CN108399431A

CN108399431A - Disaggregated model training method and sorting technique

Info

Publication number: CN108399431A
Application number: CN201810167276.7A
Authority: CN
Inventors: 孙源良; 王亚松; 刘萌; 樊雨茂
Original assignee: Guoxin Youe Data Co Ltd
Current assignee: Guoxin Youe Data Co Ltd
Priority date: 2018-02-28
Filing date: 2018-02-28
Publication date: 2018-08-14

Abstract

A kind of disaggregated model training method of the application offer and sorting technique, training method includes the aiming field common characteristic vector for the source domain common characteristic vector sum target numeric field data for capturing source domain data and inputs the first grader, obtains the first classification results of source domain data and the first classification results of target numeric field data；It captures the aiming field difference characteristic vector of the source domain difference characteristic vector sum target numeric field data of source domain data and inputs the second grader, obtain the second classification results of source domain data and the second classification results of target numeric field data；Second classification results of the second classification results of the first classification results, target numeric field data based on source domain data, the first classification results and source domain data of source domain data carry out epicycle training to neural network, common characteristic trapping layer, the first grader.This method can simultaneously use the same characteristic features and difference characteristic of source domain data and target numeric field data, and the disaggregated model that training obtains can carry out target numeric field data more accurately classification.

Description

Disaggregated model training method and sorting technique

Technical field

This application involves depth learning technology fields, in particular to a kind of disaggregated model training method and classification Method.

Background technology

Transfer learning can utilize training sample (can be described as source domain data) the training classification mould for having label in known art Type is demarcated come the data (can be described as target numeric field data) to target domain, and is not required for source domain data and target numeric field data Data distribution having the same.Transfer learning is indeed through the connection between looking for data and known label data to be calibrated Source domain and target numeric field data, are mapped in the same space, the source domain data under the space by system for example, by using the mode of kernel function Possess identical distribution with target numeric field data, so as to utilize the source domain sample data training point for having label of the space representation Class device demarcates target domain.

In existing transfer learning method, same neural network is trained usually using source domain data and target numeric field data, is obtained Obtain parameter sharing network.The network training process can find out the general character between source domain data and target numeric field data, usually by source Numeric field data and target numeric field data are mapped in the high bit space with high comparativity, to obtain the distribution characteristics in two domains.It is this Although training method can use the same characteristic features between source domain data and target numeric field data, make source domain and aiming field Between difference characteristic lose serious, cause the disaggregated model that training obtains to exist when classifying to target numeric field data certain Error.

Invention content

In view of this, the embodiment of the present application is designed to provide a kind of disaggregated model training method and sorting technique, The same characteristic features and difference characteristic of source domain data and target numeric field data can be used simultaneously, the disaggregated model that training obtains More accurately classification can be carried out to target numeric field data.

In a first aspect, providing a kind of disaggregated model training method, this method includes：

Obtain the source domain data for carrying label and not the target numeric field data of tape label；

The source domain data and the target numeric field data are inputted into same neural network, are the source domain data extraction source Characteristic of field vector, and extract target domain characterization vector for the target numeric field data；

The source domain feature vector and the target domain characterization vector are inputted into common characteristic trapping layer, and will be captured respectively Source domain data source domain common characteristic vector sum target numeric field data aiming field common characteristic vector input the first grader, obtain To the first classification results of source domain data and the first classification results of target numeric field data；And

The source domain feature vector and the target domain characterization vector are inputted into difference characteristic trapping layer, and will be captured respectively Source domain data source domain difference characteristic vector sum target numeric field data aiming field difference characteristic vector input the second grader, obtain To the second classification results of source domain data and the second classification results of target numeric field data；

Between the first classification results and the second classification results of the target numeric field data based on the target numeric field data Comparison between comparison result and the first classification results and the second classification results of the source domain data of the source domain data As a result, carrying out epicycle training to the neural network, the common characteristic trapping layer and first grader；

By carrying out more wheel training to the neural network, the common characteristic trapping layer and first grader, Obtain disaggregated model.

Second aspect, provides a kind of disaggregated model training device, which includes：

Acquisition module is used to obtain the source domain data for carrying label and not the target numeric field data of tape label；

First processing module, for by the source domain data and the same neural network of target numeric field data input, being The source domain data extraction source characteristic of field vector, and extract target domain characterization vector for the target numeric field data；

Second processing module, for catching the source domain feature vector and target domain characterization vector input common characteristic Layer is caught, and respectively by the aiming field common characteristic vector of the source domain common characteristic vector sum target numeric field data of the source domain data of capture The first grader is inputted, the first classification results of source domain data and the first classification results of target numeric field data are obtained；And

Third processing module, for catching the source domain feature vector and target domain characterization vector input difference characteristic Layer is caught, and respectively by the aiming field difference characteristic vector of the source domain difference characteristic vector sum target numeric field data of the source domain data of capture The second grader is inputted, the second classification results of source domain data and the second classification results of target numeric field data are obtained；

Training module, second point for the first classification results and the target numeric field data based on the target numeric field data First classification results of comparison result and the source domain data between class result and the second classification of the source domain data are tied Comparison result between fruit carries out epicycle to the neural network, the common characteristic trapping layer and first grader Training；By carrying out more wheel training to the neural network, the common characteristic trapping layer and first grader, obtain To disaggregated model.

The third aspect, provides a kind of sorting technique, and this method includes：

Obtain data to be sorted；

The data to be sorted are input to the classification mould that the disaggregated model training method provided by first aspect obtains In type, the classification results of the data to be sorted are obtained；

Wherein, the disaggregated model includes：The neural network, the common characteristic trapping layer and first classification Device.

Fourth aspect provides a kind of sorter, which includes：Data acquisition module to be sorted waits for point for obtaining Class data；

Sort module, for the data to be sorted to be input to the disaggregated model training side provided by first aspect In the disaggregated model that method obtains, the classification results of the data to be sorted are obtained；

In the disaggregated model training method that the embodiment of the present application is provided, when being trained to disaggregated model, profit After the source domain feature vector of neural network extraction source numeric field data and the target domain characterization vector of target numeric field data, it can be based on Source domain feature vector and target domain characterization vector, use common characteristic extract layer and difference characteristic extract layer, to source domain number respectively According to the capture for the capture and difference characteristic for carrying out common characteristic with target numeric field data, and using grader to carrying out common characteristic The classification results for capturing the feature vector generated after being captured with difference characteristic, are trained disaggregated model, so that classification Model can not only learn the common characteristic to source domain data and target numeric field data, additionally it is possible to which source domain data and aiming field number are arrived in study According to difference characteristic, to by between source domain data and target numeric field data common characteristic and difference characteristic all use, obtain The disaggregated model arrived can carry out target numeric field data more accurately classification.

In addition, when the data volume of target numeric field data is less, just with being total to for source domain data and target numeric field data There is feature to be trained disaggregated model, to classification of the guidance to target numeric field data, the shortage of target numeric field data can cause point Class model can equally cause the inaccuracy of disaggregated model without feature enough in calligraphy learning to aiming field.The embodiment of the present application exists When being trained to disaggregated model, when the data volume of target numeric field data is less, disaggregated model is to source domain data When study with the common characteristic of target numeric field data, can also it learn to the otherness of source domain data and target numeric field data spy Sign so that disaggregated model can learn into target numeric field data more features, to be carried out more to target numeric field data Accurately classification.

To enable the above objects, features, and advantages of the application to be clearer and more comprehensible, preferred embodiment cited below particularly, and coordinate Appended attached drawing, is described in detail below.

Description of the drawings

It, below will be to needed in the embodiment attached in order to illustrate more clearly of the technical solution of the embodiment of the present application Figure is briefly described, it should be understood that the following drawings illustrates only some embodiments of the application, therefore is not construed as pair The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this A little attached drawings obtain other relevant attached drawings.

Fig. 1 shows a kind of flow chart for disaggregated model training method that the embodiment of the present application is provided；

In the disaggregated model training method provided Fig. 2 shows the embodiment of the present application, the first Classification Loss determines operation Specific method flow chart；

Fig. 3 shows in the disaggregated model training method that the embodiment of the present application is provided that the second Classification Loss determines operation Specific method flow chart；

Fig. 4 shows a kind of flow chart for sorting technique that the embodiment of the present application is provided；

Fig. 5 shows a kind of structural schematic diagram for disaggregated model training device that the embodiment of the present application is provided；

Fig. 6 shows a kind of structural schematic diagram for sorter that the embodiment of the present application is provided；

Fig. 7 shows a kind of structural schematic diagram for computer equipment that the embodiment of the present application is provided.

Specific implementation mode

Unlike the prior art, the application is that source domain data and target numeric field data distinguish extraction source using same neural network After characteristic of field vector sum target domain characterization vector, source domain feature vector and target domain characterization vector input common characteristic are captured Layer is source domain data capture source domain common characteristic vector, is aiming field data capture aiming field common characteristic vector, and by source Characteristic of field vector sum target domain characterization vector inputs difference characteristic trapping layer, simultaneously for source domain data capture source domain difference characteristic vector For aiming field data capture aiming field difference characteristic vector, it is total to be then based on captured source domain common characteristic vector, aiming field There are feature vector, source domain difference characteristic vector and aiming field difference characteristic vector to be trained disaggregated model, it is obtained Disaggregated model can use the same characteristic features between source domain data and target numeric field data and by source domain data and mesh Difference characteristic between mark numeric field data uses, and then energy when classified to target numeric field data using the disaggregated model Access more accurate classification results.

For ease of understanding the present embodiment, first to a kind of disaggregated model training side disclosed in the embodiment of the present application Method describes in detail, and this method is used for the training to the disaggregated model of various data, and the disaggregated model of gained can be to correspondence Data are classified.

Shown in Figure 1, the disaggregated model training method that the embodiment of the present application is provided includes：

S101：Obtain the source domain data for carrying label and not the target numeric field data of tape label.

When specific implementation, source domain data are the data with label, and target numeric field data is the number without label According to.Source domain data and target numeric field data have certain general character, and have certain otherness.Label is the classification to source domain data As a result advance mark.

Source domain data and target numeric field data can be that image, video, language etc. may be used neural network learning and divided The data of class.

Such as when source domain data and target numeric field data are image data, source domain data can make a definite diagnosis the state of an illness Clearly medical image picture, the title of the position of the lesion that clearly medical image picture is marked and disease i.e. For the label of source domain data；Target numeric field data can not make a definite diagnosis unsharp medical image picture of the state of an illness.Training gained Disaggregated model can classify to the unintelligible medical image picture for not marking lesions position and not yet diagnosed disease name, Under the premise of obtaining whether having lesion in the unintelligible medical image picture, and have lesion in medical image picture, disease is determined The position of stove.

In another example when source domain data and target numeric field data are language data, source domain data are French vocabulary, aiming field Data are Spanish vocabulary, since French and Spanish belong to Romance, have part common between the two Feature；But the two belongs to two different language, therefore has certain difference again.Using the French that can be identified to western class The feature of tooth language is learnt, so as to identify Spanish.

In another example when source domain data and target numeric field data are language data, emotion is carried out to some vocabulary or words art Analysis；Source domain data are to be labelled with the vocabulary of affective tag, and target numeric field data is not mark art if affective tag.

S102：Source domain data and target numeric field data are inputted into same neural network, are source domain data extraction source characteristic of field Vector, and extract target domain characterization vector for target numeric field data.

When specific implementation, convolutional neural networks (Convolutional Neural may be used in neural network Network, CNN) that for source domain data extraction source characteristic of field vector, target domain characterization vector is extracted for target numeric field data.

Source domain data are to carry the data of label, which is used to indicate the correct classification results of source domain data；Target Numeric field data is not carry the data of label.After source domain data and target numeric field data are inputted the same neural network, the nerve Network carries out source domain data and target numeric field data the feature learning of shared parameter.In this process, due to neural network Supervised learning is carried out to source domain data, and unsupervised learning is carried out to target numeric field data, in the same neural network of use to source Numeric field data and target numeric field data carry out in the learning process of shared parameter, can constantly adjust the parameter used in neural network, To during to neural metwork training, allow neural network parameter by source domain data influence while, also suffer from The influence of target numeric field data, so that neural network is to source domain data and target numeric field data after carrying out feature learning, to every A source domain data, which carry out the obtained source domain feature vector of feature extraction, to be interfered by target numeric field data so that be source domain number It can be with the feature of partial target numeric field data according to the source domain feature vector extracted；Likewise, to each target numeric field data into The obtained target domain characterization vector of row feature extraction can be interfered by source domain data so that extract target for target numeric field data Domain vector can be mixed with the feature of part source domain data, final realize between source domain data and the domain of target numeric field data.

S103：Source domain feature vector and target domain characterization vector are inputted into common characteristic trapping layer, and respectively by capture The aiming field common characteristic vector of the source domain common characteristic vector sum target numeric field data of source domain data inputs the first grader, obtains First classification results of source domain data and the first classification results of target numeric field data.

When specific implementation, due to the shared spy of disaggregated model to be made study to source domain data and target numeric field data It seeks peace difference characteristic, therefore, it is necessary to when being trained to disaggregated model, using two trained branches come to source domain data It is captured respectively with the common characteristic and difference characteristic of target numeric field data.

Source domain feature vector and target domain characterization vector are inputted into common characteristic trapping layer, capture source domain in the following way The aiming field common characteristic vector of the source domain common characteristic vector sum target numeric field data of data：

Step 2 one：Source domain feature vector and target domain characterization vector are input to common characteristic trapping layer, using altogether There is Feature capturing layer extraction source domain common characteristic vector sum aiming field common characteristic vector.

Step 2 two：Source domain common characteristic vector sum aiming field common characteristic vector is subjected to gradient reverse process.

Step 2 three：It will carry out the source domain common characteristic vector sum aiming field common characteristic vector input of gradient reverse process To the first domain grader；

Step 2 four：Table is distinguished to source domain common characteristic vector sum aiming field common characteristic vector according to the first domain grader The source domain data of sign and the domain classification results of target numeric field data, adjust the parameter of neural network and common characteristic trapping layer It is whole.

When specific implementation, common characteristic trapping layer can be a full articulamentum, be on the basis of neural network The one layer of feature convergence-level added, the feature that can export neural network carry out dimension transformation.Common characteristic trapping layer is right After current source domain feature vector is handled, the corresponding source domain common characteristic vector of current source domain feature vector can be obtained；Altogether There is characteristic layer after current goal characteristic of field vector is handled, the corresponding aiming field of current goal characteristic of field vector can be obtained Common characteristic vector.

Due in the mistake being trained to neural network and common characteristic trapping layer using source domain data and target numeric field data Source domain data and target numeric field data are actually carried out the process of domain mixing by journey.It is caught using neural network and common characteristic Catch layer the source domain common characteristic vector that is obtained of feature extraction is carried out to source domain data will be by the shadow of feature in target numeric field data It rings, namely so that source domain feature vector is close to the feature of target numeric field data；Meanwhile it being caught using neural network and common characteristic Catch the shadow that layer carries out the aiming field common characteristic vector acquired in feature extraction by feature in source domain data to target numeric field data It rings, namely so that aiming field common characteristic vector is close to the feature of source domain data.Therefore, in order to realize to source domain data and The domain of target numeric field data mixes, and aiming field common characteristic vector is being extracted for each target numeric field data in target numeric field data, and After extracting source domain common characteristic vector for each source domain data in source domain data, by aiming field common characteristic vector sum source domain Common characteristic vector carries out gradient reverse process, then by the aiming field common characteristic vector sum source domain Jing Guo gradient reverse process Common characteristic vector is input to the first domain grader, shared to aiming field common characteristic vector sum source domain using the first domain grader Feature vector carries out domain classification.

The result of domain classification is correct namely the first domain grader can be correctly total to source domain common characteristic vector sum aiming field There is the probability that feature vector is correctly classified bigger, then illustrates that the degree of domain mixing is smaller；Domain classification result mistake it is general Rate is bigger namely domain grader gets over the source domain common characteristic vector sum aiming field common characteristic vector correct probability that classify It is small, illustrate that the degree of domain mixing is bigger, therefore, it is shared to aiming field common characteristic vector sum source domain to be based on the first domain grader It is that the source domain data and target numeric field data that feature vector characterizes respectively are classified as a result, being captured to neural network and common characteristic Layer carries out parameter adjustment.

Specifically, source domain common characteristic vector sum aiming field common characteristic vector is characterized respectively according to the first domain grader Source domain data and target numeric field data domain classification results, the parameter of neural network and common characteristic trapping layer is adjusted It is whole, it can specifically be realized by executing following first domain Classification Loss determination operation.The first domain Classification Loss determines operation Including：

Step 3 one：Determine the source domain number that current source domain common characteristic vector sum aiming field common characteristic vector characterizes respectively According to the domain Classification Loss of this domain classification with target numeric field data.

Herein, the degree of domain mixing is characterized by the way that domain classification is loss.The domain Classification Loss of source domain data, can pass through Domain classification is being carried out to source domain data and target numeric field data based on source domain common characteristic vector sum aiming field common characteristic vector In the process, classification results are the quantity for the source domain data being assigned in aiming field to characterize.The domain Classification Loss of target numeric field data It can be by being carried out to source domain data and target numeric field data based on source domain common characteristic vector sum aiming field common characteristic vector During domain is classified, classification results are the quantity for the target numeric field data being assigned in source domain to characterize.Using the first domain point The source domain data and target numeric field data that class device characterizes source domain common characteristic vector and aiming field common characteristic vector respectively into After the classification of row domain, it will be able to obtain domain classification results, then just according to domain classification results and source domain data and target numeric field data True domain ownership determines that source domain data close the corresponding domain Classification Loss of target numeric field data.

Step 3 two：It is not less than default differential threshold for the difference between the domain Classification Loss of nearest preset times, it is raw Parameter adjustment is carried out to neural network and common characteristic trapping layer at the first feedback information, and based on the first feedback information.

Herein, it is constrained come the degree mixed to domain using default differential threshold.It is prestored in first domain grader The distribution for having the domain that source domain common characteristic vector sum aiming field common characteristic vector belongs to respectively, when the domain of nearest preset times point When difference between class loss is not less than default differential threshold, then it is assumed that the also not up to stable state of domain classification, that is to say, that In certain domain is classified, the first domain grader can correctly distinguish source domain common characteristic vector sum aiming field common characteristic vector difference Affiliated domain, in the classification of certain domain, domain grader cannot correctly distinguish source domain common characteristic vector sum aiming field common characteristic again Domain belonging to vector difference, domain mixability is also unstable, then, need the parameter to neural network and common characteristic trapping layer It is adjusted, therefore the first excessive feedback information of domain Classification Loss difference can be generated, and feed back to neural network and shared spy Levy trapping layer.Neural network and common characteristic trapping layer are receiving the first excessive feedback information of the domain Classification Loss difference Afterwards, the parameter of itself is adjusted respectively so that the result of domain classification tends towards stability.

Step 3 three：The use of neural network and common characteristic trapping layer is that source domain data carry based on the parameter after adjustment New source domain common characteristic vector is taken, and extracts new aiming field common characteristic vector for target numeric field data, and executes domain classification The determining operation of loss, up to difference is more than default differential threshold, completion to neural network and is shared based on the first domain grader The epicycle of Feature capturing layer is trained.

Training based on the first domain grader to neural network, being will be according to the first domain grader to source domain common characteristic Domain Classification Loss maintains in certain value determined by the classification results of vector sum aiming field common characteristic vector, divides as far as possible It does not know that target numeric field data and source domain data are to belong to source domain to still fall within aiming field on earth, extracts the common characteristic of the two.

Herein, it should be noted that when the difference between the domain Classification Loss of nearest preset times is less than default difference threshold When value, the suitable feedback information of domain Classification Loss can be also generated, and it is fed back into neural network and common characteristic trapping layer.God Through network and common characteristic trapping layer after receiving the suitable feedback information of domain Classification Loss, can also to the parameter of itself into The smaller adjustment of line amplitude, makes every effort to gradient and drops to local optimum.

Herein, it should be noted that the corresponding second domain Classification Loss of following Fig. 5 such as can also be used to determine that operation is similar Method, come the source domain characterized respectively to source domain common characteristic vector sum aiming field common characteristic vector according to the first domain grader The domain classification results of data and target numeric field data, are adjusted the parameter of neural network and common characteristic trapping layer.

This method may include steps of：

Determine source domain data and target that current source domain common characteristic vector sum aiming field common characteristic vector characterizes respectively The domain Classification Loss of this domain classification of numeric field data；

For the situation of domain classification results mistake, third feedback information is generated, and based on third feedback information to nerve net Network and common characteristic trapping layer carry out parameter adjustment；

Based on the parameter after adjustment, new source domain is extracted for source domain data using neural network and difference characteristic trapping layer and be total to There is feature vector, and extract new aiming field common characteristic vector for target numeric field data, and executes the domain Classification Loss and determine behaviour Make.

Detailed process can be found in following step May Day to step 5 three and describe, and details are not described herein.

S104：Source domain feature vector and target domain characterization vector are inputted into difference characteristic trapping layer, and respectively by capture The aiming field difference characteristic vector of the source domain difference characteristic vector sum target numeric field data of source domain data inputs the second grader, obtains Second classification results of source domain data and the second classification results of target numeric field data.

When specific implementation, due to when being trained to disaggregated model, come pair using two trained branches The common characteristic and difference characteristic of source domain data and target numeric field data are captured respectively.Above-mentioned S103 be used as to source domain data and The training branch that the common characteristic of target numeric field data is captured, this step are special to the difference of source domain data and target numeric field data Levy the training branch captured.

Specifically, source domain feature vector and target domain characterization vector are inputted into difference characteristic trapping layer, in the following way Capture the aiming field difference characteristic vector of the source domain difference characteristic vector sum target numeric field data of source domain data：

Step 4 one：Source domain feature vector and target domain characterization vector are inputted into difference characteristic trapping layer, use difference The extraction of Feature capturing layer obtains source domain difference characteristic vector sum aiming field difference characteristic vector.

Step 4 two：Source domain difference characteristic vector sum aiming field difference characteristic vector is input to the second domain grader；

Step 4 three：Table is distinguished to source domain difference characteristic vector sum aiming field difference characteristic vector according to the second domain grader The source domain data of sign and the domain classification results of target numeric field data, adjust the parameter of neural network and difference characteristic trapping layer It is whole.

When specific implementation, difference characteristic trapping layer is a full articulamentum, can be on the basis of neural network The one layer of feature convergence-level added, the feature that can export neural network carry out dimension transformation.Difference characteristic trapping layer is right After current source domain feature vector is handled, the corresponding source domain difference characteristic vector of current source domain feature vector can be obtained；Difference Different characteristic layer can obtain the corresponding aiming field of current goal characteristic of field vector after current goal characteristic of field vector is handled Difference characteristic vector.

Gradient reverse process is not being carried out to source domain feature vector and target domain characterization vector, but the two is being directly inputted Difference characteristic trapping layer, and the source domain difference characteristic vector sum aiming field for the source domain data that difference characteristic trapping layer is exported respectively The aiming field difference characteristic vector of data is input to the second domain grader, using the second domain grader to source domain difference characteristic vector And the source domain data that characterize respectively of aiming field difference characteristic vector and target numeric field data carry out domain classification, obtained domain classification It loses smaller, the domain belonging to source domain data and target numeric field data can be distinguished as far as possible so that neural network and difference Feature capturing layer can capture the difference characteristic between source domain data and target numeric field data, zoom out distance between the two.

Specifically, source domain difference characteristic vector sum aiming field difference characteristic vector is characterized respectively according to the second domain grader Source domain data and target numeric field data domain classification results, the parameter of neural network and difference characteristic trapping layer is adjusted It is whole, it can specifically be realized by executing following second domain Classification Loss determination operation.The second domain Classification Loss determines operation Including：

Step 5 one：Determine the source domain number that current source domain difference characteristic vector sum aiming field difference characteristic vector characterizes respectively According to the domain Classification Loss of this domain classification with target numeric field data.

Herein, it is mixed by domain Classification Loss to characterize source domain difference characteristic vector sum aiming field difference characteristic vector progress domain The degree of conjunction.The domain Classification Loss of source domain data herein can be by based on source domain difference characteristic vector sum aiming field difference During feature vector classifies to source domain data and target numeric field data, classification results are the source domain data of target numeric field data Quantity characterize.The domain Classification Loss of target numeric field data can be by based on source domain difference characteristic vector sum aiming field difference During feature vector classifies to source domain data and target numeric field data, classification results are the target numeric field data of source domain data Quantity characterize.Table is being distinguished to source domain difference characteristic vector and aiming field difference characteristic vector using the second domain grader After the source domain data and target numeric field data of sign carry out domain classification, it will be able to domain classification results are obtained, then according to domain classification results, Determine source domain data and the corresponding domain Classification Loss of target numeric field data.

Step 5 two：For the situation of domain classification results mistake, the second feedback information is generated, and be based on the second feedback information Parameter adjustment is carried out to neural network and difference characteristic trapping layer.

Herein, due to the correctness of domain classification results to be ensured, only domain classification results are correct, can just zoom out source domain The distance between data and target numeric field data also can just extract the otherness data between source domain data and target numeric field data, Therefore to generate the second feedback information when the classification results mistake of domain, and based on the second feedback information to neural network and Difference characteristic trapping layer carries out parameter adjustment.

Step 5 three：Based on the parameter after adjustment, extracted for source domain data using neural network and difference characteristic trapping layer New source domain difference characteristic vector, and new aiming field difference characteristic vector is extracted for target numeric field data, and execute the second domain point Class loss determines operation.

Up to domain classification results are correct or the accuracy of domain classification results reaches preset threshold value.

Herein, it should be noted that above-mentioned step 3 one to three corresponding first domain of step 3 such as can also be used to classify Loss, which determines, operates similar method, according to the second domain grader to source domain difference characteristic vector sum aiming field difference characteristic to The domain classification results for measuring the source domain data and target numeric field data that characterize respectively, to the ginseng of neural network and difference characteristic trapping layer Number is adjusted.

This method may include steps of：

Determine source domain data and target that current source domain difference characteristic vector sum aiming field difference characteristic vector characterizes respectively The domain Classification Loss of this domain classification of numeric field data；

It is not less than default differential threshold for the difference between the domain Classification Loss of nearest preset times, generates the 4th feedback Information, and parameter adjustment is carried out to neural network, difference characteristic trapping layer based on the 4th feedback information；

The use of neural network and difference characteristic trapping layer is that source domain data extract new source domain based on the parameter after adjustment Difference characteristic vector, and new aiming field difference characteristic vector is extracted for target numeric field data, and execute domain Classification Loss and determine behaviour Make, until difference is more than default differential threshold, completes based on the second domain grader to neural network and difference characteristic trapping layer Epicycle training.

S105：The first classification results based on target numeric field data and the comparison between the second classification results of target numeric field data As a result, and the comparison result between the first classification results and the second classification results of source domain data of source domain data, to nerve Network, common characteristic trapping layer and the first grader carry out epicycle training.

When specific implementation, neural network and common characteristic trapping layer are instructed based on the first domain grader While white silk, the source domain common characteristic vector sum aiming field common characteristic vector that common characteristic trapping layer is extracted can also be distinguished It is input to the first grader, obtains the first classification results of source domain data and the first classification results of target numeric field data.Meanwhile While being trained to neural network and difference characteristic trapping layer based on the second domain grader, difference characteristic can also be captured The source domain difference characteristic vector sum aiming field difference characteristic vector of layer extraction is separately input into the second grader, obtains source domain data The second classification results and target numeric field data the second classification results.

Below source domain common characteristic vector is input to the first grader, the first classification results of extraction source numeric field data are Example is illustrated the work of the first grader and the second grader：

When the source domain data characterized to source domain common characteristic vector using the first grader are classified, it will usually Using to classification function, such as softmax classification functions.Classification function can obtain when source domain data are classified The source domain data belong to the probability of some classification, and the classification results by the classification of maximum probability as output.

Classified to aiming field common characteristic vector using the first grader, is special to source domain difference using the second grader The method that sign vector classified, is classified to aiming field difference characteristic vector using the second grader is similar, herein no longer It repeats.

The embodiment of the present application also provides the second of a kind of the first classification results based on target numeric field data and target numeric field data Between the first classification results of comparison result and source domain data and the second classification results of source domain data between classification results Comparison result, to neural network, common characteristic trapping layer and the first grader carry out epicycle training specific steps, packet It includes：

Step 6 one：It will be between the first classification results of target numeric field data and the second classification results of target numeric field data Difference is determined as the loss of the first probability.

Step 6 two：By the difference between the first classification results of source domain data and the second classification results of source domain data It is determined as the loss of the second probability；And

Step 6 three：According to the first probability loss and the second probability loss, to neural network, common characteristic trapping layer with And first grader carry out epicycle training.

When specific implementation, classification is carried out to data due to the use of grader and handles obtained classification results, it is real It is the probability that the data belong to some classification on border.First classification results of source domain data, the second classification results of source domain data The result classified to source domain data using the first grader and the second grader can be embodied in the form of probability.Target First classification results of numeric field data and the second classification results of target numeric field data can be embodied in the form of probability uses first point The result that class device and the second grader classify to target numeric field data.

When the difference between the first classification results of source domain data and the second classification results of source domain data is bigger, Namely first probability loss it is bigger, illustrate in Liang Ge branches, common characteristic trapping layer, the difference of the parameter of difference characteristic trapping layer Bigger, the difference of the parameter of the first grader and the second grader is also bigger, it was demonstrated that its source domain data extracted is shared Difference between feature vector and source domain data difference feature vector is also bigger so that in the process to disaggregated model training In, the difference characteristic between the source domain feature and target domain characterization of introducing is also more.

Meanwhile and cannot make difference between the two there is no limit increase, to be limited.Therefore, for The difference between the first classification results of target numeric field data and the second classification results of target numeric field data is used to be damaged as the second probability It loses, in a certain range by the second probability loss limitation.The loss of second probability is smaller, illustrates in Liang Ge branches, common characteristic is caught Catch layer, difference characteristic trapping layer parameter difference it is smaller, it was demonstrated that its aiming field common characteristic vector sum mesh extracted Mark domain difference characteristic vector between difference it is smaller so that during to having divided model training, the source domain feature of introducing and Difference characteristic between target domain characterization is also fewer.

Above-mentioned two branch synchronizes training so that is lost to neural network, common characteristic trapping layer, difference based on the first probability The training of different Feature capturing layer, the first grader and the second grader, and based on the loss of the second probability to neural network, shared spy It is mutually constrained between sign trapping layer, difference characteristic trapping layer, the first grader and the training of the second grader so that disaggregated model It is limited in certain range by the difference characteristic effect between source domain data and target numeric field data, reaches to altogether There is the balance between feature and otherness characteristic use.

The embodiment of the present application also provide it is a kind of according to the first probability loss and the second probability lose, to neural network, altogether There are Feature capturing layer and the first grader to carry out the specific method of epicycle training, including：

It executes following first and compares operation, until the loss of the first probability, which is less than preset first probability, loses threshold value；

First, which compares operation, includes：

The loss of first probability is compared with preset first probability loss threshold value；

If the loss of the first probability loses threshold value not less than preset first probability, neural network, common characteristic are captured The parameter of layer, difference characteristic trapping layer, the first grader and the second grader is adjusted；

It executes following second and compares operation, until the loss of the second probability, which is more than preset second probability, loses threshold value；

Second, which compares operation, includes：

The loss of second probability is compared with preset second probability loss threshold value；

If the loss of the second probability loses threshold value no more than preset second probability, neural network, common characteristic are captured The parameter of layer, difference characteristic trapping layer, the first grader and the second grader is adjusted.

Through the above steps, may be implemented to neural network, common characteristic trapping layer, difference characteristic trapping layer, first point It assembles, the training of the epicycle of the second grader.

In addition, the ratio between the second classification results of the first classification results and target numeric field data based on target numeric field data Pair as a result, and the comparison result between the first classification results and the second classification results of source domain data of source domain data, to god The same of epicycle training is carried out through network, common characteristic trapping layer, difference characteristic trapping layer, the first grader and the second grader When, also the first classification results of source domain data can be used to carry out neural network, common characteristic trapping layer and the first grader Training, and using source domain data the second classification results to neural network, difference characteristic trapping layer and the second grader into Row training.

Specifically, in the first classification results using source domain data to neural network, common characteristic trapping layer and first It is the target in the source domain common characteristic vector sum target numeric field data for the source domain data that will be captured when grader is trained Domain common characteristic vector executes after inputting the first grader.Specifically, following first Classification Loss can be executed and determine operation, To be trained to neural network, common characteristic trapping layer and the first grader using the first classification results of source domain data.

Shown in Figure 2, the first Classification Loss determines that operation includes：

S201：According to the first classification results of source domain data and the label of source domain data, the first Classification Loss is calculated；

S202：First Classification Loss is compared with preset first Classification Loss threshold value；

S203：Whether the first Classification Loss of detection source numeric field data is not more than preset first Classification Loss threshold value；If It is no, then jump to S204；If it is, jumping to S206.

S204：Adjust the parameter of neural network and the parameter of the second grader.

S205：Based on the parameter after adjustment, again using neural network, common characteristic trapping layer and the first grader Obtain the first classification results of source domain data；Jump to 201.

S206：It completes to train the epicycle of the parameter of neural network, common characteristic trapping layer and the first grader.

Until the first Classification Loss of source domain data is not more than preset first Classification Loss threshold value；

Herein, neural network, common characteristic extract layer and the first grader are characterized in training with the first Classification Loss In the process by the degree of aiming field data influence.So that the first Classification Loss is not more than preset first Classification Loss threshold value, It is to be influenced by the feature in target numeric field data in neural network, but this influence is limited in certain range, no It can influence the accuracy that neural network classifies to source domain data.

Usually, the first Classification Loss can be that the first grader divides the source domain common characteristic vector of source domain data The accuracy of class can be calculated the classification results of source domain data and the label of source domain data based on the first grader.

Similarly, in the second classification results using source domain data to neural network, common characteristic trapping layer and first It is the target in the source domain difference characteristic vector sum target numeric field data for the source domain data that will be captured when grader is trained Domain difference characteristic vector executes after inputting the second grader.

Specifically, operation can be determined by executing following first Classification Loss, until the first classification damage of source domain data It loses and is not more than preset first Classification Loss threshold value, to use the first classification results of source domain data to neural network, shared spy Sign trapping layer and the first grader are trained.Specifically, following second Classification Loss can be executed and determine operation, to use First classification results of source domain data are trained neural network, common characteristic trapping layer and the first grader.

Shown in Figure 3, the second Classification Loss determines that operation includes：

S301：According to the second classification results of source domain data and the label of source domain data, the second Classification Loss is calculated；

S302：Second Classification Loss is compared with preset second Classification Loss threshold value；

S303：Whether the second Classification Loss of detection source numeric field data is not more than preset second Classification Loss threshold value；If It is no, then jump to S304；If it is, jumping to S306.

S304：Adjust the parameter of neural network and the parameter of the second grader.

S305：Based on the parameter after adjustment, again using neural network, difference characteristic trapping layer and the second grader Obtain the second classification results of source domain data；Jump to 301.

S306：It completes to train the epicycle of the parameter of neural network, difference characteristic trapping layer and the second grader.

Until the second Classification Loss of source domain data is not more than preset second Classification Loss threshold value.

The determination process of second Classification Loss is similar with the determination process of the first Classification Loss, and details are not described herein.

S106：By carrying out more wheel training to neural network, common characteristic trapping layer and the first grader, classification mould is obtained Type.

When specific implementation, more wheel training refer to being inputted respectively for neural network in the multiple training of each round Source domain data and target numeric field data mesh be constant；And in the training of different wheels, it is neural network and target nerve network The source domain data and target numeric field data inputted are to be different.

In addition to the first round, in other wheel training, used initial neural network, common characteristic extract layer, difference are special It is obtained neural network, common characteristic when last round of training is completed to levy extract layer, the first grader and the second grader Extract layer, difference characteristic extract layer, the first grader and two graders, and finally by the neural networks by more wheel training, Common characteristic trapping layer and the first grader disaggregated model as the resulting.Difference characteristic trapping layer and the second grader It is only used for the training of subsidiary classification model.

Shown in Figure 4, the embodiment of the present application also provides a kind of sorting technique, and this method includes：

S401：Obtain data to be sorted；

S402：Data to be sorted are input to point obtained by disaggregated model training method provided by the embodiments of the present application In class model, the classification results of data to be sorted are obtained；

Wherein, disaggregated model includes：Neural network, common characteristic trapping layer and the first grader.

Herein, data to be sorted are target numeric field data, or have the data of same characteristic features with aiming field.

Using the obtained disaggregated model of disaggregated model training method provided by the embodiments of the present application, treat grouped data into Row classification, can obtain more accurate classification results.

Conceived based on same application, classification mould corresponding with disaggregated model training method is additionally provided in the embodiment of the present application Type training device, since the principle that the device in the embodiment of the present application solves the problems, such as is instructed with the above-mentioned disaggregated model of the embodiment of the present application It is similar to practice method, therefore the implementation of device may refer to the implementation of method, overlaps will not be repeated.

Shown in Figure 5, disaggregated model training device provided by the embodiments of the present application includes：

Acquisition module 10 is used to obtain the source domain data for carrying label and not the target numeric field data of tape label；

First processing module 20 is source domain number for source domain data and target numeric field data to be inputted same neural network Target domain characterization vector is extracted according to extraction source characteristic of field vector, and for target numeric field data；

Second processing module 30, for source domain feature vector and target domain characterization vector to be inputted common characteristic trapping layer, And the aiming field common characteristic vector of the source domain common characteristic vector sum target numeric field data of the source domain data of capture is inputted respectively First grader obtains the first classification results of source domain data and the first classification results of target numeric field data；And

Third processing module 40, for source domain feature vector and target domain characterization vector to be inputted difference characteristic trapping layer, And the aiming field difference characteristic vector of the source domain difference characteristic vector sum target numeric field data of the source domain data of capture is inputted respectively Second grader obtains the second classification results of source domain data and the second classification results of target numeric field data；

Training module 50 is used for the second classification results of the first classification results and target numeric field data based on target numeric field data Between comparison result and source domain data the first classification results and the second classification results of source domain data between comparison knot Fruit carries out epicycle training to neural network, common characteristic trapping layer and the first grader；By to neural network, shared spy It levies trapping layer and the first grader carries out more wheel training, obtain disaggregated model.

Optionally, training module 50, be specifically used for using following step to neural network, common characteristic trapping layer and First grader carries out epicycle training：By the first classification results of target numeric field data and the second classification results of target numeric field data Between difference be determined as the first probability loss；

Difference between first classification results of source domain data and the second classification results of source domain data is determined as Two probability lose；And

According to the loss of the first probability and the loss of the second probability, to neural network, common characteristic trapping layer and first point Class device carries out epicycle training.

Optionally, training module 50 are specifically used for being lost according to the first probability using following step and the second probability damage It loses, epicycle training is carried out to neural network, common characteristic trapping layer and the first grader：

First, which compares operation, includes：

Second, which compares operation, includes：

Optionally, Second processing module 30 share spy specifically for the source domain of the source domain data captured in the following way Levy the aiming field common characteristic vector of vector sum target numeric field data：

Source domain feature vector and target domain characterization vector are input to common characteristic trapping layer, captured using common characteristic Layer extraction source domain common characteristic vector sum aiming field common characteristic vector；

Source domain common characteristic vector sum aiming field common characteristic vector is subjected to gradient reverse process；

The source domain common characteristic vector sum aiming field common characteristic vector for carrying out gradient reverse process is input to the first domain Grader；

The source domain that source domain common characteristic vector sum aiming field common characteristic vector is characterized respectively according to the first domain grader The domain classification results of data and target numeric field data, are adjusted the parameter of neural network and common characteristic trapping layer.

Optionally, Second processing module 30 are specifically used for capturing neural network and common characteristic using following step The parameter of layer is adjusted：

It executes following first domain Classification Loss and determines operation：

It is not less than default differential threshold for the difference between the domain Classification Loss of nearest preset times, generates the first feedback Information, and parameter adjustment is carried out to neural network, common characteristic trapping layer based on the first feedback information；

The use of neural network and common characteristic trapping layer is that source domain data extract new source domain based on the parameter after adjustment Common characteristic vector, and new aiming field common characteristic vector is extracted for target numeric field data, and it is true to execute the first domain Classification Loss Fixed operation, until difference is more than default differential threshold, completion catches neural network and common characteristic based on the first domain grader Catch the epicycle training of layer.

Optionally, third processing module 40 is specifically used for capturing the source domain difference characteristic of source domain data using following step The aiming field difference characteristic vector of vector sum target numeric field data：Source domain feature vector and target domain characterization vector are inputted into difference Feature capturing layer is extracted using difference characteristic trapping layer and obtains source domain difference characteristic vector sum aiming field difference characteristic vector；

Source domain difference characteristic vector sum aiming field difference characteristic vector is input to the second domain grader；

The source domain that source domain difference characteristic vector sum aiming field difference characteristic vector is characterized respectively according to the second domain grader The domain classification results of data and target numeric field data, are adjusted the parameter of neural network and difference characteristic trapping layer.

Optionally, third processing module 40 is specifically used for capturing neural network and difference characteristic using following step The parameter of layer is adjusted：

It executes following second domain Classification Loss and determines operation：

For the situation of domain classification results mistake, the second feedback information is generated, and based on the second feedback information to nerve net Network and difference characteristic trapping layer carry out parameter adjustment；

Based on the parameter after adjustment, it is poor to be that source domain data extract new source domain using neural network and difference characteristic trapping layer Different feature vector, and new aiming field difference characteristic vector is extracted for target numeric field data, and execute the second domain Classification Loss and determine Operation.

Optionally, training module 50 are additionally operable to the source domain common characteristic vector sum aiming field in the source domain data that will be captured The aiming field common characteristic vector of data inputs after the first grader, executes following first Classification Loss and determines operation, until First Classification Loss of source domain data is not more than preset first Classification Loss threshold value；

First Classification Loss determines that operation includes：

According to the first classification results of source domain data and the label of source domain data, the first Classification Loss is calculated；By first Classification Loss is compared with preset first Classification Loss threshold value；If the first Classification Loss is more than preset first classification damage Threshold value is lost, then adjusts the parameter of neural network, the parameter of common characteristic trapping layer and the first grader；Based on the ginseng after adjustment Number, the first classification results of source domain data are regained using neural network, common characteristic trapping layer and the first grader, And it executes the first Classification Loss and determines operation.

Optionally, training module 50 are additionally operable to the source domain difference characteristic vector sum aiming field in the source domain data that will be captured The aiming field difference characteristic vector of data inputs after the second grader, executes following second Classification Loss and determines operation, until Second Classification Loss of source domain data is not more than preset second Classification Loss threshold value；

Second Classification Loss determines that operation includes：

According to the second classification results of source domain data and the label of source domain data, the second Classification Loss is calculated；By second Classification Loss is compared with preset second Classification Loss threshold value；If the second Classification Loss is more than preset second classification damage Threshold value is lost, then adjusts the parameter of neural network, the parameter of difference characteristic trapping layer and the second grader；Based on the ginseng after adjustment Number, the second classification results of source domain data are regained using neural network, difference characteristic trapping layer and the second grader, And it executes the second Classification Loss and determines operation.

Conceived based on same application, classification mold device corresponding with sorting technique is additionally provided in the embodiment of the present application, by The principle that device in the embodiment of the present application solves the problems, such as is similar to the above-mentioned sorting technique of the embodiment of the present application, therefore device Implementation may refer to the implementation of method, and overlaps will not be repeated.

Shown in Figure 6, sorter provided by the embodiments of the present application includes：

Data acquisition module 60 to be sorted, for obtaining data to be sorted；

Sort module 70, for data to be sorted to be input to the disaggregated model training provided by the embodiment of the present application In the disaggregated model that method obtains, the classification results of data to be sorted are obtained；

Corresponding to the disaggregated model training method in Fig. 1, the embodiment of the present application also provides a kind of computer equipments, such as scheme Shown in 7, which includes memory 1000, processor 2000 and is stored on the memory 1000 and can be in the processor 2000 The computer program of upper operation, wherein above-mentioned processor 2000 realizes above-mentioned disaggregated model instruction when executing above computer program The step of practicing method.

Specifically, above-mentioned memory 1000 and processor 2000 can be general memory and processor, not do here It is specific to limit, when the computer program of 2000 run memory 1000 of processor storage, it is able to carry out above-mentioned disaggregated model instruction Practice method, to solve using with label source domain data and the target numeric field data of tape label does not instruct disaggregated model Target numeric field data is carried out just with disaggregated model caused by the common characteristic of source domain data and target numeric field data when white silk There is error in classification, and then reach while utilizing the same characteristic features and difference characteristic of source domain data and target numeric field data Get up, the obtained disaggregated model of training can be to effect that target numeric field data is more accurately classified.

Corresponding to the disaggregated model training method in Fig. 1, the embodiment of the present application also provides a kind of computer-readable storages Medium is stored with computer program on the computer readable storage medium, is executed when which is run by processor The step of stating disaggregated model training method.

Specifically, which can be general storage medium, such as mobile disk, hard disk, on the storage medium Computer program when being run, above-mentioned disaggregated model training method is able to carry out, to solve using the source with label Just with source domain data and aiming field when the target numeric field data of numeric field data and not tape label is trained disaggregated model Disaggregated model caused by the common characteristic of data carries out classification to target numeric field data and there are problems that error, and then reaches while inciting somebody to action The same characteristic features and difference characteristic of source domain data and target numeric field data use, and the disaggregated model that training obtains can be to target The effect that numeric field data is more accurately classified.

The computer program product of disaggregated model training method and sorting technique that the embodiment of the present application is provided, including The computer readable storage medium of program code is stored, the instruction that program code includes can be used for executing previous methods embodiment In method, specific implementation can be found in embodiment of the method, details are not described herein.

It is apparent to those skilled in the art that for convenience and simplicity of description, the system of foregoing description It with the specific work process of device, can refer to corresponding processes in the foregoing method embodiment, details are not described herein.

If function is realized in the form of SFU software functional unit and when sold or used as an independent product, can store In a computer read/write memory medium.Based on this understanding, the technical solution of the application is substantially in other words to existing There is the part for the part or the technical solution that technology contributes that can be expressed in the form of software products, the computer Software product is stored in a storage medium, including some instructions are used so that a computer equipment (can be personal meter Calculation machine, server or network equipment etc.) execute each embodiment method of the application all or part of step.And it is above-mentioned Storage medium includes：USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory The various media that can store program code such as (RAM, Random Access Memory), magnetic disc or CD.

More than, the only specific implementation mode of the application, but the protection domain of the application is not limited thereto, and it is any to be familiar with Those skilled in the art can easily think of the change or the replacement in the technical scope that the application discloses, and should all cover Within the protection domain of the application.Therefore, the protection domain of the application shall be subject to the protection scope of the claim.

Claims

1. a kind of disaggregated model training method, which is characterized in that this method includes：

The source domain data and the target numeric field data are inputted into same neural network, it is special to extract source domain for the source domain data Sign vector, and extract target domain characterization vector for the target numeric field data；

The source domain feature vector and the target domain characterization vector are inputted into common characteristic trapping layer, and respectively by the source of capture The aiming field common characteristic vector of the source domain common characteristic vector sum target numeric field data of numeric field data inputs the first grader, obtains source First classification results of numeric field data and the first classification results of target numeric field data；And

The source domain feature vector and the target domain characterization vector are inputted into difference characteristic trapping layer, and respectively by the source of capture The aiming field difference characteristic vector of the source domain difference characteristic vector sum target numeric field data of numeric field data inputs the second grader, obtains source Second classification results of numeric field data and the second classification results of target numeric field data；

Comparison between the first classification results and the second classification results of the target numeric field data based on the target numeric field data As a result, and the comparison knot between the first classification results and the second classification results of the source domain data of the source domain data Fruit carries out epicycle training to the neural network, the common characteristic trapping layer and first grader；

2. according to the method described in claim 1, it is characterized in that, the first classification results based on the target numeric field data and institute State the first classification results of the comparison result and the source domain data between the second classification results of target numeric field data with it is described Comparison result between second classification results of source domain data, to the neural network, the common characteristic trapping layer, Yi Jisuo It states the first grader and carries out epicycle training, specifically include：

By the difference between the first classification results of the target numeric field data and the second classification results of the target numeric field data It is determined as the loss of the first probability；

Difference between first classification results of the source domain data and the second classification results of the source domain data is determined It is lost for the second probability；And

According to first probability loss and second probability loss, the neural network, the common characteristic are captured Layer and first grader carry out epicycle training.

3. according to the method described in claim 2, it is characterized in that, described according to first probability loss and described second Probability loses, and carries out epicycle training to the neural network, the common characteristic trapping layer and first grader, specifically Including：

Described first, which compares operation, includes：

First probability loss is compared with preset first probability loss threshold value；

If first probability loss loses threshold value not less than preset first probability, to the neural network, described shared Feature capturing layer, the difference characteristic trapping layer, the parameter of first grader and second grader are adjusted；

Described second, which compares operation, includes：

Second probability loss is compared with preset second probability loss threshold value；

If second probability loss loses threshold value no more than preset second probability, to the neural network, described shared Feature capturing layer, the difference characteristic trapping layer, the parameter of first grader and second grader are adjusted.

4. according to claim 1-3 any one of them methods, which is characterized in that by the source domain feature vector and the target Characteristic of field vector inputs common characteristic trapping layer, the source domain common characteristic vector sum mesh of the source domain data captured in the following way Mark the aiming field common characteristic vector of numeric field data：

The source domain feature vector and the target domain characterization vector are input to the common characteristic trapping layer, using described Common characteristic trapping layer extracts aiming field common characteristic vector described in the source domain common characteristic vector sum；

Aiming field common characteristic vector described in the source domain common characteristic vector sum is subjected to gradient reverse process；

Aiming field common characteristic vector described in the source domain common characteristic vector sum for carrying out gradient reverse process is input to the One domain grader；

Table is distinguished to aiming field common characteristic vector described in the source domain common characteristic vector sum according to first domain grader The domain classification results of the source domain data and the target numeric field data of sign, catch the neural network and the common characteristic The parameter for catching layer is adjusted.

5. according to the method described in claim 4, it is characterized in that, described total to the source domain according to first domain grader There is the domain for the source domain data and the target numeric field data that feature vector and the aiming field common characteristic vector characterize respectively Classification results are adjusted the parameter of the neural network and the common characteristic trapping layer, specifically include：

Determine source domain data and target that aiming field common characteristic vector described in current source domain common characteristic vector sum characterizes respectively The domain Classification Loss of this domain classification of numeric field data；

It is not less than default differential threshold for the difference between the domain Classification Loss of nearest preset times, generates the first feedback letter Breath, and parameter adjustment is carried out to the neural network, the common characteristic trapping layer based on first feedback information；

Based on the parameter after adjustment, extracted for the source domain data using the neural network and the common characteristic trapping layer New source domain common characteristic vector, and new aiming field common characteristic vector is extracted for the target numeric field data, and described in execution First domain Classification Loss determines operation, until difference is more than default differential threshold, completes based on first domain grader to institute State the epicycle training of neural network and the common characteristic trapping layer.

6. according to claim 1-3 any one of them methods, which is characterized in that by the source domain feature vector and the target Characteristic of field vector inputs difference characteristic trapping layer, captures the source domain difference characteristic vector sum target of source domain data in the following way The aiming field difference characteristic vector of numeric field data：

The source domain feature vector and the target domain characterization vector are inputted into the difference characteristic trapping layer, use the difference Different Feature capturing layer extraction obtains aiming field difference characteristic vector described in the source domain difference characteristic vector sum；

Aiming field difference characteristic vector described in the source domain difference characteristic vector sum is input to the second domain grader；

Table is distinguished to aiming field difference characteristic vector described in the source domain difference characteristic vector sum according to second domain grader The domain classification results of the source domain data and the target numeric field data of sign, catch the neural network and the difference characteristic The parameter for catching layer is adjusted.

7. according to the method described in claim 6, it is characterized in that, described poor to the source domain according to second domain grader The domain for the source domain data and the target numeric field data that different feature vector and the aiming field difference characteristic vector characterize respectively Classification results are adjusted the parameter of the neural network and the difference characteristic trapping layer, specifically include：

Determine the source domain data and aiming field number that current source domain difference characteristic vector sum aiming field difference characteristic vector characterizes respectively According to this domain classification domain Classification Loss；

For the situation of domain classification results mistake, the second feedback information is generated, and based on second feedback information to the god Parameter adjustment is carried out through network and the difference characteristic trapping layer；

The use of neural network and the difference characteristic trapping layer is that the source domain data extract new source based on the parameter after adjustment Domain difference characteristic vector, and new aiming field difference characteristic vector is extracted for the target numeric field data, and execute second domain Classification Loss determines operation.

8. according to claim 1-3 any one of them methods, which is characterized in that shared by the source domain of the source domain data of capture The aiming field common characteristic vector of feature vector and target numeric field data inputs after the first grader, and this method further includes：

It executes following first Classification Loss and determines operation, until the first Classification Loss of source domain data is not more than preset first point Class loses threshold value；

First Classification Loss determines that operation includes：

According to the first classification results of the source domain data and the label of the source domain data, the first Classification Loss is calculated；

First Classification Loss is compared with preset first Classification Loss threshold value；

If first Classification Loss be more than preset first Classification Loss threshold value, adjust the neural network parameter, The parameter of the common characteristic trapping layer and first grader；

Based on the parameter after adjustment, source is regained using neural network, the common characteristic trapping layer and the first grader First classification results of numeric field data, and execute first Classification Loss and determine operation.

9. according to claim 1-3 any one of them methods, which is characterized in that in the source domain difference for the source domain data that will be captured The aiming field difference characteristic vector of feature vector and target numeric field data inputs after the second grader, and this method further includes：

It executes following second Classification Loss and determines operation, until the second Classification Loss of source domain data is not more than preset second point Class loses threshold value；

Second Classification Loss determines that operation includes：

According to the second classification results of the source domain data and the label of the source domain data, the second Classification Loss is calculated；

Second Classification Loss is compared with preset second Classification Loss threshold value；

If second Classification Loss be more than preset second Classification Loss threshold value, adjust the neural network parameter, The parameter of the difference characteristic trapping layer and second grader；

Based on the parameter after adjustment, source is regained using neural network, the difference characteristic trapping layer and the second grader Second classification results of numeric field data, and execute second Classification Loss and determine operation.

10. a kind of sorting technique, which is characterized in that this method includes：

Obtain data to be sorted；

The data to be sorted are input to the classification obtained by the disaggregated model training method of claim 1-9 any one In model, the classification results of the data to be sorted are obtained；

Wherein, the disaggregated model includes：The neural network, the common characteristic trapping layer and first grader.