loss. For the latter, the current feature embedding methods mostly use unsupervised embedding methods such as word2vec (word steering) and item2vec (content steering). When the features extracted by these methods are applied in a specific scene, the gain effect of the features cannot be guaranteed, that is, the extracted features are universal, but the feature discrimination under specific tasks is poor.

The training method provided by the embodiment of the present application is aimed at the problems existing in the prior art, and a training method and data processing method of a feature extraction model based on deep metric learning are proposed. Based on the method provided by the embodiment of the present application, it can be Effectively improve the performance of the feature extraction model. The feature vector extracted by this model has better feature discrimination, and can be applied to specific tasks to improve the data processing effect of the task.

The solutions provided in the embodiments of the present application are implemented based on artificial intelligence (Artificial Intelligence, AI) technology. For example, the feature extraction model and the classification model are both artificial intelligence-based neural network models. The feature extraction model may be a model based on any existing feature extraction model or a model after improving the existing feature extraction model, that is to say, the training method of the feature extraction model provided in this The training of any feature extraction model can improve the performance of the model obtained by training, wherein the training of the feature extraction model can be obtained by means of machine learning (ML) based on the training set (ie, a large number of training samples).

Among them, artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making. With the research and progress of artificial intelligence technology, artificial intelligence technology has been used in common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, autonomous driving, drones, robots, smart medical, intelligent It is believed that with the development of technology, this technology will be applied in more fields and play more and more important value.

Optionally, the data processing involved in the methods provided in the embodiments of the present application may be implemented based on cloud technology. For example, in the application process of the feature extraction model (such as classifying the data to be processed) and the data calculation involved in the training process (such as the similarity calculation between the feature vectors in the model training process, the calculation of training loss, model parameters adjustment, etc.) can be implemented using cloud computing technology. Among them, cloud computing is a computing mode that distributes computing tasks on a resource pool composed of a large number of computers, so that various application systems can obtain computing power, storage space and information services as needed. The network that provides the resources is called the "cloud". The resources in the "cloud" are infinitely expandable in the eyes of users, and can be obtained at any time, used on demand, expanded at any time, and paid for according to usage.

The training method or data processing method of the feature extraction model provided by the embodiments of the present application may be executed by any electronic device, for example, may be executed by a user terminal or a server. For example, the data to be processed may be data sent by the user to the server through the user terminal, the server may deploy a trained feature extraction model, and the server executes the data processing method provided by the embodiment of the present application and invokes the model to realize the processing The extraction of the special vector of the processing data, and the subsequent processing can be performed based on the extracted feature vector according to the application requirements. The subsequent processing can include but is not limited to classifying the data to be processed, and judging the similarity between the data to be processed and other data, etc. . For example, a user can send an image set containing a large number of images to the server through his user terminal, and the server can call the trained feature extraction model to extract the feature vector of each image in the image set separately, and can extract the feature vector of each image in the image set according to the similarity between the feature vectors of the images. The degree of grouping the image set according to the image category, and can provide the grouping results to the user.

The above server may be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services. The above-mentioned user terminal (also referred to as user equipment) may be a smart phone, a tablet computer, a notebook computer, a desktop computer, an intelligent voice interaction device (such as a smart speaker), a wearable electronic device (such as a smart watch), a vehicle terminal, and a smart home appliance. (such as smart TV), AR/VR equipment, etc., but not limited to this. The terminal and the server may be directly or indirectly connected through wired or wireless communication, which is not limited in this application.

Optionally, the method provided in this embodiment of the present application may be implemented as an independent application program or a functional module/plug-in of an application program, for example, the application program may be a special data classification software or other data classification function. Application, through which the classification of the data to be processed can be achieved.

The feature extraction model trained by the training method for the feature extraction model provided in the embodiment of the present application can be applied to any application scenario that requires the extraction of discriminative feature vectors (that is, feature representations) of the data to be processed, and may include but It is not limited to scenarios such as data classification and object classification. For example, the feature extraction model provided in this embodiment of the present application can be applied to risk control management. Based on the feature extraction model, the feature vector of the business data of the related business can be extracted, and risk assessment based on the feature vector can be performed. Identify, for example, the feature extraction model can be used as the backbone network of the classification model, after the feature vector of the business data is extracted through the backbone network, the risk level corresponding to the business data can be predicted based on the business vector and the classification model of the classification model. Different risk levels in the scene are of different categories. During model training, the training samples included in the training set may be multiple training samples corresponding to each risk level, and one training sample for each risk level may include a training sample corresponding to the risk level. sample business data (that is, business data with known true risk levels).

It should be noted that, in the optional embodiments of the present application, for the data related to the objects involved, when the embodiments of the present application are applied to specific products or technologies, the object's permission or consent needs to be obtained, and the relevant The collection, use and processing of data need to comply with relevant laws, regulations and standards of relevant countries and regions. That is to say, if the data related to the object is involved in the embodiments of the present application, the data needs to be obtained with the authorization and consent of the object and in compliance with the relevant laws, regulations and standards of the country and region.

In order to better describe and understand the method provided by the embodiments of the present application, an optional implementation manner of the method provided by the present application is first described below with reference to a specific application scenario. In this scenario embodiment, taking the data processing method provided by the embodiment of the present application applied to an application program with a mobile payment function as an example, the method provided by the embodiment of the present application can be used to identify the type of the object corresponding to the application (that is, category), for example, whether the object is a risk object can be determined based on business data related to the object, that is, whether it is an object of the target category, and the business data can be one or more specified types of business (that is, the target business can be a or multiple) of business data, wherein the method for classifying object types is not limited in this embodiment of the present application.

FIG. 1 shows a schematic structural diagram of a data processing system to which an embodiment of the present application is applied. As shown in FIG. 1 , the data processing system may include a terminal device 11 , a terminal device 21 , an application server 20 and a training server 30 , wherein , the application server 20 may be a server that provides the above-mentioned mobile payment function application program. The terminal device 11 may be an electronic device that uses the application program. The terminal device 11 may communicate with the application server 20 through the network, operate through the user interface of the application program, and use the services provided by the application program. The terminal device 21 may be an electronic device on the management side of the application, and the management client of the application may run on the terminal device 21. The terminal device 21 can also communicate with the application server 20 through the network, and an authorized administrator or related personnel can The management of the application is performed through the management client.

The training server 30 may perform the training operation of the neural network model by executing the training method provided in the embodiment of the present application, so as to obtain a trained feature extraction model. Optionally, after obtaining the trained feature extraction model, the training server 30 may further train the classification model including the feature extraction model to obtain the trained classification model. The application server 20 and the training server 30 may communicate through the network, and after obtaining the trained classification model through the training server 30, the trained classification model may be deployed in the application server 20, and the application server 20 executes the embodiment of the present application by executing the The provided data processing method is to process the data to be processed by invoking the trained classification model. The processing may include, but is limited to, the identification of the category of the object corresponding to the data to be processed. Optionally, the application server 20 may also use the identification result. It is sent to the terminal device 21 to present the identification result to the manager of the application.

It should be noted that, in the above application scenario, the data processing method provided in the embodiment of the present application may be executed by the application server 20, or may be executed by other electronic devices independent of the application server, such as a cloud server. The following description takes the application server as the execution subject as an example.

FIG. 2 is a schematic flowchart of data processing based on the data processing system shown in FIG. 1 . The solution provided by the present application will be described below with reference to FIG. 1 and FIG. 2. This scenario embodiment may include steps S11 to S13, and steps S21 and S23, as follows:

Step S11: Acquire a training data set.

Step S12: The training server 20 trains the feature extraction model and the classification model.

Among them, the classification model includes a feature extraction model and a classification module, and of course, other neural network structures may also be included. Optionally, the feature extraction model can be trained first, and after the trained feature extraction model is obtained, the classification model can be further trained on the premise of fixing the model parameters of the feature extraction model (that is, the classification module of the classification model). model parameters for learning). Of course, it is also possible to directly perform end-to-end training on the entire classification model. In this case, the training loss of the classification model may include the total training loss of the feature extraction model and the classification loss of the classification model. The following describes the method of first training the feature extraction model.

Optionally, in this scenario embodiment, the training data set may include a first training set and a second training set, both the first training set and the second training set include a large number of training samples, and the first training set is used for training features. The sample set of the extraction module, and the second training set is the sample set used to train the classification model. For the convenience of description, the training samples in the first training set are referred to as first samples, and the training samples in the second training set are referred to as second samples.

This application does not limit the acquisition method of the sample. Optionally, in order to make the trained classification model more suitable for the above application scenario, the samples in the training data set may be samples corresponding to the application scenario. Optionally, training samples may be constructed based on historical business data corresponding to the application, or simulated business data may be obtained by simulating operating the application, training samples may be constructed based on the simulated business data, or other methods. Taking the purpose of training the classification model to use the trained model to identify whether the target object is an object of the target category as an example, the training samples of multiple categories in the first training set may include training samples of two categories. The categories can be target categories and non-target categories. The first training set includes multiple samples of the target category and multiple samples of non-target categories. A sample of the target category refers to the business data corresponding to a sample object of the target category or based on The service time sequence feature obtained from the service data, similarly, a sample of a non-target category refers to the service data or service sequence feature corresponding to a sample object of the non-target category. Similarly, the training samples in the second training set may also include samples corresponding to sample objects of a specified category and samples corresponding to sample objects of non-specified categories.

Taking the first training set as an example, multiple first samples can be constructed based on multiple designated types of business data of each sample object. Taking the acquisition of one sample as an example, multiple designations of one sample object in multiple time periods can be obtained. Types of business data, such as multiple specified types of transaction data (also called transaction characteristics, which can be the attribute values of various types of business attributes, for example, for payment services, can include payment The number of resources involved), according to the date order of 30 days, according to the transaction data of multiple specified types of these 30 days to obtain the business time sequence characteristics corresponding to the sample object, assuming that the above-mentioned multiple specified types of transaction data are used in total. The attribute value of the 100 business attributes, then the business time sequence feature can be expressed as a 30*100 feature matrix. Therefore, the above business time sequence feature can also be called a business time sequence feature matrix, and 30 is the dimension of the time sequence feature, that is, 30 days , 100 is the dimension of the business feature, that is, 100 attribute values. The above-mentioned business time sequence feature and the real object type of the sample object (that is, the sample label, that is, the category label) can be used as a sample.

As an example, the following table 1 shows an example of multiple specified types of business data corresponding to one sample. As shown in table 1, the example includes m days of business data of n specified types of services, and the attribute value i(i∈[1,n]) represents the attribute value of the i-th specified type of business data, and _Dj (j∈[1,m]) represents the business data of the jth day, then the business data in this example corresponds to The business time sequence feature of can be an m*n feature matrix, and the feature matrix is input into the feature extraction model as a sample, and the corresponding feature vector can be obtained.

Table 1

attribute value 1 attribute value 2 attribute value 3 attribute value 4 … attribute value n D₁ a₁₁ a₁₂ a₁₃ a₁₄ … a_1n D₂ a₂₁ a₂₂ a₂₃ a₂₄ … a_2n … … … … … … … D_m a_m1 a_m2 a_m3 a_m4 … a_mn

Among them, the dimension of the feature vector output by the feature extraction model is determined by the dimension of the output layer of the model, which can be configured according to the actual application requirements, and can be configured according to the actual application requirements. For example, it can be input into the model according to the business requirements. The dimension of the feature adjusts the dimension of the output feature vector of the model. The larger the dimension of the input feature, the larger the dimension of the output feature can be. In theory, the higher the dimension of the output vector, the denser the information contained in the feature vector, but the model training And inference speed will be affected. Among them, in practical applications, it is not recommended that the dimension of the output vector is too low, otherwise the loss of information will be large, and the ability of the feature vector to express the input data will be weak.

Assuming that in practical applications, the purpose of data processing is to identify objects of the first type (that is, objects of the above target category, such as non-compliant objects), the types of objects may include the first type (that is, the first category) and The second type, then the sample label of a sample represents the first type or the second type. If a sample object is an object of the first type, then the sample corresponding to the sample can be regarded as a sample of the first type and the second type. The samples corresponding to the sample object can be considered as samples of the second type, and multiple samples of the first type and multiple samples of the second type constitute the first training set. It can be understood that, in the model training process, the first training set may include samples of two categories or samples of more than two categories. When applying the trained feature extraction model to the classification model, if the purpose of training the model is to identify whether the object is of the first type, the second training set of the classification model can be a sample including two categories, namely the first Multiple samples corresponding to one type of sample object, and multiple samples corresponding to the second type of sample object, so that the two-classification model obtained by training can be better applied to the specific classification task.

The model structure of the feature extraction model is not limited in the embodiment of the present application, which may be any existing feature extraction network, or may be obtained by modifying the existing feature extraction network. As an optional solution, the feature extraction model may select the ResNet50 network (a deep learning network) as the basic network. Since the model is used to extract the feature vector of the input data, the feature extraction model in the embodiment of the present application is a response to the ResNet50 network. The network after the network is modified. Optionally, the modification can include: modifying the last average pooling layer, fully connected layer and Softmax layer of the ResNet50 network to fully connected layer, Batch Normalization (Batch Normalization) layer and full connection layer, in which the parameter information (including the number of neurons, etc.) of neurons in each layer in the modified model can be configured according to actual needs, for example, the parameters of the fully connected layer before and after the above modified batch normalization They can be 2048*1000 and 1024*128 respectively, where 2048 and 1024 respectively represent the dimensions of the input features of the two fully connected layers, and 1000 and 128 respectively represent the dimensions of the output features of the two fully connected layers. Using this parameter configuration, The dimension of the feature vector output by the feature extraction model is 128, that is, a feature vector including 128 feature values.

In the embodiment of the present application, the first training set includes first training samples of multiple categories, and it is assumed that the number of categories is m. First, an initial sample group (triple group) for model training should be constructed based on the first training set. Multiple batch datasets are constructed by sampling, and triples corresponding to each batch dataset are constructed based on the samples in each batch dataset. Optionally, during each sampling, the first training samples of p categories may be randomly selected from the first training set including the first training samples of M categories, and k samples may be randomly selected from each category, that is, one sampling. There are p*k samples in training, that is, the number of samples in a batch dataset is p*k. In the training process, the specific positive samples in the batch are selected as the anchor points in turn, and then the most difficult positive sample and the most difficult negative sample for the anchor point are respectively selected to form a triplet with them. Optionally, a corresponding triplet can be constructed for each sample in the batch dataset, then in one sampling, there are p*k such triples in total.

It should be noted that the initial triplet can be randomly generated, that is, before training the constructed neural network model (feature extraction model before training), for each sample in the batch dataset, it can be randomly generated. A sample belonging to the same category as the sample is selected as a positive sample, and a sample belonging to a different category from the sample is selected as a negative sample, and the three constitute a triplet. During the training process, the new triples corresponding to each sample can be re-determined based on the similarity between the feature vectors of each sample output by the model, that is, for each sample (that is, the anchor point), select the Say the hardest positive sample (the sample of the same category with the lowest similarity to the sample) and the hardest negative sample (the sample of a different category with the highest similarity to the sample), and use the three as the new sample corresponding to the sample. of triples.

Each batch data set is obtained by random sampling, and after constructing p*k initial triples in each batch data set, the neural network model can be iteratively trained based on each batch data set. The settings of the relevant parameters (such as training algebra, learning rate, etc.) used in training can be set according to requirements. As an optional solution, the epoch of training (ie, epoch, all batch datasets participate in one generation of training) can be set to t ₁ =25000, and the network optimizer used during training can be the Adam optimizer. The default learning rate ε=3e-4, and the default values of the two exponential decay rate coefficients β ₁ and β ₂ are 0.9 and 0.999, respectively. In the embodiment of this application, the relevant parameters of the Adam optimizer are optimized, β ₁ is set to 0.5, and for the learning rate, a training algebra threshold t ₀ =15000 is set, and the expression of the learning rate ε(t) is as follows:

Among them, ∈ ₀ = 3e-4, t represents the current training algebra in the training process, ∈(t) represents the current algebra in the training process, that is, when the training algebra reaches t ₀ , the learning rate increases with the training algebra. increase and change, specifically, it is positively correlated, that is, the larger t is, the larger ∈(t) is.

When iteratively training the model, one training of the model based on one batch dataset can be called one iteration. During the training process, each first training sample in the batch of data sets can be input into the neural network model to obtain the feature vector of each first training sample, and according to the features of the first training sample in the p*k triples The first similarity between vectors (for example, it can be based on the distance between the eigenvector of the anchor point and the eigenvector of the positive sample, and the distance between the eigenvector of the anchor point and the eigenvector of the negative sample in each triplet) distance) to calculate the total training loss corresponding to this iterative training, update the parameters of the model, and update the triples in the batch of data sets according to the second similarity between the feature vectors of the first training samples, to obtain p *k new triples, p*k new triples for the next generation training process. Based on each batch of data sets, through the process of continuously training the model and updating triples, until the training algebra reaches the above-mentioned set algebra t ₁ or the corresponding training total loss in the iterative training process converges, the neural network model at this time is used as the The trained feature extraction model, optionally, the model can also be verified or tested, and the model that satisfies the verification or testing conditions is regarded as a trained model. If not, the training can be continued.

After the trained feature extraction model is obtained, the classification model including the feature extraction model may be trained based on the samples in the second training set to obtain the trained classification model. The training method of the classification model is not limited in this embodiment of the present application. Optionally, when the classification model is performed, the model parameters of the trained feature extraction model may be fixed.

Step S13 : deploy the trained classification model to the application server 20 .

Step S21: The application server 20 acquires the data to be processed.

Step S22: the application server 20 identifies the data to be processed through the trained classification model, and obtains a corresponding classification result;

Step S23: The application server sends the classification result to the terminal device 21 to provide the classification result to the administrator.

After obtaining the trained classification model through training, the training server 30 can send the trained classification model to the application server 20, and the application server 20 can identify the data to be processed through the classification model, and can identify the identified data through the terminal device 21. The classification result is provided to the manager, for example, the classification result can be sent to the terminal device 21, and the terminal device 21 can display the classification result to the manager or related personnel through the management client of the application, or when the classification result is the specified result When the classification result is abnormal, send corresponding prompt information to the administrator, for example, when the classification result is abnormal, send a prompt.

In the embodiment of this scenario, the service data to be processed may be the service data of the above-mentioned multiple specified types of services corresponding to the object using the application program when the application program is used, and the application server 20 may obtain the attribute value of each service corresponding to the object to obtain The service sequence feature corresponding to the object is the data to be processed in the application scenario. Based on the service sequence feature, the application server 20 can call the trained classification model and extract the input from the feature extraction model of the classification model. Perform feature extraction on the service sequence feature of the service sequence to obtain a corresponding feature vector, and the classification module of the classification model can obtain a corresponding classification result based on the feature vector, and the classification result represents the object type of the above-mentioned object corresponding to the service sequence feature.

It can be understood that the training method of the feature extraction model and the data processing method provided by the embodiments of the present application may be applicable to but not limited to the above application scenarios. Using the feature extraction model trained by the solution provided in this application can effectively improve the degree of discrimination of the feature vector of the data to be processed extracted by the model, so that when the feature vector is further processed, the processing effect can be effectively improved, such as In the above application scenarios, the classification accuracy can be effectively improved.

The technical solutions of the embodiments of the present application and the technical effects produced by the technical solutions of the present application will be described below by describing several exemplary embodiments. It should be noted that the following embodiments may refer to, learn from, or combine with each other, and the same terms, similar features, and similar implementation steps in different embodiments will not be described repeatedly.

FIG. 3 shows a schematic flowchart of a training method for a feature extraction model provided by an embodiment of the present application. As shown in FIG. 3 , the training method may include the following steps S110 to S130 .

Step S110: Acquire a training set, where the training set includes training samples of multiple categories.

In the embodiment of the present application, the training data set includes training subsets corresponding to multiple (at least two) categories, and each training subset includes multiple training samples belonging to the same category. The present application does not limit the classification method of the categories, and the categories can be divided according to actual application scenarios and application requirements.

The method of acquiring the training set is also not limited in the embodiment of the present application, which may be a training set constructed based on an existing public data set used for classification model training, a manually designed data set, or sample data collected from actual application scenarios. set. Optionally, the training samples in the training set may be a large number of training samples corresponding to the same application scenario, so that the trained feature extraction model can be better applied to the application scenario, and the different parameters in the scenario extracted by the model can be better. The feature vector corresponding to the data to be processed has better discrimination. Of course, the training set may also adopt a large number of training samples corresponding to multiple application scenarios, so that the trained feature extraction model has better generality.

Optionally, in practical applications, the data form of the training samples can be selected according to the actual application requirements. The data that needs to be extracted for feature extraction in the model) corresponds to the data form, that is to say, the training samples can be selected according to the tasks to be solved in the actual application scenario. Optionally, the above-mentioned feature extraction model can be used as a feature extraction module in a classification model, and the classification model is used to extract the feature vector of the first data to be processed through the feature extraction model, and identify the corresponding data of the first data to be processed based on the extracted feature vector. The classification result; wherein, the classification result may be one of multiple specified categories or the corresponding probability of each specified category, and the category corresponding to the highest probability may be determined as the category corresponding to the data to be processed. In this optional solution, the above-mentioned multiple The categories include the above-mentioned multiple specified categories, and a training sample of each specified category is the second data to be processed corresponding to the specified category.

The data form of the first data to be processed is determined by the specific classification task to be solved in the actual business scenario to which the classification model is applied. For example, the classification model is used to identify the category of the object according to the business data of the object in the target application, that is, the feature extraction model is applied to the object classification task, and the feature vector for object classification is to be extracted based on the business data of the object. , the training sample should be the business data of the sample object or the processed business data characteristics obtained by processing the business data (such as the business time series feature matrix in the preceding paragraph), and the training set should include various types of sample objects corresponding to business data or business data characteristics. For another example, the classification model is used to classify text, and the training set can include sample texts (second data to be processed) of various categories (that is, the above-mentioned multiple specified categories), and the trained feature extraction model can Feature extraction is performed on the text to be processed (first data to be processed) in the classification model, and based on the extracted feature vector, the classification module of the classification model predicts the text category of the text to be processed.

It should be noted that the above classification model may be a two-class model or a multi-class model.

Optionally, the classification model may specifically identify whether the first data to be processed is data corresponding to the target category based on the extracted feature vector, that is, the above classification result indicates that the first data to be processed is data corresponding to the target category or corresponds to a non-target category. Correspondingly, the above-mentioned multiple specified categories include a target category and at least one non-target category.

In this optional solution, the classification model is a binary classification model, and its function is to identify whether the data to be processed input into the classification model is the data of the target category (a specific category in the above-mentioned multiple specified categories). During the training process of the extraction model, the training samples of multiple categories include at least training samples corresponding to the target category and training samples corresponding to other categories (categories that are not target categories), and the other categories may be one or more. For example, the classification task corresponding to the classification model is to identify whether the image to be processed is an image of category a. The training set can include multiple images of category a and multiple images that are not of category a, and each image can be used as a training sample.

It can be understood that the classification methods of the above-mentioned multiple specified categories can be pre-divided according to the actual application scenarios and application requirements. For example, the classification model is to identify whether the object is an object of the target category based on the business data of the object. The category can be an object category. Correspondingly, the above-mentioned second data to be processed, that is, the training sample, can also be the business data corresponding to the sample object of a certain object category or the business feature data obtained based on the business data. The above-mentioned first to-be-processed data The data is business data or business feature data corresponding to the target object (object to be identified), and the above classification result indicates whether the target object is an object of the target category. For another example, if the classification model is used for image classification, the above specified category is the image category.

Step S120: Construct multiple sample pairs based on the training set, where the multiple sample pairs include multiple positive sample pairs and multiple negative sample pairs, wherein the positive sample pair includes two training samples of the same category, and the negative sample pair includes different categories of training samples. two training samples.

The multiple sample pairs constructed in this step can be understood as initial sample pairs. During the training process in step S130, the sample pairs will be continuously optimized and updated based on the output of the neural network model, and new samples obtained by each training operation will be updated. The pair of samples on which subsequent training (such as the next time) is based.

The construction method of the initial sample pair is not limited in the embodiment of the present application. For example, it can be randomly generated, that is, any two training samples in the same category can be used as a positive sample pair, and two training samples in different categories can be used as a positive sample pair. as a negative sample pair.

It can be understood that "positive" and "negative" in the above-mentioned positive sample pair and negative sample pair in the embodiments of the present application are relative concepts, and "positive" and "negative" are relative to one sample. For a sample, the samples belonging to the same category as the sample can become the positive samples of the sample, and the samples belonging to different categories from the sample can become the negative samples of the sample. In the sample group described later, the positive samples and negative samples in a sample group are also relative to the anchor points in the sample group, and the samples in the sample group that belong to the same category as the anchor point are the positive samples of the anchor point. Samples that belong to a different category from the anchor point are negative samples of the anchor point.

Step S130: Repeat the training operation on the neural network model based on the training set until a preset condition is met, and use the neural network model that meets the preset condition as a trained feature extraction model.

Wherein, the above preset condition is the training end condition of the model, which may include but is not limited to the convergence of the total training loss corresponding to the neural network model or the number of training times reaching a set number of times. The process of the above training operation is shown in Figure 4, which may include the following steps S131 to step S133.

Step S131: input each training sample in the multiple sample pairs into the neural network model respectively, and obtain the feature vector of each training sample;

Step S132: Determine the total training loss based on the first similarity between the feature vectors of the training samples in each sample pair;

Step S133: If the total training loss does not converge and the number of training times does not reach the set number of times, adjust the model parameters of the neural network model, and determine a plurality of new samples based on the second similarity between the feature vectors of each training sample. and use the new multiple sample pairs as the multiple sample pairs on which subsequent training operations are based.

It is understandable that since the training samples in the positive sample pair are two samples belonging to the same category, and the two training samples in the negative sample pair are two samples belonging to different categories, the purpose of training the model should theoretically be Let the similarity between the feature vectors of the two samples in the positive sample pair (which can be simply referred to as the similarity between the two samples) be as high as possible, and the similarity between the feature vectors of the two samples in the negative sample pair is as high as possible Low, that is, the total training loss represents the degree of difference between each positive sample pair input into the model during the current training operation, and the degree of similarity between each negative sample pair, due to the feature vector of the two samples in the positive sample pair. The similarity between the pairs can reflect the degree of difference between the positive sample pairs, and the similarity between the feature vectors of the two samples in the negative sample pair can reflect the similarity between the negative sample pairs. Therefore, in the training process, it can be According to the first similarity between the two samples in each sample pair, the total training loss of the model is calculated, and the model can be judged based on the total training loss (that is, whether the total training loss converges and whether the number of training times reaches the set number of times) Whether the performance can end the training. Optionally, for each sample pair, the training loss corresponding to the sample pair can be calculated according to the first similarity between the feature vectors of the two samples in the sample pair, and then based on each sample pair The corresponding training loss is the total training loss corresponding to the model, for example, the sum or average of the corresponding training losses of each sample pair is used as the total training loss.

The specific calculation method of the above-mentioned first similarity is not limited in the embodiment of the present application. In theory, any existing method for calculating the similarity between two eigenvectors can be adopted. For example, the Euclidean equation between the two eigenvectors can be calculated. Distance, which is used to represent the first similarity, the greater the distance, the smaller the similarity.

In order to obtain a feature extraction network that meets the needs of practical applications, it is necessary to continuously train the initial neural network model based on the training samples in the training set until the preset training end condition is met, that is, the above preset conditions. It is understandable that the initial neural network model The network model is the feature extraction model to be trained. In the training phase, the input of the neural network model is the training sample, and the output is the feature vector of the training sample. When applying the trained feature extraction model, the input of the model is the data to be processed, and the output is the feature vector of the data to be processed.

The specific neural network structure of the feature extraction model is not limited in the embodiment of the present application, which may be selected according to actual application requirements, and may be a feature extraction model based on any neural network structure used for feature extraction. For example, it may include but not limited to Feature extraction model based on convolutional neural network or feature extraction model based on recurrent neural network, etc.

During continuous training of the neural network model, the preset condition for ending the training may be configured according to actual requirements, which is not limited in this embodiment of the present application. For example, the preset condition may include that the total training loss satisfies a certain condition, that is, the total training loss converges (the condition for judging the convergence of the total training loss can be set according to requirements, such as the total training loss is less than the set value or the total training loss corresponding to two consecutive training operations The difference is less than at least one of the set thresholds, etc.) or the number of training times reaches the set number of times. Optionally, the preset conditions may also include at least one of verification conditions or test conditions. Correspondingly, a verification data set and/or a test data set may also be preconfigured. During the training process of the model, the verification data set may be and/or test data set to evaluate whether the performance of the current model (model after one or more trainings) meets the verification conditions, and based on the evaluation results, it is decided whether to continue training the model. It should be noted that the number of times of training in the preset condition reaches the set number of times means that the training epoch reaches the set epoch, that is, the number of times that all samples in the training set participate in the training reaches the set epoch.

The above-mentioned training methods provided by the embodiments of the present application, after each training operation is performed, in addition to the second similarity based on the positive sample pair (that is, the similarity between the feature vectors of the two samples in the positive sample pair) and the negative In addition to calculating the total loss of the training model for the similarity of the sample pairs, the update of the sample pairs will also be performed. Specifically, the new positive sample pairs and The new negative sample pair makes the similarity between the new positive sample pair as low as possible, and the similarity between the new negative sample pair as high as possible, wherein the similarity between the new positive sample pair corresponding to a training sample is The similarity is not higher than the positive sample pair before the update corresponding to the sample (the positive sample pair corresponding to the sample input into the model in this training), and the similarity between the new negative sample pair corresponding to a training sample is not low For the negative sample pair before the update corresponding to the sample, that is to say, the re-determined positive sample pair and negative sample pair are more difficult sample pairs for the model, by adopting this method to re-determine the sample pair in each training process. Multiple sample pairs for subsequent training operations allow the model to learn the sample pairs that are difficult to learn as much as possible, so that the model can be optimized as much as possible, and the model performance of the feature extraction model obtained by the final training can be improved.

Similarly, the calculation method of the second similarity is not limited in the embodiment of the present application, and can be configured according to actual requirements. similarity.

Optionally, a plurality of new sample pairs are determined based on the second similarity between the feature vectors of each training sample, including:

For each training sample, determine the second similarity between the feature vector of the training sample and the feature vector of each first sample, and take the corresponding first sample with the lowest second similarity and the training sample as one a new pair of positive samples, wherein the first sample is a training sample belonging to the same category as the training sample in each training sample;

For each training sample, determine the second similarity between the feature vector of the training sample and the feature vector of each second sample, and use the corresponding second sample with the highest similarity and the training sample as a new Negative sample pair, wherein the second sample is a training sample belonging to a different category from the training sample in each training sample.

In this optional solution, for each training sample in the training set, a training sample and a training sample of the same category that is least similar to the sample can be used as a positive sample pair, and a training sample and a different sample that is most similar to the sample can be used as a positive sample pair. The training samples of the category are used as negative sample pairs, that is, the most difficult sample combination is used as the training sample pair for the neural network model to learn, so that the trained model can distinguish samples of different categories with high similarity. Samples of the same category with very low similarity can also be identified as the same category, that is, when the feature extraction model trained by the method provided in this application performs feature extraction of the data to be processed, even the samples with very low similarity can be identified as the same category. The similarity of the feature vectors of the two types of data to be processed by the model is relatively high. Even if the two types of data to be processed have a high similarity, the two types of data to be processed by the model output are similar. The similarity of the feature vectors of the data is also relatively low.

It can be understood that, in actual implementation, in addition to the above-mentioned method of constructing a new positive sample pair by selecting the sample with the lowest similarity and constructing a new negative sample pair by selecting the sample with the highest similarity, it can also be expanded based on the same principle. The number of new sample pairs, for example, for a training sample, at least two new positive sample pairs or negative sample pairs corresponding to the training sample can be constructed. The similarity between the training sample and other training samples of the same category can be sorted according to the order of similarity from low to high, and at least two samples with the highest ranking are selected (or according to the set Proportion to select the top sample) and the sample to construct a new positive sample pair, for example, select the first two samples with the lowest similarity, and combine these two samples with the sample to obtain two new positive sample pairs. . Similarly, a new negative sample pair can be constructed in a similar manner. For example, for a training sample, other training samples of the first two different categories with the highest similarity can be selected to construct a new negative sample pair with the training sample respectively. In this way, on the premise that the determined new sample pair is a sample pair with great learning difficulty, the number of training sample pairs can also be expanded.

As an optional solution, the above-mentioned construction of multiple sample pairs based on the training samples in the training set may include:

Each training sample in the training set is used as an anchor point, and a sample group corresponding to each anchor point is constructed. The sample group corresponding to each anchor point includes a positive sample pair and a negative sample pair corresponding to the anchor point, wherein one anchor point corresponds to The positive sample pair includes the anchor point and the anchor point's positive sample, and the negative sample pair corresponding to an anchor point includes the anchor point and the anchor point's negative sample;

Correspondingly, the above-mentioned determination of the total training loss based on the first similarity between the feature vectors of the training samples in each sample pair may include:

For each sample group, according to the first similarity between the feature vectors of the two samples in the positive sample pair of the sample group and the first similarity between the feature vectors of the two samples in the negative sample pair of the sample group degree, determine the training loss corresponding to the sample group; determine the total training loss according to the training loss corresponding to each sample group;

The above-mentioned determination of multiple new sample pairs based on the second similarity between the feature vectors of each training sample may include:

Based on the second similarity between the feature vectors of each training sample, a new sample group corresponding to each anchor point is determined, and the sample pairs in the new sample group corresponding to each anchor point are used as multiple sample pairs in subsequent training operations .

In this optional solution, the form of a sample group can be used for training the neural network model. Among them, each sample group is a triple, including an anchor point (that is, a training sample), a positive sample corresponding to the anchor point, and a negative sample corresponding to the anchor point, that is, the anchor point corresponds to The positive sample is a training sample that belongs to the same category as the anchor point, and the negative sample corresponding to the anchor point is a training sample that belongs to a different category from the anchor point, that is, a sample group is three training samples based on the anchor point The sample consists of two sample pairs.

Among them, the construction method of constructing the initial sample group (the sample group constructed before being input into the model) based on the training set is not limited in this embodiment of the present application. An anchor point can be randomly selected from the training samples belonging to the same category as the anchor point as the positive sample corresponding to the anchor point, and randomly selected from the training samples belonging to different categories with the anchor point as the anchor point. Negative samples, get the sample group corresponding to the anchor point.

For the neural network model, the purpose of training the model is to make the similarity between the feature vectors of the positive sample pairs in each sample group learned by the model as high as possible, and the similarity between the feature vectors of the negative sample pairs The similarity is as low as possible, that is, the distance between positive sample pairs is shortened as much as possible, and the distance between negative sample pairs is widened, and the learned feature vector has a good degree of class distinction.

As a schematic illustration, Figure 5 shows a schematic diagram of the principle of training a feature extraction model based on triples. As shown in Figure 5, the length of the line segment between the anchor point and the positive sample in the figure represents the anchor point and the positive sample. Similarity between samples, the length of the line segment between the anchor point and the negative sample represents the similarity between the anchor point and the negative sample, wherein, the longer the line segment, the smaller the similarity. The distance between the anchor point on the left side of the figure and its corresponding positive sample, as well as the distance between the negative sample and the anchor point, can be understood as the sample group before training, and the distance between the anchor point on the right side of the figure and the positive sample The distance between the anchor point and the negative sample can be understood as the distance between the anchor point output by the trained feature extraction model and the sample pair obtained from the feature vector of the positive sample and the negative sample. From Figure 5 It can be seen that the purpose of training the model is to improve the similarity between feature vectors of samples of the same category and reduce the similarity between feature vectors of samples of different categories.

In order to achieve the above purpose, in the training process, for each sample group, it can be based on the similarity between the positive sample pairs in the sample group (that is, the relationship between the anchor point in the sample group and the feature vector of the corresponding positive sample) similarity) and the similarity between the negative sample pairs in the sample group to calculate the training loss corresponding to the sample group, the training loss corresponding to a sample group characterizes the difference between the positive sample pairs in the sample group and the negative samples difference between pairs. In the embodiment of the present application, the form of the loss function of the neural network model is not limited in the embodiment of the present application. For example, it may include but not limited to the triplet loss (triplet loss) function, which is based on the positive sample pair in the triplet The distance between (representing the similarity between sample pairs) and the distance between negative sample pairs is used to calculate the training loss corresponding to the triplet.

Optionally, for each sample group, according to the first similarity between the feature vectors of the two samples in the positive sample pair of the sample group and the sum of the feature vectors of the two samples in the negative sample pair of the sample group. The first similarity between the samples determines the training loss corresponding to the sample group, which may include:

Determine the first distance between the feature vectors of the two samples in the positive sample pair of the sample group and the second distance between the feature vectors of the two samples in the negative sample pair of the sample group, where the first distance characterizes the the first similarity corresponding to the positive sample pair of the sample group, and the second distance represents the first similarity corresponding to the negative sample pair of the sample group;

determining the difference between the first distance and the second distance;

The training loss corresponding to the sample group is determined according to the difference, wherein the training loss corresponding to the sample group is positively correlated with the difference.

Since the larger the distance between the two feature vectors, the more dissimilar the two feature vectors are. If the distance is smaller, the more similar the feature vectors are. Therefore, the difference between the feature vectors of the two training samples in the sample pair can be calculated. The distance represents the similarity between sample pairs. The larger the distance corresponding to the sample pair, the lower the similarity. Because the goal of training the model is to make the similarity between the eigenvectors of the positive sample pair in the sample group output by the model as high as possible and the similarity between the eigenvectors of the negative sample pair as low as possible, that is, as close as possible to the positive sample pair in the sample group. The distance between the sample and the anchor point, and the distance between the negative sample and the anchor point as far as possible, therefore, for each sample group, it can be based on the first distance corresponding to the positive sample and the anchor point in the sample group and the negative sample and the anchor point. The difference between the second distances corresponding to the anchor point is used to calculate the training loss corresponding to the sample group. In the embodiment of the present application, the training loss corresponding to each sample group is positively correlated with the difference between the first distance and the second distance corresponding to the sample group. The smaller the difference, the smaller the loss, which can be understood as the smaller the loss. The smaller the distance between the positive samples and the anchor points in the sample group, the larger the distance between the negative samples and the anchor points.

The optional method provided by the embodiment of the present application provides an improved triplet loss function, and the expression of the triplet loss corresponding to an existing sample group (randomly selected from the existing sample group) is as follows:

(d(a,p)-d(a,n)+α) ₊ (1)

Among them, d(a, p) represents the first distance between the anchor point and the positive sample in the sample group, d(a, n) represents the second distance between the anchor point and the negative sample in the sample group, and α is a preset Parameter value, + means that when the value in brackets is greater than or equal to 0, the loss is the value in brackets, and the value in brackets is less than 0, the loss is 0.

It can be seen from the above expression that during the training process of the existing triplet loss function, for each triplet, if the relationship of the triplet determined based on the output of the model is correct (ie d(a, p) -d(a, n)+α is less than 0), directly set the loss corresponding to the triplet to 0, the loss calculation adopts the hard cutoff method, d(a, p)-d(a, n is not considered )+α is less than the loss corresponding to the triplet of 0, but these triples also have an impact on the training results of the model, when d(a, p)-d(a, n)+α is a negative number , although the distance between the positive sample pairs in the triplet is smaller than the distance between the negative sample pairs, if the training loss is also assigned to the triplet, that is, it is considered that the feature vector of the positive sample pair output by this model is the sum of the The distance between the distance and the distance between the feature vectors of the negative sample pair still needs to be optimized (that is, the difference between the distances can still be increased), then the performance of the model can also be improved, and the existing triplet loss ignores this. The training loss corresponding to some triples.

In view of the above problems, the above-mentioned solution for calculating the training loss corresponding to the sample group provided by the embodiment of the present application takes into account the training loss corresponding to all the sample groups input into the model when calculating the total training loss corresponding to the model. The training loss corresponding to the sample group is positively correlated with the difference between the first distance between the positive sample pairs and the second distance between the negative sample pairs in the sample group, that is, the first distance minus the second distance, the larger the value is The loss is relatively larger, and the smaller the value is, the smaller the loss is. Wherein, the difference and the loss may be a linear change relationship or a non-linear change relationship, and the specific form of the loss function may be designed and selected based on the principle of the solution provided in this application.

As an optional solution, the training loss corresponding to each sample group is determined based on the following expression:

s( ^x )=ln(1+ex ) (2)

x=d(a,p)-d(a,n)+β(3)

Among them, s(x) represents the training loss corresponding to the sample group, a, p and n represent the anchor points, positive samples and negative samples in the sample group, respectively, d(a, p) represents the first distance, d(a, n ) represents the second distance, and β represents the preset adjustment threshold.

It can be seen from expressions (2) and (3) that in this optional solution of the present application, the training loss corresponding to the sample group changes exponentially, rather than a hard cutoff, and this solution is used to calculate the total training loss corresponding to the model. , and using the loss to constrain the parameter adjustment of the model, the trained model can better shorten the distance of similar samples in the feature embedding space, and further improve the performance of the model.

In the training method provided in this embodiment of the present application, for each training operation, after calculating the training loss corresponding to each sample group based on the feature vector of each training sample output by the model, the model can be calculated based on the training loss corresponding to each sample group. The corresponding total training loss, for example, the sum or average of the training losses corresponding to each sample group can be used as the total training loss, and further it can be judged whether the total training loss satisfies the preset conditions, and if the preset conditions are met, the model training can be ended. If the preset conditions are not met, the model parameters can be adjusted. For example, the gradient descent algorithm can be used to adjust the model parameters, and the model can be trained based on the new sample group.

It can be understood that, for this optional solution, a new sample pair is determined based on the similarity between the feature vectors of the training samples output by the model, that is, a new sample group corresponding to each anchor point is determined, and a new sample group corresponding to an anchor point is determined. The sample group of includes a new positive sample pair corresponding to the anchor point and a new negative sample group corresponding to the anchor point.

Optionally, the above-mentioned determination of a new sample group corresponding to each anchor point based on the second similarity between the feature vectors of each training sample may include:

For each anchor point, the second similarity between the feature vector of the anchor point and the feature vector of each first sample is respectively determined, and the corresponding first sample with the lowest second similarity is determined as the corresponding first sample of the anchor point The new positive sample of , wherein, the first sample is the training sample that belongs to the same category as the anchor point in each training sample;

For each anchor point, determine the second similarity between the feature vector of the anchor point and the feature vector of each second sample, and determine the corresponding second sample with the highest second similarity as the new one corresponding to the anchor point A negative sample, wherein the second sample is a training sample that belongs to a different category from the anchor point in each training sample.

Similarly, in actual implementation, for each anchor point, a new sample group corresponding to the anchor point can be determined, or multiple new sample groups can be determined. Specifically, for an anchor point, the same category can be determined. At least one sample with the lowest similarity to the anchor point in the training samples of , is used as the positive sample corresponding to the anchor point, and each positive sample is combined with the anchor point to obtain the corresponding positive sample pair, and the training samples of different categories are At least one sample with the highest similarity with the anchor point is used as the negative sample corresponding to the anchor point, and each negative sample is combined with the anchor point to obtain the corresponding negative sample pair, and a positive sample pair corresponding to the anchor point and a The negative sample pair is combined to obtain a sample group. For example, if there are two positive samples and one negative sample, then there are two positive sample pairs and one negative sample pair. The two positive sample pairs can be combined with negative sample pairs respectively, Get the two sample groups corresponding to the anchor point. By using the solution provided by the embodiment of the present application, the expansion of the sample group can be realized conveniently and quickly, and it can be ensured that the expanded sample group is a relatively difficult sample combination. Further improve the expression ability of the feature vector output by the trained model.

As an optional solution, each training sample in the training set is used as an anchor point, and a sample group corresponding to each anchor point is constructed, which may include:

According to the training set, construct at least one batch data set, wherein each batch data set includes training samples of p categories, and the number of training samples of each category is k, where p≥2, k≥3;

For each batch of datasets, each training sample in the batch of datasets is used as an anchor point, and based on each training sample in the batch of datasets, a sample group corresponding to each anchor point in the batch of datasets is constructed;

Wherein, the above-mentioned repeated training operation on the neural network model based on the training set may include:

Repeatedly perform training operations on the neural network model based on each batch of data sets, and each training operation is performed based on a sample group corresponding to each anchor point in a batch of data sets;

Correspondingly, based on the second similarity between the feature vectors of each training sample, a new sample group corresponding to each anchor point is determined, including:

For each anchor point in the batch dataset corresponding to the current training operation, determine the new anchor point corresponding to the anchor point according to the second similarity between the anchor point and each training sample except the anchor point in the batch dataset. sample group.

For model training, since the number of training samples in the training set is usually relatively large, if all the samples in the training set are used for each iteration training, the computational cost will be relatively large. It divides the training set into multiple batch data sets, a batch data set is also a batch, and each iteration training can be performed based on a batch data set. By adopting this solution, the amount of calculation when calculating the total loss of the model training each time can be reduced, and the stability of the convergence of the loss function of the model can also be ensured.

In the embodiment of the present application, the size of the batch data set (that is, the total number of training samples in the batch data set) is not limited in the embodiment of the present application, as long as there are at least two categories of training data in each batch data set and each The number of training data for each category is not less than three. After the training set is divided into multiple batch data sets, the initial sample group can be constructed separately for each batch data set, that is, the three training samples in each sample group constructed are all in the same batch data set. , Similarly, when a new sample group is determined during the training process, a new sample group corresponding to each anchor point in the batch data set is also determined from the same batch data set.

It is understandable that when using batch data sets for model training operations, each batch data set must participate in the training of the model, and each batch data set participates in multiple trainings, and an iterative training is performed based on one batch data set. , input each sample in the batch of datasets into the neural network model, and obtain the feature vector of each sample in the batch of datasets. Based on the output vector of the model, the training loss corresponding to each sample group in each batch of datasets can be calculated, thereby obtaining The total training loss of the model. If the total training loss does not meet the preset conditions, the sample group corresponding to each anchor point can be updated based on the similarity between the feature vectors of each sample output by the model to obtain a new sample group, and Each new sample group obtained is used as each sample group in the batch of datasets for subsequent training operations.

Among them, when the model is trained based on the batch data set, the total training loss corresponding to the model (that is, the training loss corresponding to all triples in the batch data set) can be expressed as follows:

s(α)=ln( ¹ +ex )

表示锚点α对应的三元组中的正样本对的特征向量之间的距离，

锚点α对应的三元组中的负样本对的特征向量之间的距离，可以理解的是，一个批数据集第一次被输入到模型中时，

和

为根据模型的输出计算出的锚点a的初始的三元组中样本对所对应的距离，除了第一次输入到模型中之外，

和

则是锚点a的新的三元组(即最困难的样本组合)中样本对所对应的距离。Among them, L _th represents the total training loss, and s(α) represents the training loss corresponding to an anchor point α in the batch dataset, that is, the training loss of the triplet corresponding to the anchor point α. A batch dataset can be composed of p×k three triples, that is, each sample in the batch dataset is used as an anchor, and the triple corresponding to each anchor is constructed,

represents the distance between the feature vectors of the positive sample pair in the triplet corresponding to the anchor point α,

The distance between the feature vectors of the negative sample pairs in the triplet corresponding to the anchor point α, it can be understood that when a batch dataset is first input into the model,

and

is the distance corresponding to the sample pair in the initial triplet of anchor point a calculated according to the output of the model, except for the first input into the model,

and

is the distance corresponding to the sample pair in the new triplet of anchor point a (ie the most difficult sample combination).

It should be noted that, in practical applications, it is assumed that the nth training operation is performed based on batch data set 1. After the training operation is completed and each new sample group corresponding to batch data set 1 is obtained, the n+th One operation may be performed based on a new sample group of batch data set 1, or may be based on original sample groups corresponding to other batch data sets or new sample groups of other batch data sets. That is to say, in the training process of the neural network model, which batch of data sets each training operation is based on is not limited in this embodiment of the present application, as long as each batch of data sets can participate in the training of the model multiple times, for example , the minimum number of times each batch of data sets participates in model training can be preset, and the number of batches of data sets participating in model training can be roughly the same, so that the number of times each training sample in the training set participates in model training is basically balanced.

Optionally, suppose that the training set is divided into three batch data sets by sampling the training set, denoted as S1, S2 and S3, then when training the model, you can poll the data sets respectively. Use the initial triples in S1, S2 and S3 to train the model in turn (all three batches of data sets participate in one training operation, that is, a generation of training is completed), after which you can use S1, S2 and S3 respectively. The corresponding updated triples in the training operation are performed sequentially, and a trained feature extraction model that satisfies the preset conditions is obtained through continuous repeated training.

The trained feature extraction model obtained by using the training method provided in the embodiment of the present application can extract a feature vector with good discrimination when performing feature extraction on the data to be processed input into the model, so that the feature vector can be extracted based on the The feature vector realizes further processing of the data to be processed. For example, it can identify the type of the data to be processed. It can also judge the similarity between different data based on the feature vectors of different data to be processed, and can also be used for the classification of data sets. and many more. The data to be processed may be data in various forms, including but not limited to text, characters (such as numerical values), images, or other forms of data. The trained feature extraction model can be applied to any scene that needs to extract feature vectors with good distinguishing ability. For example, the feature extraction model can be used as a feature in other models (such as classification models, similarity judgment models, etc.) Extract modules.

In order to test the effect of the feature extraction model trained by using the training method provided in the embodiment of the present application, after obtaining a feature extraction model that satisfies the preset conditions based on the training set, that is, the embedding model, the model is tested on the cross-time test set , extract the latent vector (that is, the feature vector) of the test data in the test set. The test result shows that the iv value of the model (Information Value, which is used to measure the predictive ability of the feature) is as high as 2.7. In addition, on the cross-time test set, the feature vector extracted by the feature extraction model and the feature vector extracted by the feature extraction model trained by the existing technology are spliced together, and sent to traditional machine learning models such as xgboost. , which is used for the binary classification task to test the classification effect of the feature. The test results show that the AUC (Area Under Curve, that is, the area under the ROC curve) index of the test set under the xgboost model can be improved by 5 percentage points, and the KS value (one model evaluation metrics) increased by 4 percentage points. It has been proved by experiments that by using the training method provided by the embodiment of the present application to train a feature extraction model applied to a specific scene, a latent vector feature with a good degree of discrimination for the classification task can be extracted. The embodiment of the present application also provides a data processing method, as shown in FIG. 6 , the method may include the following steps:

Step S210: obtaining data to be processed;

Step S220: Input the data to be processed into the first feature extraction model, and extract the first feature vector of the data to be processed through the first feature extraction model, wherein the first feature extraction model adopts any optional embodiment of the present application obtained from the training method in ;

Step S230: Based on the first feature vector, determine the classification result corresponding to the data to be processed.

The form of the data to be processed is not limited in this embodiment of the present application. For different application scenarios, the form of the data to be processed may be different. For example, the data to be processed may be text, images, or feature matrices obtained by processing business data. For different application scenarios and application requirements, the form of data processing results can also be different. For example, if you want to identify the type corresponding to the data to be processed, the data classification result corresponding to the data processing result, if you want to determine the type of data to be processed Whether the data is similar to other data, the similarity can be calculated based on the first eigenvector of the data to be processed and the eigenvectors of other data, and the data processing result is the calculated similarity or is determined according to the similarity and the set threshold. The similarity judgment result. In this embodiment, the description is given by applying the feature extraction model trained by the training method provided by the embodiment of the present application to data classification as an example. The feature vector extracted by the feature extraction model can effectively improve the accuracy of the classification result. sex.

Optionally, the data processing method may further include: extracting a second feature vector of the data to be processed by using a second feature extraction model;

Correspondingly, in the above step S230, based on the first feature vector, determining the data processing result corresponding to the data to be processed may include:

fusing the first feature vector and the second feature vector;

Based on the fused features, a data processing result of the data to be processed is determined.

In order to further improve the accuracy of the classification result, in this optional method of the present application, the feature vectors of the data to be processed obtained by extraction of various feature extraction models can be fused, and the classification result can be determined by fusing the features. It has been proved by testing that the feature vector of the data to be processed extracted by the feature extraction model trained by the training method provided in the embodiment of the present application and the feature vector of the data to be processed extracted by other feature extraction models are fused and used for the classification task. , which can effectively improve the accuracy of the final classification result.

The model structure and training method of the second feature extraction model are not limited in the embodiments of the present application, and may be a feature extraction model trained by using an existing training method. The specific fusion method of the feature vectors is not limited in the embodiment of the present application, which may include, but is not limited to, at least one of splicing, adding, or calculating the mean value of the first feature vector and the second feature vector.

In an optional embodiment of the present application, the above-mentioned obtaining of the data to be processed may include:

Obtain business data of the target business in multiple time periods corresponding to the target object, and the business data corresponding to each time period includes an attribute value of at least one business attribute of the target business;

Based on the business data corresponding to the above multiple time periods, a business time series feature matrix corresponding to the target object is constructed, and the business feature matrix is used as the data to be processed. The above classification results represent the object type of the target object

The number of rows in the service sequence feature matrix is the number of time periods in multiple time periods, the number of columns is the number of attributes of at least one service attribute, and each element value in the service sequence feature matrix represents the attribute of a service attribute corresponding to a time period value.

It can be known from the application scenario embodiments provided in the foregoing that the data processing method provided in the embodiments of the present application can be applied to the type recognition scenario of the target object. Optionally, the above classification can be obtained through a classification model including the above-mentioned first feature extraction model. The identification result, for example, the above-mentioned business sequence feature matrix can be input into the classification model, the first feature vector of the feature matrix can be extracted through the first feature extraction model of the classification model, and the second feature vector of the feature matrix can be extracted through the second feature extraction model. , the first feature vector and the second feature vector are spliced together, and a classification result is obtained through the classification module, and the specific structure of the classification module is not limited in this embodiment of the present application. Optionally, the classification model may include, but is not limited to, a classification model based on traditional machine learning models such as xgboost.

Based on the same principle as the training method provided by the embodiment of the present application, the embodiment of the present application further provides a training device for a feature extraction model. As shown in FIG. 7 , the training device 100 includes a training data acquisition module 110, a training data processing module module 120 and model training module 130.

A training data acquisition module 110, configured to acquire a training set, where the training set includes training samples of multiple categories;

The training data processing module 120 is configured to construct multiple sample pairs based on the training set, where the multiple sample pairs include multiple positive sample pairs and multiple negative sample pairs, wherein the positive sample pairs include two training samples of the same category, and the negative sample pairs include For two training samples including different classes;

The model training module 130 is used to repeatedly perform training operations on the neural network model based on the training set until a preset condition is met, and the neural network model that meets the preset condition is used as a trained feature extraction model; wherein, the above-mentioned preset conditions include neural network models. The total training loss corresponding to the network model converges or the number of training times reaches the set number of times. The above training operations include:

Input each training sample in the multiple sample pairs into the neural network model, respectively, to obtain the feature vector of each training sample; determine the total training loss based on the first similarity between the feature vectors of the training samples in each sample pair; If the total training loss does not converge and the number of training times does not reach the set number of times, the model parameters of the neural network model are adjusted, and multiple new sample pairs are determined based on the second similarity between the feature vectors of each training sample, and Use the new multiple sample pairs as the multiple sample pairs on which subsequent training operations are based.

Optionally, the model training module can be used to: for each training sample, respectively determine the second similarity between the feature vector of the training sample and the feature vector of each first sample, and set the corresponding second similarity to be the lowest. The first sample and the training sample are taken as a new positive sample pair, wherein the first sample is a training sample belonging to the same category as the training sample in each training sample; for each training sample, determine the training sample respectively The second similarity between the feature vector of the second sample and the feature vector of each second sample, the corresponding second sample with the highest second similarity and the training sample are taken as a new negative sample pair, where the second sample is each The training samples in the training samples belong to different categories from the training samples.

Optionally, the training data processing module can be used to: use each training sample in the training set as an anchor point, and construct a sample group corresponding to each anchor point, and the sample group corresponding to each anchor point includes a positive sample pair corresponding to the anchor point. and a negative sample pair, wherein the positive sample pair corresponding to an anchor point includes the anchor point and the positive sample of the anchor point, and the negative sample pair corresponding to an anchor point includes the anchor point and the negative sample of the anchor point;

Correspondingly, when determining the total training loss, the model training module can be used to: for each sample group, according to the first similarity between the feature vectors of the two samples in the positive sample pair of the sample group, and the The first similarity between the feature vectors of the two samples in the negative sample pair determines the training loss corresponding to the sample group; the total training loss is determined according to the training loss corresponding to each sample group;

When determining a plurality of new sample pairs, the model training module can be used to: determine a new sample group corresponding to each anchor point based on the second similarity between the feature vectors of each training sample, and assign the new sample group corresponding to each anchor point. The sample pairs in the sample group are used as multiple sample pairs in subsequent training operations.

Optionally, when determining the new sample group corresponding to each anchor point, the model training module can be used to: for each anchor point, respectively determine the number 1 between the feature vector of the anchor point and the feature vector of each first sample. The second similarity is to determine the first sample with the lowest second similarity as the new positive sample corresponding to the anchor point, wherein the first sample is the training sample that belongs to the same category as the anchor point in each training sample ; For each anchor point, determine the second similarity between the feature vector of the anchor point and the feature vector of each second sample, and determine the second sample with the highest second similarity as the new anchor point corresponding to the second sample. The negative sample of , wherein the second sample is a training sample that belongs to a different category from the anchor point in each training sample.

Optionally, the training data processing module can be used to construct at least one batch data set according to the training set when constructing the sample group corresponding to each anchor point, wherein each batch data set includes training samples of p categories, and each batch data set includes training samples of p categories. The number of training samples of each category is k, where p≥2, k≥3; for each batch of data sets, each training sample in the batch of data sets is used as an anchor point, based on the Training samples to construct a sample group corresponding to each anchor point in the batch of datasets;

Correspondingly, the model training module can be used to: repeatedly perform training operations on the neural network model based on each batch of data sets, wherein each training operation is performed based on a sample group corresponding to each anchor point in a batch of data sets;

When determining a new sample group corresponding to each anchor point, the model training module can be used to: for each anchor point in the batch data set corresponding to the current training operation, according to the anchor point and the batch data set except the anchor point The second similarity between the training samples of , determines the new sample group corresponding to the anchor point.

Optionally, for each sample group, when determining the training loss corresponding to the sample group, the model training module can be used to:

Determine the first distance between the feature vectors of the two samples in the positive sample pair of the sample group and the second distance between the feature vectors of the two samples in the negative sample pair of the sample group, where the first distance characterizes the the first similarity corresponding to the positive sample pair of the sample group, and the second distance represents the first similarity corresponding to the negative sample pair of the sample group;

Determine the difference between the first distance and the second distance; determine the training loss corresponding to the sample group according to the difference, wherein the training loss corresponding to the sample group is positively correlated with the difference.

Optionally, the training loss corresponding to each sample group is determined based on the following expression:

s( ^x )=ln(1+ex )

x=d(a,p)-d(a,n)+β

Among them, s(x) represents the training loss corresponding to the sample group, a, p and n represent the anchor points, positive samples and negative samples in the sample group, respectively, d(a, p) represents the first distance, d(a, n ) represents the second distance, and β represents the preset adjustment threshold.

Based on the same principle as the data processing method provided by the embodiment of the present application, the embodiment of the present application further provides a data processing apparatus based on a neural network model. As shown in FIG. 8 , the data processing apparatus 200 includes a data acquisition module 210 and a Data processing model 220 .

a data acquisition module 210, configured to acquire data to be processed;

The data processing module 220 is configured to input the data to be processed into the first feature extraction model, extract the first feature vector corresponding to the data to be processed through the first feature extraction model, and obtain the corresponding data to be processed through the classification module based on the first feature vector. , wherein the first feature extraction model is obtained by training using the training method provided in any optional embodiment of the present application.

Optionally, the data processing module 220 may be specifically configured to: extract the second feature vector of the data to be processed by using the second feature extraction model; fuse the first feature vector and the second feature vector; The fused features determine the classification result corresponding to the data to be processed.

Optionally, the data acquisition module can be used to: acquire the business data of the target business in multiple time periods corresponding to the target object, and the business data corresponding to each time period includes the attribute value of at least one business attribute of the target business; The business data of the target object is constructed, the business feature matrix corresponding to the target object is constructed, and the business feature matrix is used as the data to be processed. The above classification results represent the object type of the target object; wherein, the number of rows of the business feature matrix is the number of time periods in multiple time periods, The number of columns is the attribute number of the at least one service attribute, and each element value in the service feature matrix represents an attribute value of a service attribute corresponding to a time period.

The apparatuses in the embodiments of the present application can execute the methods corresponding to the apparatuses provided in the embodiments of the present application, and the implementation principles thereof are similar. Corresponding to the steps in the corresponding method in , for the detailed functional description of each module of the apparatus, reference may be made to the description in the corresponding method shown above, and details are not repeated here. The training apparatus and the data processing apparatus provided in the embodiments of the present application may be any electronic device.

The embodiment of the present application also provides an electronic device, the electronic device may include a memory, a processor, and a computer program stored in the memory, and when the processor executes the computer program stored in the memory, any optional optional device of the present application may be implemented. methods in the examples.

Optionally, FIG. 9 shows a schematic structural diagram of an electronic device to which an embodiment of the present invention is applicable. As shown in FIG. 9 , the electronic device may be a server or a user terminal, and the electronic device may be used to implement any The method provided in an embodiment.

As shown in FIG. 9 , the electronic device 2000 may include at least one component such as a processor 2001 , a memory 2002 , a communication module 2003 and an input/output interface 2004 . It should be noted that the structure of the electronic device 2000 shown in FIG. 9 is only schematic, and does not constitute a limitation on the electronic device to which the methods provided in the embodiments of the present application are applicable.

The memory 2002 may be used to store an operating system and an application program, etc. The application program may include a computer program for implementing the method shown in the embodiment of the present invention when called by the processor 2001, and may also include a program for implementing other functions or services. . The memory 2002 can be ROM (Read Only Memory, read only memory) or other types of static storage devices that can store static information and instructions, RAM (Random Access Memory, random access memory) or other types that can store information and computer programs It can also be EEPROM (Electrically Erasable Programmable Read Only Memory, Electrically Erasable Programmable Read Only Memory), CD-ROM (CompactDisc Read Only Memory, CD-ROM) or other CD-ROM storage, CD-ROM storage (including compressed CD-ROM) , laser disc, compact disc, digital versatile disc, Blu-ray disc, etc.), magnetic disk storage medium or other magnetic storage device, or any other device capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer Other media, but not limited to this.

The processor 2001 is connected to the memory 2002 through the bus 2005, and implements corresponding functions by calling the application program stored in the memory 2002. The processor 2001 may be a CPU (Central Processing Unit, central processing unit), a general-purpose processor, a DSP (Digital Signal Processor, data signal processor), an ASIC (Application Specific Integrated Circuit, an application-specific integrated circuit), an FPGA (FieldProgrammable Gate) Array, field programmable gate array) or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof, which can implement or execute the various exemplary logic blocks, modules, and logic blocks described in connection with the present disclosure. circuit. The processor 2001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

The electronic device 2000 can be connected to the network through the communication module 2003 (which may include but not limited to components such as a network interface), so as to communicate with other devices (such as user terminals or servers, etc.) through the network to realize data interaction, such as sending data to other devices. data or receive data from other devices. The communication module 2003 may include a wired network interface and/or a wireless network interface, etc., that is, the communication module may include at least one of a wired communication module or a wireless communication module.

The electronic device 2000 can be connected to required input/output devices, such as keyboards, display devices, etc., through the input/output interface 2004 . Optionally, a storage device, such as a hard disk, can also be connected through the interface 2004, so that the data in the electronic device 2000 can be stored in the storage device, or the data in the storage device can be read, and the data in the storage device can also be stored. stored in the memory 2002. It can be understood that the input/output interface 2004 may be a wired interface or a wireless interface. According to different actual application scenarios, the device connected to the input/output interface 2004 may be a component of the electronic device 2000, or may be an external device connected to the electronic device 2000 when needed.

The bus 2005 used to connect the components may include a path to transfer information between the components. The bus 2005 may be a PCI (Peripheral Component Interconnect, Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture, Extended Industry Standard Architecture) bus or the like. According to different functions, the bus 2005 can be divided into an address bus, a data bus, a control bus, and the like.

Optionally, for the solutions provided by the embodiments of the present invention, the memory 2002 may be used to store a computer program for executing the solutions of the present invention, and be executed by the processor 2001. When the processor 2001 runs the computer program, the present invention is implemented. Examples of acts of the method or apparatus provided.

Embodiments of the present application further provide a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the corresponding content of the foregoing method embodiments can be implemented.

Embodiments of the present application further provide a computer program product, the product including a computer program, and when the computer program is executed by a processor, the corresponding content of the foregoing method embodiments can be implemented.

It should be noted that the terms "first", "second", "third", "fourth", "1", "2", etc. in the description and claims of the present application and the above drawings (if existence) is used to distinguish similar objects and is not necessarily used to describe a particular order or sequence. It is to be understood that the data so used are interchangeable under appropriate circumstances so that the embodiments of the application described herein can be practiced in sequences other than those illustrated or described in the text.

It should be understood that, although the respective operation steps are indicated by arrows in the flowcharts of the embodiments of the present application, the execution order of these steps is not limited to the order indicated by the arrows. Unless explicitly stated herein, in some implementation scenarios of the embodiments of the present application, the implementation steps in each flowchart may be performed in other sequences as required. In addition, some or all of the steps in each flowchart are based on actual implementation scenarios, and may include multiple sub-steps or multiple stages. Some or all of these sub-steps or stages may be executed at the same time, and each of these sub-steps or stages may also be executed at different times respectively. In scenarios with different execution times, the execution order of these sub-steps or stages may be flexibly configured according to requirements, which is not limited in this embodiment of the present application.

The above are only optional implementations of some implementation scenarios of the present application. It should be pointed out that for those of ordinary skill in the art, without departing from the technical concept of the solution of the present application, the application based on the technical concept of the present application is adopted. Other similar implementation means also belong to the protection scope of the embodiments of the present application.

Claims (15)

1. a training method of a feature extraction model, is characterized in that, comprises: obtaining a training set, the training set includes training samples of multiple categories; Construct a plurality of sample pairs based on the training set, the plurality of sample pairs include a plurality of positive sample pairs and a plurality of negative sample pairs, wherein the positive sample pairs include two training samples of the same category, and the negative sample pairs for two training samples including the different classes; Repeat the training operation on the neural network model based on the training set until a preset condition is met, and use the neural network model that satisfies the preset condition as a trained feature extraction model; wherein the preset condition includes the neural network model The total training loss corresponding to the network model converges or the number of training times reaches the set number of times, and the training operations include: Input each training sample in the multiple sample pairs into the neural network model respectively, and obtain the feature vector of each training sample; Determine the total training loss based on the first similarity between the feature vectors of the training samples in each of the sample pairs; If the total training loss does not converge and the number of training times does not reach the set number of times, the model parameters of the neural network model are adjusted, and based on the second similarity between the feature vectors of the respective training samples, a plurality of new sample pairs, and use the new multiple sample pairs as multiple sample pairs on which subsequent training operations are based. 2. The method according to claim 1, wherein, determining a plurality of new sample pairs based on the second similarity between the feature vectors of the respective training samples, comprising: For each training sample, determine the second similarity between the feature vector of the training sample and the feature vector of each first sample, respectively, and compare the corresponding first sample with the lowest second similarity to the training sample As a new positive sample pair, wherein the first sample is a training sample belonging to the same category as the training sample in each of the training samples; For each training sample, the second similarity between the feature vector of the training sample and the feature vector of each second sample is determined respectively, and the corresponding second sample with the highest second similarity and the training sample are regarded as one A new pair of negative samples, wherein the second sample is a training sample belonging to a different category from the training sample among the training samples. 3. The method according to claim 1, wherein the constructing a plurality of sample pairs based on the training samples in the training set comprises: Each training sample in the training set is used as an anchor point, and a sample group corresponding to each anchor point is constructed, and the sample group corresponding to each anchor point includes a positive sample pair and a negative sample pair corresponding to the anchor point, wherein, The positive sample pair corresponding to an anchor point includes the anchor point and the positive sample of the anchor point, and the negative sample pair corresponding to an anchor point includes the anchor point and the negative sample of the anchor point; Determining the total training loss based on the first similarity between the feature vectors of the training samples in each of the sample pairs, including: For each sample group, according to the first similarity between the feature vectors of the two samples in the positive sample pair of the sample group and the first similarity between the feature vectors of the two samples in the negative sample pair of the sample group degree, determine the training loss corresponding to the sample group; Determine the total training loss according to the training loss corresponding to each of the sample groups; Determining a plurality of new sample pairs based on the second similarity between the feature vectors of the respective training samples, including: Based on the second similarity between the feature vectors of the respective training samples, a new sample group corresponding to each of the anchor points is determined, and the sample pair in the new sample group corresponding to each of the anchor points is used as a subsequent training operation multiple sample pairs. 4. The method according to claim 3, wherein the determining a new sample group corresponding to each of the anchor points based on the second similarity between the feature vectors of the respective training samples, comprising: For each anchor point, the second similarity between the feature vector of the anchor point and the feature vector of each first sample is determined respectively, and the corresponding first sample with the lowest second similarity is determined as the anchor A new positive sample corresponding to the point, wherein the first sample is a training sample belonging to the same category as the anchor point in each of the training samples; For each anchor point, determine the second similarity between the feature vector of the anchor point and the feature vector of each second sample, and determine the corresponding second sample with the highest second similarity as the corresponding second sample of the anchor point A new negative sample, wherein the second sample is a training sample belonging to a different category from the anchor point among the training samples. 5. The method according to claim 3 or 4, characterized in that, using each training sample in the training set as an anchor point, and constructing a sample group corresponding to each anchor point, comprising: According to the training set, construct at least one batch data set, wherein each batch data set includes training samples of p categories, and the number of training samples of each category is k, where p≥2, k ≥3; For each of the batch data sets, each training sample in the batch data set is used as an anchor point, and based on each training sample in the batch data set, a sample group corresponding to each anchor point in the batch data set is constructed; The repeatedly performing training operations on the neural network model based on the training set includes: Repeatedly perform training operations on the neural network model based on each of the batch data sets, wherein each training operation is performed based on a sample group corresponding to each anchor point in one of the batch data sets; Determining a new sample group corresponding to each anchor point based on the second similarity between the feature vectors of each training sample, including: For each anchor point in the batch dataset corresponding to the current training operation, determine the new anchor point corresponding to the anchor point according to the second similarity between the anchor point and each training sample except the anchor point in the batch dataset. sample group. 6. The method according to claim 3 or 4, wherein, for each sample group, the first similarity between the feature vectors of two samples in the positive sample pair according to the sample group, and the The first similarity between the feature vectors of the two samples in the negative sample pair of the sample group determines the training loss corresponding to the sample group, including: Determine the first distance between the feature vectors of the two samples in the positive sample pair of the sample group and the second distance between the feature vectors of the two samples in the negative sample pair of the sample group, where the first distance represents the first similarity corresponding to the positive sample pair of the sample group, and the second distance represents the first similarity corresponding to the negative sample pair of the sample group; determining the difference between the first distance and the second distance; The training loss corresponding to the sample group is determined according to the difference, wherein the training loss corresponding to the sample group is positively correlated with the difference. 7. The method according to claim 6, wherein the training loss corresponding to each of the sample groups is determined based on the following expression: s( ^x )=ln(1+ex ) x=d(a,p)-d(a,n)+β Among them, s(x) represents the training loss corresponding to the sample group, a, p and n represent the anchor points, positive samples and negative samples in the sample group, respectively, d(a, p) represents the first distance, d(a, n ) represents the second distance, and β represents the preset adjustment threshold. 8. A data processing method based on a neural network model, characterized in that, comprising: Get data to be processed; Input the data to be processed into a first feature extraction model, and extract the first feature vector corresponding to the data to be processed through the first feature extraction model, wherein the first feature extraction model is based on claims 1 to Obtained by the method training described in any one of 7; Based on the first feature vector, a classification result corresponding to the data to be processed is determined. 9. The method according to claim 8, wherein the method further comprises: Extract the second feature vector of the data to be processed by using the second feature extraction model; The determining the classification result corresponding to the data to be processed based on the first feature vector includes: fusing the first feature vector and the second feature vector; Based on the fused features, a classification result corresponding to the data to be processed is determined. 10. The method according to claim 8 or 9, wherein the acquiring the data to be processed comprises: obtaining business data of the target business in multiple time periods corresponding to the target object, and the business data corresponding to each time period includes an attribute value of at least one business attribute of the target business; Based on the business data corresponding to the multiple time periods, construct a business time series feature matrix corresponding to the target object, and use the business time series feature matrix as the data to be processed, wherein the classification result represents the target object's characteristic matrix. object type; Wherein, the number of rows of the service sequence feature matrix is the number of time periods of the plurality of time periods, the number of columns is the number of attributes of the at least one service attribute, and each element value in the service sequence feature matrix represents a The attribute value of a business attribute corresponding to the time period. 11. A training device for a feature extraction model, comprising: a training data acquisition module for acquiring a training set, the training set including training samples of multiple categories; A training data processing module, configured to construct a plurality of sample pairs based on the training set, the plurality of sample pairs include a plurality of positive sample pairs and a plurality of negative sample pairs, wherein the positive sample pairs include two samples of the same category training samples, the negative sample pair includes two training samples of the different categories; A model training module, configured to repeatedly perform training operations on the neural network model based on the training set until a preset condition is met, and use the neural network model that satisfies the preset condition as a trained feature extraction model; wherein the preset The set conditions include that the total training loss corresponding to the neural network model converges or the number of training times reaches a set number of times, and the training operations include: Input each training sample in the multiple sample pairs into the neural network model respectively, and obtain the feature vector of each training sample; Determine the total training loss based on the first similarity between the feature vectors of the training samples in each of the sample pairs, where the total training loss represents the degree of difference between the positive sample pairs and the negative samples in the plurality of sample pairs degree of similarity between pairs; If the total training loss does not converge and the number of training times does not reach the set number of times, the model parameters of the neural network model are adjusted, and based on the second similarity between the feature vectors of the respective training samples, a plurality of new sample pairs, and use the new multiple sample pairs as multiple sample pairs on which subsequent training operations are based. 12. A data processing device based on a neural network model, comprising: The data acquisition module is used to acquire the data to be processed; a data processing module, configured to input the data to be processed into a first feature extraction model, extract a first feature vector corresponding to the data to be processed through the first feature extraction model, and determine based on the first feature vector the classification result corresponding to the data to be processed; Wherein, the first feature extraction model is obtained by training using the method of any one of claims 1 to 7. 13. An electronic device, characterized in that the electronic device comprises a memory and a processor, wherein a computer program is stored in the memory, and the processor executes the computer program to realize any one of claims 1 to 7 The method described or the method described in any one of claims 8 to 10 is implemented. 14. A computer-readable storage medium, characterized in that, a computer program is stored in the storage medium, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is realized or the right The method of any one of claims 8 to 10. 15. A computer program product, characterized in that the computer product comprises a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7 or implements claims 8 to 10 The method of any of the above.

Images