CN112257806B - Heterogeneous user-oriented migration learning method - Google Patents
Heterogeneous user-oriented migration learning method Download PDFInfo
- Publication number
- CN112257806B CN112257806B CN202011195428.8A CN202011195428A CN112257806B CN 112257806 B CN112257806 B CN 112257806B CN 202011195428 A CN202011195428 A CN 202011195428A CN 112257806 B CN112257806 B CN 112257806B
- Authority
- CN
- China
- Prior art keywords
- data
- classification
- matrix
- sample
- sample data
- 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.)
- Active
Links
- 238000013508 migration Methods 0.000 title claims abstract description 15
- 230000005012 migration Effects 0.000 title claims abstract description 15
- 238000000034 method Methods 0.000 title claims abstract description 14
- 230000009467 reduction Effects 0.000 claims abstract description 16
- 238000007635 classification algorithm Methods 0.000 claims abstract description 8
- 239000011159 matrix material Substances 0.000 claims description 45
- 238000013507 mapping Methods 0.000 claims description 12
- 239000013598 vector Substances 0.000 claims description 9
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 4
- 238000012549 training Methods 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 claims description 3
- 230000015572 biosynthetic process Effects 0.000 claims description 3
- 125000004122 cyclic group Chemical group 0.000 claims description 3
- 230000006870 function Effects 0.000 claims description 3
- 238000003786 synthesis reaction Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 2
- 238000012847 principal component analysis method Methods 0.000 claims 1
- 238000013526 transfer learning Methods 0.000 abstract description 8
- 230000000694 effects Effects 0.000 abstract description 5
- 238000012216 screening Methods 0.000 abstract description 4
- 238000010801 machine learning Methods 0.000 description 5
- 238000013145 classification model Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention discloses a heterogeneous user-oriented migration learning method, wherein a server and other participants cannot obtain original data, so that the risk of privacy disclosure is reduced to a certain extent. Secondly, through domain delimitation and secondary dimension reduction screening, the correlation between sample data and a classification target is higher, the method can adapt to user isomerism, the classification effect is better, and the requirement of high classification accuracy can be met to a great extent. On the other hand, the circulation double-classification algorithm of Softmax and CNN has supervised learning to guide unsupervised learning, and improves the classification accuracy of the insufficient label data. The invention selects and delimits the data acquired by the local end multi-channel to ensure that the transfer learning has enough data quantity. On the basis, the requirement of multi-target output is met, and the classification accuracy is improved.
Description
Technical Field
The invention relates to the technical field of machine learning, in particular to a heterogeneous user-oriented migration learning method.
Background
With the continued development and maturation of traditional machine learning, it has been relatively easy to train a good classification model from a large amount of tagged data. However, in a real application scenario, the conventional machine learning method still cannot fully apply the requirements. On the one hand, it is relatively difficult to obtain tagged data. Most of the data generated in life do not contain labels, and the cost of manual labels is too high; and the data acquisition often also considers personal privacy and security problems, which also increases the difficulty of data acquisition. On the other hand, conventional machine learning requires re-modeling and training each time data is updated, thereby consuming a lot of time and resources.
The transfer learning relieves the data pressure of the traditional machine learning to a certain extent, but the transfer learning can not be performed in any situation, and the effect of the transfer is also influenced by a plurality of factors. Most of the research today uses random source domain data resulting in poor classification accuracy and inability to accommodate user heterogeneity, i.e., inability to meet multi-objective classification requirements. When using the data acquired by multiple channels, the accuracy of the classification result is reduced due to the large difference of the data correlation, and the source domain and the target domain may be randomly determined, so that the transfer learning cannot exert the advantages of sufficient data quantity, but the learning efficiency is not high and the accuracy cannot be ensured. Under the limitation of factors in various aspects, the application of transfer learning is not very wide, most researches only propose specific algorithms for transaction classification in a certain field, and a complete model architecture is not provided.
In summary, the existing classification model does not have a complete flow from data acquisition and data processing to a classification algorithm, the problem of multi-target output cannot be met, and the classification accuracy is difficult to guarantee.
Disclosure of Invention
The invention aims to provide a heterogeneous user-oriented migration learning method.
The technical scheme adopted by the invention is as follows:
a heterogeneous user-oriented migration learning method comprises the following steps:
step 1, a participant performs data acquisition and primary processing on a local side to realize first data dimension reduction.
And 2, selecting and delimiting a source domain and a target domain according to the participant demand server side, and realizing second data dimension reduction.
And 3, classifying by using an S-CNN cyclic classification algorithm.
Further, the specific steps of step 1 are as follows:
step 1-1, participants locally based on the raw data X n×h Calculating covariance matrix F of the data:
where n is the number of entries of the participant's local data, h is the data dimension;
step 1-2, calculating all eigenvalues lambda and corresponding eigenvectors mu thereof according to |lambda E-F|=0, wherein E is an identity matrix;
step 1-3, for the eigenvalue lambda i (λ i Epsilon lambda), and selecting the number of principal components according to a preset threshold r;
step 1-4, outputting a feature vector set (mu) corresponding to the first r feature values 1 ,μ 2 ,…,μ r ) Calculating the modulus of the feature vectors, unitizing the corresponding r feature vectors to form a feature matrix A;
step 1-5, calculating a projection matrix X' n×r ＝X n×h A(r<h) Obtaining a new data sample X';
step 1-6, the server receives and stores the local reduced-dimension data set uploaded by all participants to form a sample data pool
wherein X′v A sample data matrix uploaded by the v-th participant in the data pool is represented, and N represents the number of the participants;
further, the specific steps of the step 2 are as follows:
step 2-1, participant u uploads its classification requirement D u ＝(N u ,M u ,acc u), wherein Nu For the number of source domains, M u For the number of categories, acc u Representing a minimum classification accuracy requirement;
step 2-2, the server calculates each data sample X in the data pool v ' data sample X uploaded with participant u u ' correlation I (X) v ',X u '), the calculation method is as follows:
wherein X 'represents a set of data of a data matrix X'; p (x) v ',x u ' indicates x v ' and x u ' joint probability distribution of two sets of data; probability distribution of data x 'represented by P (x'); p (x) v '|x u ') represents data x v At data x u ' sProbability distribution; KL represents distance and is short for the difference of Kullback-Leibler;
step 2-3, based on correlation I (X v ',X u ') to X v 'ranking from high to low, selecting top N with high relevance according to the participant's needs u Source field X used as current migration learning S Sample data X of participants u ' as target Domain X T ；
Step 2-4, performing secondary dimension reduction: and mapping the data in a plurality of fields to the same dimensional space by using a migration component analysis TCA algorithm, wherein feature mapping is accompanied with the reduction of feature numbers, and finally, obtaining a new data feature sample matrix.
Further, as a preferred embodiment, the specific steps of steps 2 to 4 are as follows:
step 2-4-1, defining a kernel matrix K: respectively calculate X S 、X T And a kernel matrix K of two synthesis domains S,S 、K T,T 、K T,S and KS,T Then constructing a kernel matrix K by using the formula (1);
wherein K is one (n 1 +n 2 )×(n 1 +n 2 ) N of the matrix of (a) 1 and n2 Respectively X S and XT The number of samples mapped to the Regenerated Kernel Hilbert Space (RKHS);
step 2-4-2, decomposing the kernel matrix K into k= (KK) using an empirical kernel mapping function -1/2 )(K -1/2 K)；
Step 2-4-3, calculating the characteristic distance between the source domain and the target domain according to the formula 2 according to K:
Dist(X S ,X T )＝tr(KL) (2)
where tr (KL) represents the trace of matrix KL.
Step 2-4-4, calculating W εR according to equation 3 (n1+n2)×m (m<n1+n2)：
wherein ,
x represents S and XT Maximum mean error (MMD) distance between the empirical mean of two domains, i.e. X S and XT KL distance of two domains; />
Is the result of empirical kernel mapping of the m-dimensional space of W;
step 2-4-5, outputting the final source domain matrix W XS And a target domain matrix W XT ；
Further, as a preferred embodiment, the specific classification step of step 3 is as follows:
step 3-1, initializing: first training W XS Obtaining an initialized Softmax classifier;
step 3-2, from W XT Selecting part of samples, and initializing the cycle discrimination frequency q of the batch of data to be 0;
step 3-3, predicting and marking unlabeled sample data in the labeled sample data through a Softmax classifier, and adding a pseudo label to the labeled sample data to obtain a primary classification result;
step 3-4, the batch of samples are also subjected to CNN classifier to obtain a classification result;
step 3-5, comparing the primary classification result with the secondary classification result; when the two are inconsistent, let q=q+1, and judge whether Q is greater than threshold value Q, if yes, delete this batch of data and return to step 3-2, otherwise return to step 3-3;
step 3-6, using the sample data to train a Softmax classifier, obtaining classification accuracy acc and matching the participant's required accuracy acc u Comparing; when acc is greater than acc u Outputting acc and the classification result; otherwise, returning to the step 3-3.
By adopting the technical scheme, firstly, the server and other participants cannot obtain the original data, so that the risk of privacy disclosure is reduced to a certain extent. Secondly, through domain delimitation and secondary dimension reduction screening, the correlation between sample data and a classification target is higher, the method can adapt to user isomerism, the classification effect is better, and the requirement of high classification accuracy can be met to a great extent. On the other hand, the circulation double-classification algorithm of Softmax and CNN has supervised learning to guide unsupervised learning, and improves the classification accuracy of the insufficient label data. The invention selects and delimits the data acquired by the local end multi-channel to ensure that the transfer learning has enough data quantity. On the basis, the requirement of multi-target output is met, and the classification accuracy is improved.
Drawings
The invention is described in further detail below with reference to the drawings and detailed description;
FIG. 1 is a flow chart of a model for transitional learning adapted to user profile according to the present invention;
fig. 2 is a flow chart of the loop classification of the S-CNN algorithm of the present invention.
Detailed Description
For the purposes, technical solutions and advantages of the embodiments of the present application, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.
As shown in fig. 1 or fig. 2, the invention discloses a heterogeneous user-oriented migration learning method, which comprises the following steps:
step 1, a participant performs data acquisition and primary processing on a local side to realize first data dimension reduction.
And 2, selecting and delimiting a source domain and a target domain according to the participant demand server side, and realizing second data dimension reduction.
And 3, classifying by using an S-CNN cyclic classification algorithm.
Further, as a preferred embodiment, the specific steps of step 1 are as follows:
step 1-1, participants locally use raw data X n×h Meter (D)Calculating a data covariance matrix F:
n is the number of entries of the local original data of the participant, and h is the dimension of the data;
step 1-2, calculating all eigenvalues lambda and corresponding eigenvectors mu thereof according to |lambda E-F|=0, wherein E is an identity matrix;
step 1-3, for the eigenvalue lambda i (λ i Epsilon lambda), and selecting the number of principal components according to a preset threshold r;
step 1-4, outputting a feature vector set (mu) corresponding to the first r feature values 1 ,μ 2 ,…,μ r ) Calculating the modulus of the feature vectors, unitizing the corresponding r feature vectors to form a feature matrix A;
step 1-5, calculating a projection matrix X' n×r ＝X n×h A(r<h) Obtaining a new data sample X';
and step 1-6, the server receives and stores the local dimensionality reduced data set uploaded by all the participants. Forming a data pool
wherein X′v A sample data matrix uploaded by the v-th participant in the data pool is represented, and N represents the number of the participants;
the specific steps of the step 2 are as follows:
step 2-1, participant u uploads its classification requirement (N u ,M u ,acc u), wherein Nu For the number of source domains, M u For the number of categories, acc u Representing the lowest classification accuracy;
step 2-2, the server calculates each data sample matrix X in the data pool v ' data sample X uploaded with participant u u The 'correlation' is calculated as follows:
wherein I represents a data matrix X v ' and X u ' correlation between; a set of data of a data matrix X 'denoted by X'; p (x) v ',x u ' indicates x v ' and x u ' joint probability distribution of two sets of data; probability distribution of data x represented by P (x); p (x) v '|x u ') represents data x v At data x u ' probability distribution; KL represents distance and is short for the difference of Kullback-Leibler;
step 2-3, based on correlation I (X v ',X u ') to X v 'ranking from high to low, selecting top N with high relevance according to the participant's needs u Source field X used as current migration learning S Sample data X of participants u ' as target Domain X T ；
Step 2-4, performing secondary dimension reduction: and mapping the data in a plurality of fields to the same dimensional space by using a migration component analysis TCA algorithm, wherein feature mapping is accompanied with the reduction of feature numbers, and finally, obtaining a new data feature sample matrix.
Further, as a preferred embodiment, the specific steps of steps 2 to 4 are as follows:
step 2-4-1, defining a kernel matrix K: respectively calculate X S 、X T And a kernel matrix K of two synthesis domains S,S 、K T,T 、K T,S and KS,T Then constructing a kernel matrix K by using the formula (1);
wherein K is one (n 1 +n 2 )×(n 1 +n 2 ) N of the matrix of (a) 1 and n2 Respectively X S and XT The number of samples mapped to the Regenerated Kernel Hilbert Space (RKHS);
step 2-4-2, decomposing the kernel matrix K into k= (KK) using an empirical kernel mapping function -1/2 )(K -1/2 K)；
Step 2-4-3, calculating the characteristic distance between the source domain and the target domain according to the formula 2 according to K:
Dist(X S ,X T )＝tr(KL) (2)
where tr (KL) represents the trace of matrix KL.
Step 2-4-4, calculating W according to equation 3:
wherein ,
x represents S and XT Maximum mean error (MMD) distance between the empirical mean of two domains, i.e. X S and XT KL distance of two domains; />
Is the result of empirical kernel mapping of the m-dimensional space of W;
step 2-4-5, outputting the final source domain matrix W XS And a target domain matrix W XT ；
Further, as a preferred embodiment, the specific classification step of step 3 is as follows:
step 3-1, initializing: first training W XS Obtaining an initialized Softmax classifier;
step 3-2, from W XT Selecting part of samples, and initializing the cycle discrimination frequency q of the batch of data to be 0;
step 3-3, predicting and marking unlabeled sample data in the labeled sample data through a Softmax classifier, and adding a pseudo label to the labeled sample data to obtain a primary classification result;
step 3-4, the batch of samples are also subjected to CNN classifier to obtain a classification result;
step 3-5, comparing the primary classification result with the secondary classification result; when the two are inconsistent, let q=q+1, and judge whether Q is greater than threshold value Q, if yes, delete this batch of data and return to step 3-2, otherwise return to step 3-3;
step 3-6, using the sample data to train a Softmax classifier, obtaining a classification accuracy acc and combining the classification accuracy acc with the accuracy acc required by participant i u Comparing; when the classification accuracy acc is greater than acc u Outputting classification accuracy acc and classification results; otherwise, returning to the step 3-3.
By adopting the technical scheme, firstly, the server and other participants cannot obtain the original data, so that the risk of privacy disclosure is reduced to a certain extent. Secondly, through domain delimitation and secondary dimension reduction screening, the correlation between sample data and a classification target is higher, the method can adapt to user isomerism, the classification effect is better, and the requirement of high classification accuracy can be met to a great extent. On the other hand, the circulation double-classification algorithm of Softmax and CNN has supervised learning to guide unsupervised learning, and improves the classification accuracy of the insufficient label data. The invention selects and delimits the data acquired by the local end multi-channel to ensure that the transfer learning has enough data quantity. On the basis, the requirement of multi-target output is met, and the classification accuracy is improved.
By adopting the technical scheme, firstly, the server and other participants cannot obtain the original data, so that the risk of privacy disclosure is reduced to a certain extent. Secondly, through domain delimitation and secondary dimension reduction screening, the correlation between sample data and a classification target is higher, the method can adapt to user isomerism, the classification effect is better, and the requirement of high classification accuracy can be met to a great extent. On the other hand, the circulation double-classification algorithm of Softmax and CNN has supervised learning to guide unsupervised learning, and improves the classification accuracy of the insufficient label data. The invention selects and delimits the data acquired by the local end multi-channel to ensure that the transfer learning has enough data quantity. On the basis, the requirement of multi-target output is met, and the classification accuracy is improved.
It will be apparent that the embodiments described are some, but not all, of the embodiments of the present application. Embodiments and features of embodiments in this application may be combined with each other without conflict. The components of the embodiments of the present application, which are generally described and illustrated in the figures herein, may be arranged and designed in a wide variety of different configurations. Thus, the detailed description of the embodiments of the present application is not intended to limit the scope of the application, as claimed, but is merely representative of selected embodiments of the application. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without making any inventive effort, are intended to be within the scope of the present application.
Claims (2)
1. A heterogeneous user-oriented migration learning method is characterized by comprising the following steps of: which comprises the following steps:
step 1, a participant performs data acquisition and primary processing at a local end, and performs primary dimension reduction on original data by using a principal component analysis method to obtain sample data;
step 2, calculating the correlation between the sample data provided by the participants and other sample data in the data pool by the server side according to the requirements of the participants; and according to the correlation, all sample data are arranged in descending order; finally, determining a source domain and a target domain according to the number of source domains defined by a user, and realizing second data dimension reduction; the specific steps of the step 2 are as follows:
step 2-1, participant u uploads its classification requirement D u ＝(N u ,M u ,acc u), wherein Nu For the number of source domains, M u For the number of categories, acc u Representing a minimum classification accuracy requirement;
step 2-2, the server calculates each data sample X in the data pool v ' data sample X uploaded with participant u u ' correlation I (X) v ',X u '), the calculation method is as follows:
wherein X 'is a set of data for X'; p (x) v ',x u ' indicates x v ' and x u ' joint probability distribution; p (x ') represents the probability distribution of x'; KL represents distance, which is the difference between Kullback and LeiblerIs abbreviated as (1);
step 2-3, according to I (X v ',X u ') to X v ' order from high to low; according to D u Selecting front N with high correlation u Source field X used as current migration learning S ，X u ' as target Domain X T ；
Step 2-4, performing secondary dimension reduction: mapping data in multiple fields to the same dimensional space by using a migration component analysis TCA algorithm, wherein feature mapping is accompanied with feature number reduction, and finally a new data feature sample matrix W is obtained; the specific steps of the steps 2-4 are as follows:
step 2-4-1, defining a kernel matrix K: respectively calculate X S 、X T And a kernel matrix K of two synthesis domains S,S 、K T,T 、K T,S and KS,T Constructing K by using the formula (1);
wherein K is one (n 1 +n 2 )×(n 1 +n 2 ) N of the matrix of (a) 1 and n2 Respectively X S and XT The number of samples mapped to the Regenerated Kernel Hilbert Space (RKHS);
step 2-4-2, decomposing the kernel matrix K into k= (KK) using an empirical kernel mapping function -1/2 )(K -1/2 K)；
Step 2-4-3, calculating the characteristic distance between the source domain and the target domain according to the formula (2) according to K:
Dist(X S ,X T )＝tr(KL) (2)
where tr (KL) represents the trace of matrix KL;
step 2-4-4, calculating W according to equation (3):
wherein ,
x represents S and XT Maximum mean error (MMD) distance between the empirical mean of two domains, i.e. X S and XT KL distance of two domains; />
Is the empirical kernel mapping result of the m-dimensional space of W;
step 2-4-5, outputting the final source domain matrix W XS And a target domain matrix W XT ；
Step 3, classifying by using an S-CNN cyclic classification algorithm: firstly, inputting sample data in batches to train a Softmax classifier to obtain a data primary classification result; then, inputting the sample data of the same batch into a CNN classifier to obtain a classification result, and comparing the two classification results; if the sample data are consistent, circularly sending the sample data of the batch back to the Softmax classifier, further optimizing the classifier until the accuracy meets the requirement, and finally outputting a classification result; if the sample data are inconsistent, the batch of sample data are returned to the Softmax classifier for reclassifying, and a new round of judgment is carried out; the specific classification step of the step 3 is as follows:
step 3-1, initializing: first training W XS Obtaining an initialized Softmax classifier;
step 3-2, from W XT Selecting part of samples, and initializing the cycle discrimination frequency q of the batch of data to be 0;
step 3-3, predicting and marking unlabeled sample data in the labeled sample data through a Softmax classifier, and adding a pseudo label to the labeled sample data to obtain a primary classification result;
step 3-4, the batch of samples are also subjected to CNN classifier to obtain a classification result;
step 3-5, comparing the primary classification result with the secondary classification result; when the two are inconsistent, let q=q+1, and judge whether Q is greater than threshold value Q, if yes, delete this batch of data and return to step 3-2, otherwise return to step 3-3;
step 3-6, using the sample dataTraining a Softmax classifier to obtain classification accuracy acc and to match the accuracy acc required by participant i u Comparing; if acc is>acc u Outputting acc and the classification result; otherwise, returning to the step 3-3.
2. The heterogeneous user-oriented migration learning method of claim 1, wherein the method comprises the steps of: the specific steps of the step 1 are as follows:
step 1-1, participants locally based on the raw data X n×h Calculating covariance matrix F of the data:
where n is the number of entries of the participant's local data, h is the data dimension;
step 1-2, calculating all eigenvalues lambda and corresponding eigenvectors mu thereof according to |lambda E-F|=0, wherein E is an identity matrix;
step 1-3, for the eigenvalue lambda i (λ i Epsilon lambda), and selecting the number of principal components according to a preset threshold r;
step 1-4, outputting a feature vector set (mu) corresponding to the first r feature values 1 ,μ 2 ,…,μ r ) Calculating the modulus of the feature vectors, unitizing the corresponding r feature vectors to form a feature matrix A;
step 1-5, calculating a projection matrix X' n×r ＝X n×h A(r<h) Obtaining a new data sample X';
step 1-6, the server receives and stores the local reduced-dimension data set uploaded by all participants to form a sample data pool
wherein X′v Representing a sample data matrix uploaded by the v-th participant in the data pool, and N represents the number of participants.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011195428.8A CN112257806B (en) | 2020-10-30 | 2020-10-30 | Heterogeneous user-oriented migration learning method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011195428.8A CN112257806B (en) | 2020-10-30 | 2020-10-30 | Heterogeneous user-oriented migration learning method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112257806A CN112257806A (en) | 2021-01-22 |
| CN112257806B true CN112257806B (en) | 2023-06-20 |
Family
ID=74267558
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011195428.8A Active CN112257806B (en) | 2020-10-30 | 2020-10-30 | Heterogeneous user-oriented migration learning method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112257806B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106095099A (en) * | 2016-06-12 | 2016-11-09 | 南京邮电大学 | A kind of user behavior motion detection recognition methods |
| CN110134868A (en) * | 2019-05-14 | 2019-08-16 | 辽宁工程技术大学 | A kind of recommended method based on the analysis of user preference isomerism |
| CN111400560A (en) * | 2020-03-10 | 2020-07-10 | 支付宝(杭州)信息技术有限公司 | Method and system for predicting based on heterogeneous graph neural network model |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10296846B2 (en) * | 2015-11-24 | 2019-05-21 | Xerox Corporation | Adapted domain specific class means classifier |
-
2020
- 2020-10-30 CN CN202011195428.8A patent/CN112257806B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106095099A (en) * | 2016-06-12 | 2016-11-09 | 南京邮电大学 | A kind of user behavior motion detection recognition methods |
| CN110134868A (en) * | 2019-05-14 | 2019-08-16 | 辽宁工程技术大学 | A kind of recommended method based on the analysis of user preference isomerism |
| CN111400560A (en) * | 2020-03-10 | 2020-07-10 | 支付宝(杭州)信息技术有限公司 | Method and system for predicting based on heterogeneous graph neural network model |
Non-Patent Citations (1)
| Title |
|---|
| 基于半监督的SVM迁移学习文本分类算法;谭建平;刘波;肖燕珊;;无线互联科技(04);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112257806A (en) | 2021-01-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107515895B (en) | Visual target retrieval method and system based on target detection | |
| CN110321926B (en) | Migration method and system based on depth residual error correction network | |
| CN108596154B (en) | Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion | |
| CN113326731A (en) | Cross-domain pedestrian re-identification algorithm based on momentum network guidance | |
| CN110766042B (en) | Multi-mark feature selection method and device based on maximum correlation minimum redundancy | |
| CN101140623A (en) | Video frequency objects recognition method and system based on supporting vectors machine | |
| CN113706547B (en) | Unsupervised domain adaptive semantic segmentation method based on category dissimilarity guidance | |
| CN115410088B (en) | Hyperspectral image field self-adaption method based on virtual classifier | |
| Guo et al. | Deep embedded k-means clustering | |
| Miao et al. | Hyperspectral imagery classification via random multigraphs ensemble learning | |
| CN112257806B (en) | Heterogeneous user-oriented migration learning method | |
| Sadeghi et al. | Deep multirepresentation learning for data clustering | |
| Hsieh et al. | Adaptive structural co-regularization for unsupervised multi-view feature selection | |
| Zhang et al. | Query-aware quantization for maximum inner product search | |
| Pinhas et al. | A continuous and probabilistic framework for medical image representation and categorization | |
| JP4302799B2 (en) | Document search apparatus, method, and recording medium | |
| CN114610884A (en) | Classification method based on PCA combined feature extraction and approximate support vector machine | |
| Chen et al. | Similarity-based retrieval of time-series data using multi-scale histograms | |
| CN111931757A (en) | Finger vein quick sorting method and device based on MDLBP block histogram and PCA dimension reduction | |
| Zhang et al. | Deep least squares alignment for unsupervised domain adaptation | |
| Papp et al. | Difference based query strategies in active learning | |
| Lei et al. | Extensions of Local-Mean K-Nearest Neighbor Classifier with Various Distance Metrics | |
| Wu et al. | Bayesian active learning in relevance feedback for image retrieval | |
| Baek et al. | Multi-choice wavelet thresholding based binary classification method | |
| Shen | Comparison and Evaluation of Classical Dimensionality Reduction Methods |
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 |