CN110210625A - Modeling method, device, computer equipment and storage medium based on transfer learning - Google Patents
Modeling method, device, computer equipment and storage medium based on transfer learning Download PDFInfo
- Publication number
- CN110210625A CN110210625A CN201910418820.5A CN201910418820A CN110210625A CN 110210625 A CN110210625 A CN 110210625A CN 201910418820 A CN201910418820 A CN 201910418820A CN 110210625 A CN110210625 A CN 110210625A
- Authority
- CN
- China
- Prior art keywords
- feature
- sample
- dimensionality reduction
- information
- reduction feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Image Analysis (AREA)
Abstract
This application involves artificial intelligence, a kind of modeling method based on transfer learning, device, computer equipment and storage medium are provided, the described method includes: exemplar to be learned and target labels sample are carried out kernel principal component analysis, obtain first dimensionality reduction feature corresponding with exemplar to be learned, the second dimensionality reduction feature corresponding with target labels sample；The generic features that first dimensionality reduction feature and the input of the second dimensionality reduction feature have been trained are obtained in model, general column feature is obtained；First dimensionality reduction feature is inputted basic model corresponding with target labels sample to test, obtains weight information corresponding with the first dimensionality reduction feature, weight information is higher than the first dimensionality reduction feature of default weight threshold as general row feature；General column feature and general row feature are inputted in basic model corresponding with target labels sample and carry out model training, obtains object module, effective model can be constructed in the case where a small amount of tape label sample based on transfer learning realization.
Description
Technical field
This application involves field of computer technology, more particularly to it is a kind of by the modeling method of transfer learning, device, based on
Calculate machine equipment and storage medium.
Background technique
With the high speed development of field of computer technology, the data exponentially grade obtained in actual life increases.It is how right
The data of magnanimity are quickly and effectively handled, and then extract valuable information required for user, are that researchers are general
All over concern.With constantly bringing forth new ideas for machine learning field, researchers propose transfer learning, transfer learning refer to by
The knowledge migration acquired in one scene is into another scene, so that model can also be made very well in a large amount of completely new scenes
Prediction.
The sample for largely having business to show traditionally is required for the foundation of model, but the business of certain new developments can
It can be difficult to construct effective model according to conventional method without enough samples；If a small amount of current service data is used only
Modeling, model are easy over-fitting and unstable；If using the model of the sample building by other business, in view of different business visitor
There may be bigger difference, modelling effect may be remarkably decreased group, can not be constructed in the case where only a small amount of tape label sample
Effective model.
Summary of the invention
Based on this, it is necessary in view of the above technical problems, provide it is a kind of by the modeling method of transfer learning, device, based on
Machine equipment and storage medium are calculated, effective mould can be constructed in the case where a small amount of tape label sample based on transfer learning realization
Type.
A kind of modeling method based on transfer learning, which comprises
Obtain exemplar and target labels sample to be learned；
The exemplar to be learned and the target labels sample are subjected to kernel principal component analysis, obtained with described wait learn
Practise the corresponding first dimensionality reduction feature of exemplar, the second dimensionality reduction feature corresponding with the target labels sample；
The generic features that the first dimensionality reduction feature and the second dimensionality reduction feature input have been trained are obtained in model, are obtained
To general column feature；
The first dimensionality reduction feature is inputted corresponding with target labels sample basic model to test, obtain and
The corresponding weight information of the first dimensionality reduction feature makees the first dimensionality reduction feature that the weight information is higher than default weight threshold
For general row feature；
The general column feature and the general row feature are inputted into basic model corresponding with the target labels sample
Middle carry out model training, obtains object module.
In one of the embodiments, the method also includes:
Sample to be evaluated is obtained, the sample to be evaluated is inputted in the object module, output and the test sample to be evaluated
This corresponding sample label information；
The sample label information is shown, label correct information corresponding with the sample label information is obtained；
The weight in the object module is adjusted according to the label correct information, according to the power after each adjust
Value is updated the object module, obtains updated object module.
In one of the embodiments, the method also includes:
The first dimensionality reduction feature and the second dimensionality reduction feature are subjected to aspect ratio pair, obtain characteristic similarity；
The characteristic similarity is higher than the first dimensionality reduction feature when presetting similar threshold value as general column feature.
In one of the embodiments, the method also includes:
The first dimensionality reduction feature is inputted corresponding with target labels sample basic model to test, output and
The corresponding sample label information of the exemplar to be learned；
The sample label information is shown, the positive false information of label corresponding with the sample label information is obtained；
It is corrected errors the first dimensionality reduction feature described in information evaluation, is obtained corresponding with the first dimensionality reduction feature according to the label
Signature contributions degree information；
The weight information of the first dimensionality reduction feature is determined according to the signature contributions degree information.
In one of the embodiments, the method also includes:
The general column feature and the general row feature are divided into the training characteristics collection of predetermined quantity part；
Successively the training characteristics collection is inputted in the input variable of the basic model and be trained, until all training
Feature set training finishes, the object module trained.
A kind of model building device based on transfer learning, described device include:
Sample acquisition module, for obtaining exemplar and target labels sample to be learned；
Feature Dimension Reduction module, for the exemplar to be learned and the target labels sample to be carried out kernel principal component point
Analysis obtains first dimensionality reduction feature corresponding with the exemplar to be learned, the second drop corresponding with the target labels sample
Dimensional feature；
Column feature obtains module, logical for having trained the first dimensionality reduction feature and the second dimensionality reduction feature input
It is obtained in model with feature, obtains general column feature；
Row feature obtains module, for the first dimensionality reduction feature to be inputted basis corresponding with the target labels sample
Model is tested, and weight information corresponding with the first dimensionality reduction feature is obtained, and the weight information is higher than default weight
First dimensionality reduction feature of threshold value is as general row feature；
Model training module, for inputting and the target labels sample the general column feature and the general row feature
Model training is carried out in this corresponding basic model, obtains object module.
Described device in one of the embodiments, further include:
The sample to be evaluated is inputted the object module for obtaining sample to be evaluated by label information output module
In, output sample label information corresponding with the sample to be evaluated；
Correct information obtains module, for showing the sample label information, obtains and believes with the sample label
Cease corresponding label correct information；
Model modification module, for the weight in the object module to be adjusted according to the label correct information,
The object module is updated according to the weight after each adjust, obtains updated object module.
Described device in one of the embodiments, further include:
Feature comparison module is obtained for the first dimensionality reduction feature and the second dimensionality reduction feature to be carried out aspect ratio pair
To characteristic similarity；
Similarity judgment module is made for the first dimensionality reduction feature when the characteristic similarity to be higher than to default similar threshold value
For general column feature.
A kind of computer equipment can be run on a memory and on a processor including memory, processor and storage
The step of computer program, the processor realizes the above method when executing described program.
A kind of computer readable storage medium, is stored thereon with computer program, realization when which is executed by processor
The step of above method.
The above-mentioned modeling method based on transfer learning, device, computer equipment and storage medium, server is by treating
It practises exemplar and target labels sample carries out kernel principal component analysis, obtain the first dimensionality reduction feature and the second dimensionality reduction feature, it can
Originally very high-dimensional data are indicated with seldom some representative dimensions, and the first dimensionality reduction feature and the second dimensionality reduction is special
The generic features that sign input has been trained obtain in model, obtain general column feature, and further, it is special that server obtains general row
Sign, and object module is established according to general row feature and general column feature.Server arrives exemplar transfer learning to be learned
Target labels sample, building can identify the object module of target labels sample, can be realized based on transfer learning in a small amount of band
Effective model is constructed in the case where exemplar.
Detailed description of the invention
Fig. 1 is the applied environment figure of the modeling method based on transfer learning in one embodiment；
Fig. 2 is the method flow diagram of the modeling method based on transfer learning in one embodiment；
Fig. 3 is the method flow for carrying out object module update in one embodiment in the modeling method based on transfer learning
Figure；
Fig. 4 is in one embodiment based on the method flow diagram for determining weight information in the modeling method of transfer learning；
Fig. 5 is the structural schematic diagram of the modeling method device based on transfer learning in one embodiment；
Fig. 6 is the schematic diagram of internal structure of computer equipment in one embodiment.
Specific embodiment
It is with reference to the accompanying drawings and embodiments, right in order to which the objects, technical solutions and advantages of the application are more clearly understood
The application is further elaborated.It should be appreciated that specific embodiment described herein is only used to explain the application, not
For limiting the application.
Modeling method provided in the embodiment of the present invention based on transfer learning can be applied to application as shown in Figure 1
In environment, server 120 obtains exemplar and target labels sample to be learned from terminal 110, and server 120 is by mark to be learned
Signed-off sample sheet and target labels sample carry out kernel principal component analysis, obtain first dimensionality reduction feature corresponding with exemplar to be learned,
The second dimensionality reduction feature corresponding with target labels sample, server 120 input the first dimensionality reduction feature and the second dimensionality reduction feature
Trained generic features obtain in model, obtain general column feature, and server 120 inputs the first dimensionality reduction feature and target labels
The corresponding basic model of sample is tested, and weight information corresponding with the first dimensionality reduction feature is obtained, and server 120 believes weight
Breath is higher than the first dimensionality reduction feature of default weight threshold as general row feature, and server 120 is by general column feature and general row
Feature inputs in basic model corresponding with target labels sample and carries out model training, obtains object module.
Following embodiments are illustrated so that the modeling method based on transfer learning is applied to the server of Fig. 1 as an example, but
It should be noted that in practical application this method and not only limit be applied to above-mentioned server.
As shown in Fig. 2, for the flow chart of the modeling method based on transfer learning in one embodiment, this method is specifically wrapped
Include following steps:
Step 202, exemplar and target labels sample to be learned is obtained.
Wherein, the exemplar of exemplar and target labels sample representation different business type to be learned, mark to be learned
Signed-off sample sheet has performance sample for business A's, and target labels sample has performance sample for minimal amount of business B's.It is understood that
It is that exemplar and target labels sample to be learned is all the sample with label information.
Specifically, transfer learning refers to the knowledge migration that will be acquired in a scene into another scene.It is learned in migration
In habit, existing knowledge is called source domain, and the new knowledge to be learnt is called aiming field, by the knowledge migration learnt to another
The study of the unknown knowledge of kind, i.e., move to aiming field from source domain.It is understood that source domain can be exemplar to be learned,
Aiming field can be target labels sample.
It illustrates, it is assumed that there is one can differentiate the model of cat and dog with pinpoint accuracy, if thinking one energy of training
The object module of the dog of enough respectively different kinds, the not instead of not from the beginning training data for needing to do are passed through with obtaining object module
General row feature and general column feature are extracted, using general row feature and the last several layers of neurons of general column feature training, is obtained
The object module of the kind of dog can be differentiated, here it is transfer learnings.
In one embodiment, source domain can be the exemplar to be learned that user carries out repaying ability when loan for vehicle,
Aiming field can be the target labels sample that user carries out repaying ability when petty load, and server is borrowed by transfer learning vehicle
The modeling method of money class business, the object module of petty load class business is established with this, enable object module to user into
Repaying ability when row petty load is assessed.
Server can obtain exemplar and target labels sample to be learned from other servers, can also obtain from terminal
Exemplar and target labels sample to be learned.Sample with label information refers to the label for having predefined in the sample
Information.For example, the label for example when exemplar to be learned is the picture of a doggie, in the exemplar to be learned
Information is " doggie ".
Step 204, exemplar to be learned and target labels sample are subjected to kernel principal component analysis, obtained and mark to be learned
The corresponding first dimensionality reduction feature of signed-off sample sheet, the second dimensionality reduction feature corresponding with target labels sample.
Wherein, server will carry out kernel principal component analysis, kernel principal component to exemplar to be learned and target labels sample
Analysis is transformed into the data of Nonlinear separability on the new lower-dimensional subspace that one is suitble to alignment to carry out linear classification.That is,
Exemplar to be learned and target labels sample are subjected to dimension-reduction treatment, kernel principal component analysis is very effective in machine learning
Dimensionality reduction can indicate originally very high-dimensional data with seldom some representative dimensions, for example, 1000 multidimensional with 100 tie up come
It indicates, without losing crucial data information.
Specifically, server uses kernel principal component analysis for the sample of source domain (i.e. business A) and aiming field (i.e. business B)
All samples are mapped to the subspace to common across a data field subspace by acquistion, obtain new character representation.
That is, server will obtain first dimensionality reduction feature corresponding with exemplar to be learned, and corresponding with target labels sample
Two dimensionality reduction features.
In one embodiment, it when exemplar to be learned and target labels sample need to carry out dimensionality reduction to K, can carry out
Following steps: 1) going average value (i.e. decentralization), i.e. each feature subtracts respective average value.2) covariance square is calculated
Battle array.3) eigen vector of covariance matrix is sought with Eigenvalues Decomposition method.4) characteristic value is sorted from large to small, is selected
Select maximum k.Then using its corresponding k feature vector as row vector composition characteristic vector matrix P.5) will
Data are transformed into the new space of k feature vector building, i.e. Y=PX.I.e. by obtaining fisrt feature average value and the second spy
Average value is levied, then carries out average value, obtains characteristic value corresponding with target signature and feature vector, server is further according to feature
Value is ranked up feature vector, obtains ranking results, while establishing according to the feature vector that ranking results are greater than preset threshold
Across data field subspace, then exemplar to be learned and target labels sample are mapped in across data field subspace, Neng Gouzhun
Really obtain first dimensionality reduction feature corresponding with exemplar to be learned, the second dimensionality reduction feature corresponding with target labels sample.
Step 206, the generic features that the first dimensionality reduction feature and the input of the second dimensionality reduction feature have been trained are obtained in model, is obtained
To general column feature.
Specifically, it is to solve source domain and aiming field distribution different problems in transfer learning that generic features, which obtain model,
Study obtains in above-mentioned new subspace, and marginal probability distribution is poor between reducing domain using the method for minimizing Largest Mean difference
It is different, while Largest Mean difference will be minimized and expand to conditional probability distribution between domain, joint edges matched probability distribution and item
Part probability distribution.Minimum Largest Mean difference, which refers to, to be projected and is summed to each sample, and the size of sum is utilized to state
The distributional difference of two data.It is understood that across data field subspace is identical by the way that source domain and aiming field to be mapped to
Space (or one of them is mapped in another space) simultaneously minimizes the distance of source domain and aiming field to complete knowledge
Migration.
It is understood that assuming there is stochastic variable X and Y, at this point, P(X=a, Y=b)For indicating that X=a and Y=b's is general
Rate, this kind of probability comprising multiple conditions and all conditions establishment simultaneously is joint probability, and the list of joint probability is known as joining
Close distribution.It is corresponding with joint probability, P(X=a)Or P(Y=b)This kind of probability only related with single stochastic variable is marginal probability,
The list of marginal probability is known as edge distribution.In the case where condition Y=b is set up, the probability of X=a is denoted as P(X=a | Y=b)Or
P(a|b).Distribution, that is, conditional probability distribution of conditional probability, i.e., known two relevant stochastic variable X and Y, stochastic variable Y is in item
Conditional probability distribution under part { X=x } refer to when known X value be some particular value x when, the probability distribution of Y.Server
It is obtained in model, can be accurately obtained by the generic features for having trained the first dimensionality reduction feature and the input of the second dimensionality reduction feature
General column feature.
Step 208, the first dimensionality reduction feature corresponding with target labels sample basic model is inputted to test, obtain and
Weight information is higher than the first dimensionality reduction feature of default weight threshold as general row by the corresponding weight information of the first dimensionality reduction feature
Feature.
Step 210, by general column feature and general row feature input in basic model corresponding with target labels sample into
Row model training, obtains object module.
Wherein, each example since exemplars to be learned certain in source domain are unrelated with aiming field sample, i.e., in source domain
Different to the contribution of aiming field model training, the example in source domain is high to target domain model relevance grade, and weight is just high, relevance grade
Low, weight is with regard to low.Server, to the contribution degree of aiming field model training, is further obtained according to each example obtained in source domain
General row feature is taken, using L2,1 norm selects the related example in source domain to carry out model training, obtains the mesh suitable for business B
Mark model, it is to be understood that L2,1 norm refer to the sparse selection feature of row, obtain general row feature for server.
Specifically, the general column feature and the input of general row feature that server will acquire are corresponding with target labels sample
Basic model in carry out model training, that is, complete the process of transfer learning, preferably object module can be obtained.
In the present embodiment, exemplar to be learned and target labels sample kernel principal component analysis are obtained first by server
Dimensionality reduction feature and the second dimensionality reduction feature can indicate original high-dimensional data with seldom some representativeness dimensions, without
Crucial data information is lost, then the generic features that the first dimensionality reduction feature and the input of the second dimensionality reduction feature have been trained are obtained into model
In, general column feature is obtained, the first dimensionality reduction feature is inputted into basic model corresponding with target labels sample and is tested, is obtained
Weight information is higher than the first dimensionality reduction feature of default weight threshold as general by weight information corresponding with the first dimensionality reduction feature
Row feature can realize source domain to mesh by general column feature and general row feature in the case where only a small amount of tape label sample
The transfer learning for marking domain, to complete the foundation of object module.
In one embodiment, as shown in figure 3, this method is further comprising the steps of:
Step 302, sample to be evaluated is obtained, sample to be evaluated is inputted in object module, output and sample pair to be evaluated
The sample label information answered.
Wherein, sample to be evaluated refers to that sample to be evaluated is inputted mesh by the sample verified to object module, server
It marks in model, sample label information corresponding with sample to be evaluated can be exported.It is understood that sample to be evaluated be without
There is the sample of label information.
Step 304, sample label information is shown, obtains label correct information corresponding with sample label information.
Step 306, the weight in object module is adjusted according to label correct information, according to the power after each adjust
Value is updated object module, obtains updated object module.
Wherein, the mode that sample label information is shown is included but is not limited to Real time displaying and is sent to pair by server
The terminal answered is shown, after server is shown sample label information, will acquire corresponding with sample label information
Label correct information.
For example, and user's loan for vehicle is held for example when sample to be evaluated is user's loan for vehicle ability to bear sample
When sample label information by ability is " middle grade ", which is shown, if server receiving terminal is returned
When the label correct information returned is " high-grade ", server will be adjusted the weight in object module according to label correct information
Section, and object module is updated according to the weight after each adjust, obtain updated object module.
In the present embodiment, the on-line study and real-time update to object module can be reached, according to end by the intervention of terminal
The label correct information that end returns further updates object module, improves the processing capacity to sample of object module.User is real
Modified result is also incorporated into training set, again training pattern, more new model in use, after user's correction result by border, is carried out
Next round prediction.Server inputs in object module by obtaining sample to be evaluated, then by sample to be evaluated, output with it is to be evaluated
The corresponding sample label information of sample, then shows sample label information, and obtain label correct information, according to label
The weight in object module is adjusted in correct information, is updated, is obtained to object module according to the weight after each adjust
To updated object module, the real-time update of object module can be realized.
In one embodiment, this method further include: the first dimensionality reduction feature and the second dimensionality reduction feature are subjected to aspect ratio pair,
Obtain characteristic similarity；Characteristic similarity is higher than the first dimensionality reduction feature when presetting similar threshold value as general column feature.
Specifically, and the corresponding first dimensionality reduction feature of source domain and the second dimensionality reduction feature corresponding with aiming field is subjected to feature
Similarity compares, and the first dimensionality reduction feature when characteristic similarity to be higher than to default similar threshold value is as general column feature, general
Column feature is used to be trained basic model in conjunction with general row feature, obtains object module.
In the present embodiment, the first dimensionality reduction feature and the second dimensionality reduction feature are carried out aspect ratio pair by server, obtain feature phase
Like degree；Characteristic similarity is higher than the first dimensionality reduction feature when presetting similar threshold value as general column feature, general column feature is used
In the transfer learning for carrying out model, object module is further obtained.
In one embodiment, as shown in figure 4, this method is further comprising the steps of:
Step 402, the first dimensionality reduction feature corresponding with target labels sample basic model is inputted to test, output and
The corresponding sample label information of exemplar to be learned.
Step 404, sample label information is shown, obtains the positive false information of label corresponding with sample label information.
Specifically, sample label information refers to that label information corresponding to exemplar to be learned, the positive false information of label are
Discriminant information of correcting errors to being made based on the label information in exemplar to be learned to the sample label information.Server is by
One dimensionality reduction feature inputs basic model corresponding with target labels sample and is tested, and exports corresponding with exemplar to be learned
Sample label information, and carry out Real time displaying by sample label information or be sent to corresponding terminal to show.
For example, the mark for example when exemplar to be learned is the picture of a doggie, in the exemplar to be learned
Signing information is " doggie ", is surveyed when the first dimensionality reduction feature is inputted basic model corresponding with target labels sample by server
Examination, when obtained sample label information is " kitten ", server will be believed according to based on the label in exemplar to be learned at this time
Breath differentiates correcting errors for sample label information, and label information of correcting errors includes but is not limited to correct and mistake.
Step 406, it is corrected errors information evaluation the first dimensionality reduction feature according to label, obtains feature corresponding with the first dimensionality reduction feature
Contribution degree information.
Step 408, the weight information of the first dimensionality reduction feature is determined according to signature contributions degree information.
Wherein, basic model is used to obtain contribution degree of each example to aiming field model training in source domain, server
Weight information is further determined that further according to the contribution degree got, when weight is high, it is meant that this feature relevance grade is high, works as weight
When low, it is meant that this feature relevance grade is low.By the high Feature Selection of relevance grade come out obtain general row feature, with for it is subsequent into
One step establishes object module.
Specifically, server according to label correct errors information evaluation the first dimensionality reduction feature for aiming field model training feature
Contribution degree obtains signature contributions degree information, and the weight information of the first dimensionality reduction feature is determined according to signature contributions degree information, and will
Weight information is higher than the first dimensionality reduction feature of default weight threshold as general row feature.
In the present embodiment, the first dimensionality reduction feature is inputted basic model corresponding with target labels sample and surveyed by server
Examination exports corresponding with exemplar to be learned sample label information, and sample label information is shown, can obtain and
The positive false information of the corresponding label of sample label information further judges the first dimensionality reduction feature to aiming field by the positive false information of label
The contribution degree of model training is corrected errors information evaluation the first dimensionality reduction feature according to label, is obtained corresponding with the first dimensionality reduction feature
Signature contributions degree information determines the weight information of the first dimensionality reduction feature according to signature contributions degree information, can further determine that logical
With row feature, server carries out transfer learning according to general row feature and general column feature, and then establishes and be suitable for aiming field
Object module.
In one embodiment, this method further include: general column feature and general row feature are divided into predetermined quantity part
Training characteristics collection；It will be successively trained in the input variable of training characteristics collection input basic model, until all training characteristics
Collection training finishes, the object module trained.
Wherein, the training characteristics collection of predetermined quantity part is for being trained basic model, the target mould trained
Type realizes the transfer learning from source domain to aiming field.
In the present embodiment, general column feature and general row feature are divided into the training characteristics collection of predetermined quantity part by server,
And will be successively trained in the input variable of training characteristics collection input basic model, until all training characteristics collection training are complete
Finish, the object module trained, can be realized based on transfer learning and be constructed effectively in the case where a small amount of tape label sample
Model.
As shown in figure 5, for the schematic diagram of the model building device based on transfer learning in an embodiment, which includes:
Sample acquisition module 502, for obtaining exemplar and target labels sample to be learned；
Feature Dimension Reduction module 504 is obtained for exemplar to be learned and target labels sample to be carried out kernel principal component analysis
To first dimensionality reduction feature corresponding with exemplar to be learned, the second dimensionality reduction feature corresponding with target labels sample；
Column feature obtains module 506, for the first dimensionality reduction feature and the second dimensionality reduction feature to be inputted the general spy trained
Sign obtains in model, obtains general column feature；
Row feature obtains module 508, for the first dimensionality reduction feature to be inputted basic model corresponding with target labels sample
It is tested, obtains weight information corresponding with the first dimensionality reduction feature, weight information is higher than to the first drop of default weight threshold
Dimensional feature is as general row feature；
Model training module 510, for general column feature and the input of general row feature is corresponding with target labels sample
Model training is carried out in basic model, obtains object module.
In one embodiment, model training module includes: label information output module, for obtaining sample to be evaluated,
Sample to be evaluated is inputted in object module, sample label information corresponding with sample to be evaluated is exported；Correct information obtains mould
Block obtains label correct information corresponding with sample label information for showing sample label information；Model modification mould
Block, for the weight in object module to be adjusted according to label correct information, according to the weight after each adjust to target
Model is updated, and obtains updated object module.
In one embodiment, it includes: feature comparison module that column feature, which obtains module, for by the first dimensionality reduction feature and the
Two dimensionality reduction features carry out aspect ratio pair, obtain characteristic similarity；Similarity judgment module, it is default for characteristic similarity to be higher than
The first dimensionality reduction feature when similar threshold value is as general column feature.
In one embodiment, it includes: by the input of the first dimensionality reduction feature and target labels sample pair that row feature, which obtains module,
The basic model answered is tested, and sample label information corresponding with exemplar to be learned is exported；By sample label information into
Row display, obtains the positive false information of label corresponding with sample label information；It is corrected errors information evaluation the first dimensionality reduction feature according to label,
Obtain signature contributions degree information corresponding with the first dimensionality reduction feature；The power of the first dimensionality reduction feature is determined according to signature contributions degree information
Weight information.
In one embodiment, model training module includes: that general column feature and general row feature are divided into predetermined quantity
The training characteristics collection of part；It will be successively trained in the input variable of training characteristics collection input basic model, until all training
Feature set training finishes, the object module trained.
Specific restriction about the model building device based on transfer learning may refer to above for based on transfer learning
The restriction of modeling method, details are not described herein.Modules in the above-mentioned model building device based on transfer learning can whole or portion
Divide and is realized by software, hardware and combinations thereof.Above-mentioned each module can be embedded in the form of hardware or independently of computer equipment
In processor in, can also be stored in a software form in the memory in computer equipment, in order to processor calling hold
The corresponding operation of the above modules of row.The processor can be central processing unit (CPU), microprocessor, single-chip microcontroller etc..On
Stating the model building device based on transfer learning can be implemented as a kind of form of computer program.
In one embodiment, a kind of computer equipment is provided, which can be server, be also possible to
Terminal.When the computer equipment is terminal, internal structure chart can be as shown in Figure 6.The computer equipment includes passing through to be
Processor, memory and the network interface of bus of uniting connection.Wherein, the processor of the computer equipment is calculated and is controlled for providing
Ability processed.The memory of the computer equipment includes non-volatile memory medium, built-in storage.The non-volatile memory medium is deposited
Contain operating system and computer program.The built-in storage is operating system and computer program in non-volatile memory medium
Operation provide environment.The network interface of the computer equipment is used to communicate with external terminal by network connection.The calculating
To realize a kind of modeling method based on transfer learning when machine program is executed by processor.It will be understood by those skilled in the art that
Structure shown in Fig. 6, only the block diagram of part-structure relevant to application scheme, is not constituted to application scheme institute
The restriction for the computer equipment being applied thereon, specific computer equipment may include than more or fewer portions as shown in the figure
Part perhaps combines certain components or with different component layouts.
Wherein, it is performed the steps of when processor executes program and obtains exemplar to be learned and target labels sample；It will
Exemplar and target labels sample to be learned carries out kernel principal component analysis, obtains first drop corresponding with exemplar to be learned
Dimensional feature, the second dimensionality reduction feature corresponding with target labels sample；First dimensionality reduction feature and the input of the second dimensionality reduction feature have been instructed
Experienced generic features obtain in model, obtain general column feature；First dimensionality reduction feature is inputted corresponding with target labels sample
Basic model is tested, and weight information corresponding with the first dimensionality reduction feature is obtained, and weight information is higher than default weight threshold
The first dimensionality reduction feature as general row feature；General column feature and the input of general row feature is corresponding with target labels sample
Model training is carried out in basic model, obtains object module.
The above-mentioned restriction for computer equipment may refer to the tool above for the modeling method based on transfer learning
Body limits, and details are not described herein.
Please continue to refer to Fig. 6, a kind of computer readable storage medium is also provided, is stored thereon with computer program, such as Fig. 6
Shown in non-volatile memory medium, wherein the program performs the steps of when being executed by processor obtains label to be learned
Sample and target labels sample；Exemplar to be learned and target labels sample are subjected to kernel principal component analysis, obtain with wait learn
Practise the corresponding first dimensionality reduction feature of exemplar, the second dimensionality reduction feature corresponding with target labels sample；By the first dimensionality reduction feature
The generic features trained with the input of the second dimensionality reduction feature obtain in model, obtain general column feature；First dimensionality reduction feature is defeated
Enter basic model corresponding with target labels sample to be tested, obtains weight information corresponding with the first dimensionality reduction feature, will weigh
Weight information is higher than the first dimensionality reduction feature of default weight threshold as general row feature；General column feature and general row feature is defeated
Enter in basic model corresponding with target labels sample and carry out model training, obtains object module.
The above-mentioned restriction for computer readable storage medium may refer to above for the modeling based on transfer learning
The specific restriction of method, details are not described herein.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the program can be stored in a non-volatile computer and can be read
In storage medium, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, the storage is situated between
Matter can be magnetic disk, CD, read-only memory (Read-OnlyMemory, ROM) etc..
Each technical characteristic of embodiment described above can be combined arbitrarily, for simplicity of description, not to above-mentioned reality
It applies all possible combination of each technical characteristic in example to be all described, as long as however, the combination of these technical characteristics is not deposited
In contradiction, all should be considered as described in this specification.
The embodiments described above only express several embodiments of the present invention, and the description thereof is more specific and detailed, but simultaneously
It cannot therefore be construed as limiting the scope of the patent.It should be pointed out that coming for those of ordinary skill in the art
It says, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to protection of the invention
Range.Therefore, the scope of protection of the patent of the invention shall be subject to the appended claims.
Claims (10)
1. a kind of modeling method based on transfer learning, which comprises
Obtain exemplar and target labels sample to be learned；
The exemplar to be learned and the target labels sample are subjected to kernel principal component analysis, obtained and the mark to be learned
The corresponding first dimensionality reduction feature of signed-off sample sheet, the second dimensionality reduction feature corresponding with the target labels sample；
The generic features that the first dimensionality reduction feature and the second dimensionality reduction feature input have been trained are obtained in model, are led to
With column feature；
The first dimensionality reduction feature is inputted corresponding with target labels sample basic model to test, obtain with it is described
The weight information is higher than the first dimensionality reduction feature of default weight threshold as logical by the corresponding weight information of the first dimensionality reduction feature
With row feature；
By the general column feature and the general row feature input in basic model corresponding with the target labels sample into
Row model training, obtains object module.
2. the method according to claim 1, wherein described by the general column feature and the general row feature
It inputs in corresponding with target labels sample basic model and carries out model training, after obtaining object module, further includes:
Sample to be evaluated is obtained, the sample to be evaluated is inputted in the object module, output and the sample pair to be evaluated
The sample label information answered；
The sample label information is shown, label correct information corresponding with the sample label information is obtained；
The weight in the object module is adjusted according to the label correct information, according to the weight pair after each adjust
The object module is updated, and obtains updated object module.
3. the method according to claim 1, wherein described by the first dimensionality reduction feature and second dimensionality reduction
The generic features that feature input has been trained obtain in model, obtain general column feature, comprising:
The first dimensionality reduction feature and the second dimensionality reduction feature are subjected to aspect ratio pair, obtain characteristic similarity；
The characteristic similarity is higher than the first dimensionality reduction feature when presetting similar threshold value as general column feature.
4. the method according to claim 1, wherein described by the first dimensionality reduction feature input and the target
The corresponding basic model of exemplar is tested, and weight information corresponding with the first dimensionality reduction feature is obtained, comprising:
The first dimensionality reduction feature is inputted corresponding with target labels sample basic model to test, export with it is described
The corresponding sample label information of exemplar to be learned；
The sample label information is shown, the positive false information of label corresponding with the sample label information is obtained；
It is corrected errors the first dimensionality reduction feature described in information evaluation according to the label, obtains feature corresponding with the first dimensionality reduction feature
Contribution degree information；
The weight information of the first dimensionality reduction feature is determined according to the signature contributions degree information.
5. the method according to claim 1, wherein described by the general column feature and the general row feature
It inputs basic model corresponding with the target labels sample and carries out model training, obtain object module, comprising:
The general column feature and the general row feature are divided into the training characteristics collection of predetermined quantity part；
Successively the training characteristics collection is inputted in the input variable of the basic model and be trained, until all training characteristics
Collection training finishes, the object module trained.
6. a kind of model building device based on transfer learning, which is characterized in that described device includes:
Sample acquisition module, for obtaining exemplar and target labels sample to be learned；
Feature Dimension Reduction module, for the exemplar to be learned and the target labels sample to be carried out kernel principal component analysis,
First dimensionality reduction feature corresponding with the exemplar to be learned is obtained, the second dimensionality reduction corresponding with the target labels sample is special
Sign；
Column feature obtains module, for the first dimensionality reduction feature and the second dimensionality reduction feature to be inputted the general spy trained
Sign obtains in model, obtains general column feature；
Row feature obtains module, for the first dimensionality reduction feature to be inputted basic model corresponding with the target labels sample
It is tested, obtains weight information corresponding with the first dimensionality reduction feature, the weight information is higher than default weight threshold
The first dimensionality reduction feature as general row feature；
Model training module, for inputting and the target labels sample pair the general column feature and the general row feature
Model training is carried out in the basic model answered, and obtains object module.
7. device according to claim 6, which is characterized in that the model training module includes:
Label information output module inputs the sample to be evaluated in the object module for obtaining sample to be evaluated, defeated
Sample label information corresponding with the sample to be evaluated out；
Correct information obtains module, for showing the sample label information, obtains and the sample label information pair
The label correct information answered；
Model modification module, for the weight in the object module to be adjusted according to the label correct information, according to
Weight after adjusting every time is updated the object module, obtains updated object module.
8. device according to claim 6, which is characterized in that the column feature obtains module and includes:
Feature comparison module obtains spy for the first dimensionality reduction feature and the second dimensionality reduction feature to be carried out aspect ratio pair
Levy similarity；
Similarity judgment module, for the first dimensionality reduction feature when the characteristic similarity to be higher than to default similar threshold value as logical
With column feature.
9. a kind of computer equipment, including memory and processor, the memory are stored with computer program, feature exists
In the step of processor realizes any one of claims 1 to 5 the method when executing the computer program.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
The step of any one of claims 1 to 5 the method is realized when being executed by processor.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910418820.5A CN110210625B (en) | 2019-05-20 | 2019-05-20 | Modeling method and device based on transfer learning, computer equipment and storage medium |
| PCT/CN2019/102740 WO2020232874A1 (en) | 2019-05-20 | 2019-08-27 | Modeling method and apparatus based on transfer learning, and computer device and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910418820.5A CN110210625B (en) | 2019-05-20 | 2019-05-20 | Modeling method and device based on transfer learning, computer equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110210625A true CN110210625A (en) | 2019-09-06 |
| CN110210625B CN110210625B (en) | 2023-04-07 |
Family
ID=67787850
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910418820.5A Active CN110210625B (en) | 2019-05-20 | 2019-05-20 | Modeling method and device based on transfer learning, computer equipment and storage medium |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN110210625B (en) |
| WO (1) | WO2020232874A1 (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110929877A (en) * | 2019-10-18 | 2020-03-27 | 平安科技（深圳）有限公司 | Model establishing method, device, equipment and storage medium based on transfer learning |
| CN112581250A (en) * | 2019-09-30 | 2021-03-30 | 深圳无域科技技术有限公司 | Model generation method and device, computer equipment and storage medium |
| WO2022100491A1 (en) * | 2020-11-11 | 2022-05-19 | 中兴通讯股份有限公司 | Model training method and apparatus, and electronic device and computer-readable storage medium |
| CN116910573A (en) * | 2023-09-13 | 2023-10-20 | 中移(苏州)软件技术有限公司 | Training method and device for abnormality diagnosis model, electronic equipment and storage medium |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113159085B (en) * | 2020-12-30 | 2024-05-28 | 北京爱笔科技有限公司 | Classification model training and image-based classification method and related device |
| CN115396831A (en) * | 2021-05-08 | 2022-11-25 | 中国移动通信集团浙江有限公司 | Interaction model generation method, device, equipment and storage medium |
| CN114021180B (en) * | 2021-10-11 | 2024-04-12 | 清华大学 | Dynamic security domain determining method and device for power system, electronic equipment and readable medium |
| CN117708592B (en) * | 2023-12-12 | 2024-05-07 | 清新文化艺术有限公司 | Art science and technology integration digital platform based on cultural creative |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104091602A (en) * | 2014-07-11 | 2014-10-08 | 电子科技大学 | Speech emotion recognition method based on fuzzy support vector machine |
| CN104523269A (en) * | 2015-01-15 | 2015-04-22 | 江南大学 | Self-adaptive recognition method orienting epilepsy electroencephalogram transfer environment |
| EP2993618A1 (en) * | 2014-09-04 | 2016-03-09 | Xerox Corporation | Domain adaptation for image classification with class priors |
| WO2019015461A1 (en) * | 2017-07-18 | 2019-01-24 | 中国银联股份有限公司 | Risk identification method and system based on transfer deep learning |
| US20190107404A1 (en) * | 2017-06-13 | 2019-04-11 | Beijing Didi Infinity Technology And Development Co., Ltd. | Systems and methods for determining estimated time of arrival |
| CN109710512A (en) * | 2018-12-06 | 2019-05-03 | 南京邮电大学 | Neural network software failure prediction method based on geodesic curve stream core |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106030608A (en) * | 2013-11-06 | 2016-10-12 | 理海大学 | Diagnostic system and method for biological tissue analysis |
| CN107506775A (en) * | 2016-06-14 | 2017-12-22 | 北京陌上花科技有限公司 | model training method and device |
| CN106326214A (en) * | 2016-08-29 | 2017-01-11 | 中译语通科技（北京）有限公司 | Method and device for cross-language emotion analysis based on transfer learning |
| CN107292246A (en) * | 2017-06-05 | 2017-10-24 | 河海大学 | Infrared human body target identification method based on HOG PCA and transfer learning |
-
2019
- 2019-05-20 CN CN201910418820.5A patent/CN110210625B/en active Active
- 2019-08-27 WO PCT/CN2019/102740 patent/WO2020232874A1/en active Application Filing
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104091602A (en) * | 2014-07-11 | 2014-10-08 | 电子科技大学 | Speech emotion recognition method based on fuzzy support vector machine |
| EP2993618A1 (en) * | 2014-09-04 | 2016-03-09 | Xerox Corporation | Domain adaptation for image classification with class priors |
| CN104523269A (en) * | 2015-01-15 | 2015-04-22 | 江南大学 | Self-adaptive recognition method orienting epilepsy electroencephalogram transfer environment |
| US20190107404A1 (en) * | 2017-06-13 | 2019-04-11 | Beijing Didi Infinity Technology And Development Co., Ltd. | Systems and methods for determining estimated time of arrival |
| WO2019015461A1 (en) * | 2017-07-18 | 2019-01-24 | 中国银联股份有限公司 | Risk identification method and system based on transfer deep learning |
| CN109710512A (en) * | 2018-12-06 | 2019-05-03 | 南京邮电大学 | Neural network software failure prediction method based on geodesic curve stream core |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112581250A (en) * | 2019-09-30 | 2021-03-30 | 深圳无域科技技术有限公司 | Model generation method and device, computer equipment and storage medium |
| CN112581250B (en) * | 2019-09-30 | 2023-12-29 | 深圳无域科技技术有限公司 | Model generation method, device, computer equipment and storage medium |
| CN110929877A (en) * | 2019-10-18 | 2020-03-27 | 平安科技（深圳）有限公司 | Model establishing method, device, equipment and storage medium based on transfer learning |
| CN110929877B (en) * | 2019-10-18 | 2023-09-15 | 平安科技（深圳）有限公司 | Model building method, device, equipment and storage medium based on transfer learning |
| WO2022100491A1 (en) * | 2020-11-11 | 2022-05-19 | 中兴通讯股份有限公司 | Model training method and apparatus, and electronic device and computer-readable storage medium |
| CN116910573A (en) * | 2023-09-13 | 2023-10-20 | 中移(苏州)软件技术有限公司 | Training method and device for abnormality diagnosis model, electronic equipment and storage medium |
| CN116910573B (en) * | 2023-09-13 | 2023-12-05 | 中移(苏州)软件技术有限公司 | Training method and device for abnormality diagnosis model, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110210625B (en) | 2023-04-07 |
| WO2020232874A1 (en) | 2020-11-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110210625A (en) | Modeling method, device, computer equipment and storage medium based on transfer learning | |
| CN110363138A (en) | Model training method, image processing method, device, terminal and storage medium | |
| CN110619059B (en) | Building marking method based on transfer learning | |
| CN103258210B (en) | A kind of high-definition image classification method based on dictionary learning | |
| CN108021930B (en) | Self-adaptive multi-view image classification method and system | |
| CN110516095A (en) | Weakly supervised depth Hash social activity image search method and system based on semanteme migration | |
| CN105354595A (en) | Robust visual image classification method and system | |
| CN103177265B (en) | High-definition image classification method based on kernel function Yu sparse coding | |
| CN110765882A (en) | Video tag determination method, device, server and storage medium | |
| CN105469063A (en) | Robust human face image principal component feature extraction method and identification apparatus | |
| CN108228684A (en) | Training method, device, electronic equipment and the computer storage media of Clustering Model | |
| CN109255029A (en) | A method of automatic Bug report distribution is enhanced using weighted optimization training set | |
| CN110781970A (en) | Method, device and equipment for generating classifier and storage medium | |
| CN113822264A (en) | Text recognition method and device, computer equipment and storage medium | |
| CN108877947A (en) | Depth sample learning method based on iteration mean cluster | |
| Wang et al. | Research on maize disease recognition method based on improved resnet50 | |
| Intrator | Making a low-dimensional representation suitable for diverse tasks | |
| CN110348287A (en) | A kind of unsupervised feature selection approach and device based on dictionary and sample similar diagram | |
| CN112420125A (en) | Molecular attribute prediction method and device, intelligent equipment and terminal | |
| CN113705092B (en) | Disease prediction method and device based on machine learning | |
| CN107480627A (en) | Activity recognition method, apparatus, storage medium and processor | |
| Bi et al. | Critical direction projection networks for few-shot learning | |
| CN117371511A (en) | Training method, device, equipment and storage medium for image classification model | |
| Nguyen et al. | Robust product classification with instance-dependent noise | |
| CN111144453A (en) | Method and equipment for constructing multi-model fusion calculation model and method and equipment for identifying website data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |