CN104463208A - Multi-view semi-supervised collaboration classification algorithm with combination of agreement and disagreement label rules - Google Patents
Multi-view semi-supervised collaboration classification algorithm with combination of agreement and disagreement label rules Download PDFInfo
- Publication number
- CN104463208A CN104463208A CN201410742319.1A CN201410742319A CN104463208A CN 104463208 A CN104463208 A CN 104463208A CN 201410742319 A CN201410742319 A CN 201410742319A CN 104463208 A CN104463208 A CN 104463208A
- Authority
- CN
- China
- Prior art keywords
- sample
- view
- sorter
- algorithm
- unmarked
- 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.)
- Pending
Links
Classifications
-
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
Abstract
A multi-view semi-supervised collaboration classification algorithm with the combination of agreement and disagreement label rules is put forward to improve the performance of a multi-view semi-supervised collaboration algorithm and solve the problem that the application range of the algorithm is limited. The algorithm combines the agreement label rules with the disagreement label rules; if classifiers have identical labels, corresponding samples are added to corresponding sample sets; if the classifiers have different labels and the difference value between confidence coefficients of labels corresponding to the two classifiers exceeds a certain threshold value, the label result of the high-confidence-coefficient classifier is adopted, and the samples are added to the corresponding sample sets. Combined labeling is conducted on unlabeled samples by judging whether the two classifiers agree with corresponding samples and judging a difference threshold value, the classifier difference judgment principle is used for updating the classification principle, useful information in the unlabeled samples is fully utilized, and the performance of the classifiers is improved by more than 5%.
Description
Technical field
The present invention relates to pattern-recognition and machine learning techniques field, particularly semisupervised classification algorithm, specifically refer to that a kind of multi views of composite marking rule works in coordination with semisupervised classification algorithm.
Background technology
Multi views mainly refers to approach to the description of things or angle, and therefore namely multi views describes the property set (data set) of things.Two views describing same thing have certain independence each other, and have the function that can describe things self character, and under many circumstances, two views also can be added by attribute and merge into a view, can reflect things itself better.The method of multi views can be applied in supervision, semi-supervised and unsupervised learning simultaneously.In the semi-supervised learning of multi views, data have multiple view to describe on the one hand; On the other hand, each viewdata comprises a small amount of flag data and a large amount of Unlabeled data two parts, and fundamental purpose utilizes the multi-level view with Unlabeled data to go to increase marker samples data set.More particularly, the multiple remarkable view of each example can be used for, for same task trains different models, then increasing the training sample of other sorter with the prediction of each sorter to Unlabeled data.In general, multi views method mainly utilizes consistance between the learner that to carry out training on the different views of same problem to improve the performance of sorter.
Standard C o-training algorithm is a kind of typical multi views semi-supervised learning method, and its proposition is that the further research of the semi-supervised coorinated training of multi views provides a theory fully and open learning framework.After standard C o-training algorithm, create two research directions, one is the semi-supervised learning of multi views, and two is semi-supervised learnings of single-view.The former mainly raw data concentrate and comprise multiple view, utilize the multiple sorter of multiple view generation, then by sorter model that the synergy between sorter trains.Work in coordination with in semi-supervised training process at multi views, how view publishing method, Unlabeled data are all the problem of scholars' primary study by the problem such as multi-classifier cooperate mark, model modification in multi views.
By the analysis to standard multi views Co-Training, can find out that problem main in Co-training algorithm comprises: the restriction of abundant redundancy condition, the generation method of multi-categorizer, confidence calculations method, sorter are worked in coordination with labeling method and apply for different application domain algorithms.
This problem of marking convention that multi views multi-categorizer consistency checking and nonuniformity judge was obtaining extensive attention in recent years.The consistance judgment principle of the utilization variance sorters such as Hahn S and non-uniform judgement principle marks for treatment data and Unlabeled data imbalance problem.Umit Guz etc., according to self-training and coorinated training two kinds of algorithms, in conjunction with consistance and nonuniformity two kinds of sample labeling strategies, propose a kind of self-combined algorithm to process sentence boundary classification problem, obtain good classification performance.Zhou Z H, in semi-supervised learning process, utilizes the inconsistency between multiple learner training mission and learner, marks unmarked sample, improves sample labeling ability.Bousmalis etc. propose a kind of based on consistance and nonconforming spontaneous non-semantic multiple mode model, and sets of video data has good performance.Jacob Andreas etc., for a statement consistance and nonconforming corpus, propose a kind of coding response method and are applied to statement annotation.Christoudias utilizes the inconsistency of conditional entropy to propose a kind of multi views learning method, classifies to video data.Guangxia Li etc. proposes two novel view transductive SVM methods, makes full use of a large amount of unmarked samples to improve classifier performance.As can be seen from above-mentioned document, researcher pays much attention to the marking convention of the unmarked sample of multi-categorizer, the research majority of the rule marked with ambiguousness are marked all from considering in a certain respect, when indivedual researcher's unification is considered also under special applications for multi-categorizer consistance.Therefore the marking convention of different situations is considered in research comprehensively, will play an important role to the performance improving multi views multi-categorizer.
Summary of the invention
This chapter is for Problems existing in standard C o-training algorithm, primary study Unlabeled data consistance and nonuniformity labeling method, the semi-supervised Cooperative Study algorithm of a kind of multi views based on composite marking rule (A Semi-supervisedCollaboration Classification Algorithm with the combination of Agreement and Disagreement labelrules is called for short Co-AgDiag algorithm) is proposed.This algorithm, first by view publishing, ensure that each view is independent, and increase judges link to the degree of confidence of sample labeling further, utilizes composite marking rule to sample labeling.In addition by the assessment of sorter otherness, further research and improvement have been done to model modification strategy, thus improved the performance of sorter model.The feasibility of the last algorithm herein by experimental verification and validity.The concrete algorithm flow chart of the present invention as shown in Figure 1.
Technical scheme provided by the invention comprises the steps:
1. the multi views based on composite marking rule works in coordination with a semisupervised classification algorithm, it is characterized in that, described method comprises the following steps:
(1) raw data set is divided into: originally mark small data set L; Original unmarked large data sets U; Test data set T:
(2) view publishing is carried out: raw data set L and U generates two views by view publishing, arrange from big to small by the association relationship between each attribute and class label, attribute is averaged segmentation, makes the mutual information sum between two parts attribute and class label as far as possible close; Marker samples collection L generates view 1 marker samples collection L
1={ (x
1.1, y
1), (x
2.1, y
2) ..., ((x
m.1, y
m)) and view 2 marker samples collection sample L
2={ (x
1.2, y
1), (x
2.2, y
2) ..., ((x
m.2, y
m)); Unmarked sample set U generates the unmarked sample set U of view 1
1={ (x
1.1, x
2.1..., x
n.1) and view 2 unmarked sample set sample U
2={ (x
1.2, x
2.2..., x
n.2); Wherein, the length of m and n difference representative sample collection L and U, each sample x
i(i=1,2 ..., m, m+1 ..., n) by the x of feature set 1
i.1with the x of feature set 2
i.2replace, y
i(i=1,2 ..., m, m+1 ..., n) representative sample classification;
(3) from original Unlabeled data collection U, the random individual unmarked sample of uSize that takes out defines sample set u;
(4) respectively by the data set L of two views
1, L
2study is to two sorter H
1and H
2, and utilize two sorters respectively to data set U
1and U
2classify, select the sample that degree of confidence is high;
(5) iterative loop is until reach update condition:
1) H
1and H
2respectively unmarked sample u is marked:
Judge H
1(x
i.1) whether equal H
2(x
i.2), if equal, then u=u-{x
i, L=L ∪ { (x
i, H
1(x
i.1)); If unequal, make the following judgment: when
time, then u=u-{x
i, L
1=L
1∪ { (x
i, 1,h
1(x
i.1)), L
2=L
2∪ { (x
i, 2, H
1(x
i.1)), L=L ∪ { (x
i, H
1(x
i.1)); When
time, then u=u-{x
i, L
1=L
1∪ { (x
i, 1, H
2(x
i.2)), L
2=L
2∪ { (x
i, 2, H
1(x
i.1)), L=L ∪ { (x
i, H
2(x
i.2)); Wherein, f (x
i) be classifier confidence function, θ is confidence threshold value;
2) new marker samples collection is utilized to upgrade sorter model;
3) otherness between sorter is calculated;
4) satisfied following model modification condition is judged whether,
Condition 1: sorter H
1and H
2error rate no longer reduce;
Condition 2: unmarked sample all marks;
Otherness between condition 3: two sorter meets certain threshold value;
As long as 3 conditions meet one, then enter step 1), carry out next round iteration, otherwise skip to step (6);
(6) output category device model;
(7) sorter model is utilized to classify to test sample book collection T.
In algorithm, view publishing is the major issue needing to discuss.The basic thought of view publishing is correlativity (degree of share of attributive character information) between any two attributes of analytical calculation view, and what correlativity was strong belongs to a view, otherwise then adheres to different views separately.
What mutual information was portrayed is quantity of information total between two stochastic variables, and this value is larger, illustrates that the degree of correlation between Two Variables is higher.If the mutual information of Two Variables is zero, then illustrate that Two Variables is completely incoherent.
Mutual information between X and Y is one measurement X (or Y) being included in the information in Y (or X), is defined as follows:
Probability is wherein estimated by corresponding histogram, that is:
Wherein N is the total quantity of sample, N (x
i) represent x
ithe number of times occurred.
A data set comprising two views, the attribute comprised in each sample adheres to two views separately.According to definition and the concept of view, the attribute in each view sample the gross information content that can provide for respective labels be roughly the same because each view can both describe practical judgment.The mutual information of data attribute in original view between category attribute is estimated in conjunction with Bootstrap and histogrammic method by one.Calculate the mutual information between each attribute and category attribute, can be understood as this attribute the quantity of information size that can provide for class label, two parts substantially identical sized by finally attribute being divided, and make its quantity of information summation identical.Mutual information (quantity of information that attribute provides for category attribute) sum in each view between each attribute and category attribute should be the minimal information amount required for determination of this category attribute.
Suppose to comprise 10 marker samples in original view, then each data attribute can represent with the vector of one 10 dimension, and category attribute is also like this, with Z=[Z
1, Z
2..., Z
10] form represent.Method is as follows:
(1) obtain the vector representation of data attribute X and category attribute Y, be respectively X=[X
1, X
2, X
n] and Y=[Y
1, Y2 ..., Y
n],
Wherein N is original tally sample size.
(2) to n=1,2 ..., m
1) X and Y is carried out to the Bootstrap sampling of corresponding time point, obtain respectively
with
2) association relationship of two attributes is estimated by histogram frequency distribution diagram method
(3) the Mutual Information Estimation value obtained is arranged by increasing
obtain
(4) expect (1-α), fiducial interval is
wherein
(5) estimated value of mutual information is finally obtained
for the mean value of all association relationship in fiducial interval.
Wherein need to carry out some explanation following:
(1) due to be solved be the multi views classification problem in semi-supervised learning field, raw data is concentrated and is only comprised a small amount of marker samples, and carry out view publishing according to marker samples in advance, therefore respectively two attributes are processed as discrete variable when calculating the mutual information between certain attribute and category attribute.
(2) when finally view being split according to mutual information, take and on average split and the principle making both mutual informations as far as possible close, to simplify procedures, it may not be best dividing method, but owing to being the basic definition based on multi views, therefore do not have obvious negative effect.
(3) view publishing obtains the method that two views mainly wish to use when meeting the condition of multi views multi views classification, may can not meet the condition of abundant redundancy completely, but the assessment carrying out otherness due to sorter generated upon splitting judges (sorter otherness appraisal procedure), therefore, it is possible to address this problem well.
Uneven to Unlabeled data distribution, in the process of mark, if sorter model not yet reaches good accuracy rate, then likely the mark made mistake is done to unmarked sample, use composite marking strategy can improve classification performance to greatest extent.As shown in Figure 2, its sample labeling rule is as follows: after the sorter model of each view has been carried out classify for its unmarked sample, if two sorter models are consistent to its mark result, then determine its mark result, if two sorter models are different to its mark result, and both degree of confidence gaps have exceeded the threshold value preset, then using the classification results of high confidence level sample as its classification.
Accompanying drawing explanation
Fig. 1 Co-AgDiag algorithm flow chart
Fig. 2 rule of combination flow process
The semi-supervised synergetic classification model of Fig. 3 bridge structure health multi views
Fig. 4 emerging bridge strain sample set category distribution
The wide bridge strain in Fig. 5 capital sample set category distribution
Embodiment
1. experimental data source
Utilize the coorinated training algorithm of composite marking rule, the semi-supervised synergetic classification model of the bridge structure health multi views that the present invention sets up as shown in Figure 3 this model mainly comprises primitive bridge structured data and imports data prediction coorinated training model prediction and result and export data prediction part mainly forms standard sample set by preprocess method, mainly comprise the attribute arranging and quantize each parameter, utilize the redundancy feature in feature selection approach removal bridge structure health data, noise data is processed, Data distribution8 analysis etc.
In the training process, first selection sort device, the multi views that the bridge structure set of data samples marked is divided into abundant redundancy is carried out training classifier, between multi-categorizer, Cooperative Study realizes being marked in forecasting process to Unlabeled data, selects the sorter that in training process, classification performance is good to carry out test also Output rusults to test data.
Bridge structure parameter mainly comprises static parameter and dynamic parameter, reflect dynamic parameter raw data that bridge health situation and operation state collect mainly from a certain moment, the continually varying data sequence arrived with certain frequency collection, namely vibration signal, static parameter raw data mainly comprise the bridge structure health characteristic attribute collection such as measuring point monitoring time, monitoring location, environment temperature, parameter value and comprise connection attribute (as temperature, strain value etc.) and Category Attributes (position attribution, time interval attribute etc.).
This experimental data comprises the Monitoring Data in emerging bridge capital, Beijing Guang Qiao (writing a Chinese character in simplified form Bri1 and Bri2), carry out in experiments experiment using the data that measuring point collects between 11, on February 6 zero point to 2011 on November 19th, 2010 as sample set, the parameter that emerging bridge uses is for strain and tilt, the parameter that the wide bridge in capital uses is strain and sedimentation, data sample classification is divided into 4 classes, normal respectively, early warning, report to the police, extreme value, sample size between different classes of differs greatly, Fig. 4 and Fig. 5 illustrates for strain parameter UCI data set spect and wdbc (writing a Chinese character in simplified form UCI1 and UCI2) that bridge structure sample set feature also have selected two standards in addition and proves that the efficacy data feature of algorithm is as table 1 further, shown in 2:
Table 1UCI data set features statistics
Table Bridge 2 beam data set features statistics
2. Setup Experiments
(1) sample distributes
In an experiment, for each data set, the data of Stochastic choice 75% are as training sample set, and the data of residue 25% are as test sample book collection
(2) validity of the coorinated training algorithm in order to verify composite marking rule selected by sorter, the coorinated training algorithm that have selected standard is identical as the learning algorithm of two sorters in the training of comparison algorithm standard in combination, therefore have selected two sorter learning algorithms in the coorinated training algorithm of radial basis function (Radial Basis Function, RBF) sorter and J48 decision tree composite marking rule in an experiment respectively and have selected RBF sorter and the combination of J48 decision tree
(3) evaluation index adopts the classification error rate of test set as algorithm evaluation index:
R=Nerr/N × 100% (formula 11)
Wherein: R refers to algorithm classification error rate, Nerr refers to the number of samples of classification error, and N refers to total number of samples.
3. experimental result and analysis
In order to prove the validity of innovatory algorithm in bridge structure health classification, in an experiment the synergetic of classification experiments standard and the coorinated training algorithm of the composite marking rule classification results in not marker samples quantity situation is carried out under different marker samples quantity as shown in Table 3, 4 to multiple sorting algorithm:
Under table 3 varying number marker samples, the classification error rate of Bri1 and Bri2 compares
The classification error rate of UCI1 and UCI2 under table 4 varying number marker samples
Experimental result shows, when unmarked sample size is different, the general substandard coorinated training algorithm of error rate of the coorinated training algorithm of composite marking rule increases sorter otherness and is conducive to strengthening the synergy between synergetic classification device, the consistency of performance of further raising algorithm and nonuniformity composite marking rule can utilize the implicit information in Unlabeled data, reduce marked erroneous, the coorinated training algorithm of error rate composite marking rule and the comparative result of standard in combination algorithm that reduce classification can be found out, algorithm has comparatively superior performance herein, no matter be for bridge structure data set, or the UCI data set of standard, the coorinated training algorithm of composite marking rule all has lower classification error rate, obviously be better than other learning method, administering and maintaining of bridge structure health can be instructed well.
Claims (1)
1. the multi views based on composite marking rule works in coordination with a semisupervised classification algorithm, it is characterized in that, described method comprises the following steps:
(1) raw data set is divided into: originally mark small data set L; Original unmarked large data sets U; Test data set T:
(2) view publishing is carried out: raw data set L and U generates two views by view publishing, arrange from big to small by the association relationship between each attribute and class label, attribute is averaged segmentation, makes the mutual information sum between two parts attribute and class label as far as possible close; Marker samples collection L generates view 1 marker samples collection L
1={ (x
1.1, y
1), (x
2.1, y
2) ..., ((x
m.1, y
m)) and view 2 marker samples collection sample L
2={ (x
1.2, y
1), (x
2.2, y
2) ..., ((x
m.2, y
m)); Unmarked sample set U generates the unmarked sample set U of view 1
1={ (x
1.1, x
2.1..., x
n.1) and view 2 unmarked sample set sample U
2={ (x
1.2, x
2.2..., x
n.2); Wherein, the length of m and n difference representative sample collection L and U, each sample x
i(i=1,2 ..., m, m+1 ..., n) by the x of feature set 1
i.1with the x of feature set 2
i.2replace, y
i(i=1,2 ..., m, m+1 ..., n) representative sample classification;
(3) from original Unlabeled data collection U, the random individual unmarked sample of uSize that takes out defines sample set u;
(4) respectively by the data set L of two views
1, L
2study is to two sorter H
1and H
2, and utilize two sorters respectively to data set U
1and U
2classify, select the sample that degree of confidence is high;
(5) iterative loop is until reach update condition:
1) H
1and H
2respectively unmarked sample u is marked:
Judge H
1(x
i.1) whether equal H
2(x
i.2), if equal, then u=u-{x
i, L=L ∪ { (x
i, H
1(x
i.1)); If unequal, make the following judgment: when
time, then u=u-{x
i, L
1=L
1∪ { (x
i, 1, H
1(x
i.1)), L
2=L
2∪ { (x
i, 2, H
1(x
i.1)), L=L ∪ { (x
i, H
1(x
i.1)); When
time, then u=u-{x
i, L
1=L
1∪ { (x
i, 1, H
2(x
i.2)), L
2=L
2∪ { (x
i, 2, H
1(x
i.1)), L=L ∪ { (x
i, H
2(x
i.2)); Wherein, f (x
i) be classifier confidence function, θ is confidence threshold value;
2) new marker samples collection is utilized to upgrade sorter model;
3) otherness between sorter is calculated;
4) satisfied following model modification condition is judged whether,
Condition 1: sorter H
1and H
2error rate no longer reduce;
Condition 2: unmarked sample all marks;
Otherness between condition 3: two sorter meets certain threshold value;
As long as 3 conditions meet one, then enter step 1), carry out next round iteration, otherwise skip to step (6);
(6) output category device model;
(7) sorter model is utilized to classify to test sample book collection T.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410742319.1A CN104463208A (en) | 2014-12-09 | 2014-12-09 | Multi-view semi-supervised collaboration classification algorithm with combination of agreement and disagreement label rules |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410742319.1A CN104463208A (en) | 2014-12-09 | 2014-12-09 | Multi-view semi-supervised collaboration classification algorithm with combination of agreement and disagreement label rules |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104463208A true CN104463208A (en) | 2015-03-25 |
Family
ID=52909222
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410742319.1A Pending CN104463208A (en) | 2014-12-09 | 2014-12-09 | Multi-view semi-supervised collaboration classification algorithm with combination of agreement and disagreement label rules |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104463208A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104809468A (en) * | 2015-04-20 | 2015-07-29 | 东南大学 | Multi-view classification method based on indefinite kernels |
| CN106909540A (en) * | 2015-12-23 | 2017-06-30 | 神州数码信息系统有限公司 | A kind of smart city citizen's preference discovery technique based on Cooperative Study |
| CN107463996A (en) * | 2017-06-05 | 2017-12-12 | 西安交通大学 | From step coorinated training learning method |
| CN108021985A (en) * | 2016-11-03 | 2018-05-11 | 北京智能管家科技有限公司 | A kind of model parameter training method and device |
| CN108597602A (en) * | 2018-04-28 | 2018-09-28 | 中南大学 | A kind of label correction method towards dermatology data |
| CN109241770A (en) * | 2018-08-10 | 2019-01-18 | 深圳前海微众银行股份有限公司 | Information value calculating method, equipment and readable storage medium storing program for executing based on homomorphic cryptography |
| CN109325357A (en) * | 2018-08-10 | 2019-02-12 | 深圳前海微众银行股份有限公司 | Information value calculating method, equipment and readable storage medium storing program for executing based on RSA |
| CN110254438A (en) * | 2018-03-12 | 2019-09-20 | 松下知识产权经营株式会社 | Information processing unit and program recorded medium |
| CN110363415A (en) * | 2019-06-29 | 2019-10-22 | 上海淇馥信息技术有限公司 | The method and apparatus of fraud label based on multiple view study |
| CN110431543A (en) * | 2017-02-28 | 2019-11-08 | 弗劳恩霍夫应用研究促进协会 | The method and classification processor of classification information |
| CN110458245A (en) * | 2019-08-20 | 2019-11-15 | 图谱未来(南京)人工智能研究院有限公司 | A kind of multi-tag disaggregated model training method, data processing method and device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101902744A (en) * | 2010-07-28 | 2010-12-01 | 南京航空航天大学 | Intrusion detection system of wireless sensor network based on sniffer |
| CN103345922A (en) * | 2013-07-05 | 2013-10-09 | 张巍 | Large-length voice full-automatic segmentation method |
| CN103679677A (en) * | 2013-12-12 | 2014-03-26 | 杭州电子科技大学 | Dual-model image decision fusion tracking method based on mutual updating of models |
-
2014
- 2014-12-09 CN CN201410742319.1A patent/CN104463208A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101902744A (en) * | 2010-07-28 | 2010-12-01 | 南京航空航天大学 | Intrusion detection system of wireless sensor network based on sniffer |
| CN103345922A (en) * | 2013-07-05 | 2013-10-09 | 张巍 | Large-length voice full-automatic segmentation method |
| CN103679677A (en) * | 2013-12-12 | 2014-03-26 | 杭州电子科技大学 | Dual-model image decision fusion tracking method based on mutual updating of models |
Non-Patent Citations (1)
| Title |
|---|
| 于重重等: "组合标记的多视图半监督协同分类算法", 《计算机应用》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104809468A (en) * | 2015-04-20 | 2015-07-29 | 东南大学 | Multi-view classification method based on indefinite kernels |
| CN106909540A (en) * | 2015-12-23 | 2017-06-30 | 神州数码信息系统有限公司 | A kind of smart city citizen's preference discovery technique based on Cooperative Study |
| CN108021985B (en) * | 2016-11-03 | 2022-05-10 | 北京如布科技有限公司 | Model parameter training method and device |
| CN108021985A (en) * | 2016-11-03 | 2018-05-11 | 北京智能管家科技有限公司 | A kind of model parameter training method and device |
| CN110431543A (en) * | 2017-02-28 | 2019-11-08 | 弗劳恩霍夫应用研究促进协会 | The method and classification processor of classification information |
| CN110431543B (en) * | 2017-02-28 | 2024-03-15 | 弗劳恩霍夫应用研究促进协会 | Method for classifying information and classification processor |
| CN107463996A (en) * | 2017-06-05 | 2017-12-12 | 西安交通大学 | From step coorinated training learning method |
| CN110254438A (en) * | 2018-03-12 | 2019-09-20 | 松下知识产权经营株式会社 | Information processing unit and program recorded medium |
| CN108597602A (en) * | 2018-04-28 | 2018-09-28 | 中南大学 | A kind of label correction method towards dermatology data |
| CN109241770A (en) * | 2018-08-10 | 2019-01-18 | 深圳前海微众银行股份有限公司 | Information value calculating method, equipment and readable storage medium storing program for executing based on homomorphic cryptography |
| CN109325357A (en) * | 2018-08-10 | 2019-02-12 | 深圳前海微众银行股份有限公司 | Information value calculating method, equipment and readable storage medium storing program for executing based on RSA |
| CN109241770B (en) * | 2018-08-10 | 2021-11-09 | 深圳前海微众银行股份有限公司 | Information value calculation method and device based on homomorphic encryption and readable storage medium |
| CN109325357B (en) * | 2018-08-10 | 2021-12-14 | 深圳前海微众银行股份有限公司 | RSA-based information value calculation method, device and readable storage medium |
| CN110363415A (en) * | 2019-06-29 | 2019-10-22 | 上海淇馥信息技术有限公司 | The method and apparatus of fraud label based on multiple view study |
| CN110458245A (en) * | 2019-08-20 | 2019-11-15 | 图谱未来(南京)人工智能研究院有限公司 | A kind of multi-tag disaggregated model training method, data processing method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104463208A (en) | Multi-view semi-supervised collaboration classification algorithm with combination of agreement and disagreement label rules | |
| CN105824802A (en) | Method and device for acquiring knowledge graph vectoring expression | |
| CN109615014A (en) | A kind of data sorting system and method based on the optimization of KL divergence | |
| CN103984959A (en) | Data-driven and task-driven image classification method | |
| CN108470022B (en) | Intelligent work order quality inspection method based on operation and maintenance management | |
| CN105389583A (en) | Image classifier generation method, and image classification method and device | |
| CN106682696A (en) | Multi-example detection network based on refining of online example classifier and training method thereof | |
| CN104834940A (en) | Medical image inspection disease classification method based on support vector machine (SVM) | |
| CN102842043B (en) | Particle swarm classifying method based on automatic clustering | |
| CN103136540B (en) | A kind of Activity recognition method based on implicit structure reasoning | |
| CN106778878A (en) | A kind of character relation sorting technique and device | |
| CN104156945A (en) | Method for segmenting gray scale image based on multi-objective particle swarm optimization algorithm | |
| CN106156315A (en) | A kind of data quality monitoring method judged based on disaggregated model | |
| CN107679550A (en) | A kind of appraisal procedure of data set classification availability | |
| CN113362299B (en) | X-ray security inspection image detection method based on improved YOLOv4 | |
| CN106203296A (en) | The video actions recognition methods of one attribute auxiliary | |
| CN107273505A (en) | Supervision cross-module state Hash search method based on nonparametric Bayes model | |
| CN106228027A (en) | A kind of semi-supervised feature selection approach of various visual angles data | |
| CN113901229A (en) | Syntactic graph convolution-based Chinese-Yue bilingual news event causal relationship extraction method | |
| CN109919236A (en) | A kind of BP neural network multi-tag classification method based on label correlation | |
| CN113554100A (en) | Web service classification method for enhancing attention network of special composition picture | |
| Kenyon-Dean et al. | Clustering-oriented representation learning with attractive-repulsive loss | |
| CN103123685B (en) | Text mode recognition method | |
| CN113704082A (en) | Model evaluation method and device, electronic equipment and storage medium | |
| CN110111365B (en) | Training method and device based on deep learning and target tracking method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20150325 |
|
| WD01 | Invention patent application deemed withdrawn after publication |