CN105046236A - Iterative tag noise recognition algorithm based on multiple voting - Google Patents
Iterative tag noise recognition algorithm based on multiple voting Download PDFInfo
- Publication number
- CN105046236A CN105046236A CN201510490699.9A CN201510490699A CN105046236A CN 105046236 A CN105046236 A CN 105046236A CN 201510490699 A CN201510490699 A CN 201510490699A CN 105046236 A CN105046236 A CN 105046236A
- Authority
- CN
- China
- Prior art keywords
- noise
- sample
- voting
- iteration
- value
- 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
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Complex Calculations (AREA)
Abstract
The invention discloses an iterative tag noise recognition algorithm based on multiple voting, and belongs to the field of machine learning and data mining. The recognition algorithm is based on integrated learning thoughts; a multiple voting mode is adopted; the noise is judged through the voting of a plurality of classifiers; before the voting in each time, the sequence of samples to be detected is randomly disturbed, so that a group of noise recognition results with possible differences can be generated during the voting in each time; and finally, multiple groups of results generated by multiple voting are integrated to generate the final noise recognition result. The recognition algorithm also adopts an iterative mode, and the samples to be detected input during iteration in each time are residual samples subjected to noise filtering during the iteration in the last time. Compared with the conventional single voting mode, a multiple voting method has the advantages that flexibility and accuracy are higher; the results of the single voting result can be summarized in the other level; and the requirements of different types of data and noise ratios can be met. The iterative recognition method can be used for comprehensively and thoroughly recognize all noise data.
Description
Technical field
The present invention relates to data mining and machine learning techniques field, specifically based on the iterative label noise identification algorithm of repeatedly voting.
Background technology
A lot of training datas that machine learning uses in the middle of practical application are all that band is noisy, and the reason wherein caused comprises artificial mistake, the mistake of hardware device, the mistake etc. of data-gathering process.Traditional way is exactly before those machine learning algorithms of application, by manually carrying out data prediction work to source data, obtain pure source data, but, these artificial job difficult, loaded down with trivial details, consuming time, and the total correctness of data can not be ensured, this causes very important impact to follow-up algorithm application.Data noise generally includes two classes: attribute noise and classification noise, and it is inaccurate that attribute noise refers to sample attribute value, and classification noise refers to the label inaccurate [1] of sample.Compared to attribute noise, the impact of classification noise is larger.
The disposal route of classification noise is comprised: the algorithm [2,3] that design is healthy and strong and noise detection algorithm [4,5,6,7].The algorithm of design stalwartness mainly improves existing algorithm, makes existing algorithm less by the impact of classification noise.And noise detection algorithm detected and erased noise before use comprises the data of noise.By contrast, the effect of noise like detection algorithm and versatility stronger.
Existing noise like detection algorithm mainly comprises two classes: based on k neighbour [4] with based on integrated study [5,6,7].Basic thought based on k neighbour is the class label comparing a sample and its neighbours' sample, if these labels are obviously inconsistent, then thinks that this sample label is noise.This method affects by the limitation of k nearest neighbor algorithm, and not all Data distribution8 is all applicable to the method based on k next-door neighbour.By contrast, the algorithm based on integrated study uses more extensive.The representative of this class algorithm is that great majority filter and consistance filters [7].In these algorithms, first training data is divided into multiple subset by random, and then each subset can be carried out walkaway by independent.The basic thought detected is the ballot by taking residuary subset as multiple sorters that training sample obtains.This kind of algorithm mainly comprises two steps: sample divides and multi-categorizer ballot.Because sample divides and multi-categorizer ballot Exactly-once, therefore belong to the label noise detecting method based on single ballot.
There are two deficiencies in the existing label noise detecting method based on single ballot: the impact that the result of single ballot divides by sample is comparatively large, and the possibility of omitting noise is larger.
List of references:
[1]Zhu,Xingquan,andXindongWu."Classnoisevs.attributenoise:Aquantitativestudy."ArtificialIntelligenceReview22.3(2004):177-210.
[2]J.Bootkrajang,A.Kaban,Classificationofmislabelledmicroarraysusingrobustsparselogisticregression,Bioinformatics29(7)(2013)870–877.
[3]J.Saez,M.Galar,J.Luengo,F.Herrera,Afirststudyondecompositionstrategieswithdatawithclassnoiseusingdecisiontrees,in:HybridArtificialIntelligentSystems,LectureNotesinComputerScience,vol.7209,2012,pp.25–35.
[4]D.L.Wilson,Asymptoticpropertiesofnearestneighborrulesusingediteddata,IEEETrans.Syst.ManCybernet.2(3)(1992)431–433.
[5]J.Young,J.Ashburner,S.Ourselin,Wrappermethodstocorrectmislabeledtrainingdata,in:3rdInternationalWorkshoponPatternRecognitioninNeuroimaging,2013,pp.170–173.
[6]D.Guan,W.Yuan,etal.,Identifyingmislabeledtrainingdatawiththeaidofunlabeleddata,Appl.Intell.35(3)(2011)345–358.
[7]C.E.Brodley,M.A.Friedl,Identifyingmislabeledtrainingdata,J.Artif.Intell.Res.11(1999)131–167.
Summary of the invention
The problem to be solved in the present invention is to provide a kind of iterative label noise identification algorithm based on repeatedly voting, this algorithm the method adopts the mode of repeatedly voting, repeatedly ballot and single vote link corresponding parameter and strategy can be set according to real data collection situation, the problem that the impact that the result avoiding single ballot is divided by sample is larger, effectively can improve the accuracy rate of identification, by the mode of iteration, noise data can be found more thoroughly.
As the iterative label noise identification algorithm based on repeatedly voting disclosed by the invention, comprise the following steps:
Step 1) determine algorithm input variable, comprise pending sample set D, maximum iteration time maxIter, repeatedly vote frequency n umVote, minimum ballot frequency n umFinalPass needed for final noise identification, random block count numCross, single ballot sorter frequency n umClassifier, minimum ballot frequency n umPass needed for single noise identification, initialization is repeatedly voted iterations t=1, peripheral iterations m=1, the pending sample set E=D of initialization;
Step 2) E is divided into numCross subset of the same size at random
initiation parameter i=1;
Step 3) use
in set, sample does training data, selects the sorting algorithm that numClassifier different, the sorter H that training numClassifier is different
1, H
2..., H
numClassifier;
Step 4) use H
1, H
2..., H
numClassifierto sample set
middle sample classification, adds up each sample by the frequency n umWrong of mis-classification, if numWrong is more than or equal to specify threshold value numPass, then this sample is classified as suspicious noise by this ballot;
Step 5) iteration execution step 2) to 4), after each iteration, i value adds 1, until i value equals numCross, stops iteration, generates suspicious noise set;
Step 6) iteration execution step 2) to 5), after each iteration, t value adds 1, until t=numVote, generates numVote suspicious noise set;
Step 7) comprehensive analysis numVote suspicious noise set, if the frequency n umExist of sample appearance in numVote set is more than or equal to specify threshold value numFinalPass, then basis repeatedly voting results, assert that this sample is noise, if based on the m time iteration, the noise set of generation is
Step 8)
iteration performs step 2) to 7), after each iteration, m value adds 1, until
or till m=maxIter;
Step 9) return E value, E is the purified sample collection after erased noise, and algorithm terminates.
Further, described step 3) in, numClassifier is chosen to be odd number, selects odd number to be conducive to the realization of voting.Sorting algorithm is k next-door neighbour, decision tree, Bayes, neural network, one or more in Support Vector Machine.And the selected of numClassifier affects by data set.During Small Sample Database collection, for ensureing otherness between multi-categorizer, larger numClassifier value should be taked.When sample set label noise is higher, larger numClassifier value also should be taked.Namely larger numClassifier can ensure the high label noise identification rate of each iteration, contributes to again reducing iterations, improves efficiency of algorithm.On the other hand, when sample set number is comparatively large and sample label noise ratio is lower, less numClassifier can be selected.As can numClassifier=3 be arranged.
Another kind of improvement, described step 4) in, described numPass value is chosen to be numClassifier/2 or numClassifier.It is larger that numPass value is arranged, and detects stricter.Accordingly, detect stricter, holding data, to be used as the possibility of noise less, label noise when the possibility that perform data is larger.
Another kind of to improve, described step 7) in numFinalPass value can select some conventional value, as numVote/2 or numVote.Also can by independently verifying sample, the numFinalPass numerical value of calculation optimization.Concrete steps comprise: noise ratio a) estimating pending noise data according to priori, b) in verification sample, random noise is added, c) traversal institute likely numFinalPass numerical value under calculating this numerical value, this algorithm is to the recognition accuracy of noise in verification sample, d) selection has the numFinalPass of the highest recognition accuracy.It is larger that numVote value is arranged, and detects stricter, and accordingly, holding data, to be used as the possibility of noise less, and the possibility that label noise ought be performed data is larger.NumFinalPass value should be echoed with numPass herein, if numPass is too small, then numFinalPass should strengthen, in order to avoid too much good sample is taken as noise, in like manner, if numPass value is excessive, then numFinalPass should reduce, in order to avoid too much noise sample has been taken as sample.
The invention has the beneficial effects as follows: the iterative label noise identification algorithm based on repeatedly voting of the present invention adopts mode of repeatedly voting to carry out noise identification, before each ballot, upset sample order at random, therefore ensure that the otherness of ballot, compare with traditional single ballot mode, repeatedly voting method has more dirigibility and accuracy, single temporal voting strategy is crossed often and is sent or tension, and repeatedly vote to do single voting results in another aspect and gather, therefore, it is possible to meet the requirement of different types of data and noise ratio.In addition, additionally use iterative recognition methods in recognizer, the sample to be detected inputted during each iteration, the purified sample exported for filter out noise during last iteration, can identify institute's noise data more comprehensively thoroughly.Recognizer of the present invention solves the not high problem of existing label noise identification algorithm recognition accuracy, ensure that the pin-point accuracy of noise identification.
Accompanying drawing explanation
Fig. 1 is the process flow diagram that the present invention is based on the iterative label noise identification calculation of repeatedly voting.
Embodiment
Below in conjunction with accompanying drawing, a kind of iterative label noise identification algorithm based on repeatedly voting that the present invention proposes is described in detail.
As shown in Figure 1, the iterative label noise identification algorithm based on repeatedly voting of the present invention, comprises the following steps:
Step 1) determine algorithm input variable, comprise pending sample set D, maximum iteration time maxIter, repeatedly vote frequency n umVote, minimum ballot frequency n umFinalPass needed for final noise identification, random block count numCross, single ballot sorter frequency n umClassifier, minimum ballot frequency n umPass needed for single noise identification, initialization is repeatedly voted iterations t=1, peripheral iterations m=1, the pending sample set E=D of initialization;
Step 2) E is divided into numCross subset of the same size at random
initiation parameter i=1;
Step 3) use
in set, sample does training data, selects the sorting algorithm that numClassifier different, the sorter H that training numClassifier is different
1, H
2..., H
numClassifier; NumClassifier is chosen to be odd number, such as 3,5,7 etc., certainly, be not limited to these cited odd numbers; Sorting algorithm is k next-door neighbour, decision tree, Bayes, neural network, one or more in Support Vector Machine.
Step 4) use H
1, H
2..., H
numClassifierto sample set
middle sample classification, adds up each sample by the frequency n umWrong of mis-classification, if numWrong is more than or equal to specify threshold value numPass, then this sample is classified as suspicious noise by this ballot.It is larger that numPass value is arranged, and detects stricter, and accordingly, holding data, to be used as the possibility of noise less, and the possibility that label noise ought be performed data is larger.Therefore, numPass value is preferably numClassifier/2 or numClassifier, and this is preferable examples, and other suitable numerical value all can be used as a kind of selection.
Step 5) iteration execution step 2) to 4), after each iteration, i value adds 1, until i value equals numCross, stops iteration, generates suspicious noise set;
Step 6) iteration execution step 2) to 5), after each iteration, t value adds 1, until t=numVote, generates numVote suspicious noise set;
Step 7) comprehensive analysis numVote suspicious noise set, if the frequency n umExist of sample appearance in numVote set is more than or equal to specify threshold value numFinalPass, then basis repeatedly voting results, assert that this sample is noise, if based on the m time iteration, the noise set of generation is
nalPass value be preferably for numVote/2 or numVote, numVote value arrange larger, detect stricter, accordingly, holding data, to be used as the possibility of noise less, and the possibility that label noise ought be performed data is larger.NumFinalPass value should be echoed with numPass herein, if numPass is too small, then numFinalPass should strengthen, in order to avoid too much good sample is taken as noise, in like manner, if numPass value is excessive, then numFinalPass should reduce, in order to avoid too much noise sample has been taken as sample.
Step 8)
iteration performs step 2) to 7), after each iteration, m value adds 1, until
or till m=maxIter;
Step 9) return E value, E is the purified sample collection after erased noise, and algorithm terminates.
Below describe the present invention in detail to the test result of 2 groups of data in UCI database and the improvement comparing performance with other label noise recognizing methods.Identification in this paper compares with MajorityFiltering and ConsensusFiltering the most popular at present.Because data do not exist label noise in original UCI database, in the present embodiment, people, for adding noise, considers different noise ratios, comprises 10%, 20%, 30%, 40%.In this example, the marked erroneous number measurement by mistake of label noise detection algorithm performance.This error number comprises two parts, and a part is the diagnosis as well data of noise data mistake, represents with E1, another part be error in data be diagnosed as noise data, represent with E2.E1+E2 value is less, shows that algorithm accuracy rate is higher.
Table 1-data set
| Data set | Sample number | Characteristic number |
| breast | 699 | 9 |
| wdbc | 569 | 31 |
Middle optimum configurations is as follows: numCross=3, numClassifier=3 (three kinds of sorting algorithms comprise naive Bayesian, decision tree and arest neighbors), maxIter=100, numVote=5; It is numPass=2, numFinalPass=5 (being called IMFCF algorithm) that numPass and numFinalPass has two kinds to combine one, and another kind of combination is numPass=3, numFinalPass=3 (being called ICFMF algorithm).
Table 2-breast data set, result under 10% noise ratio
Table 3-breast data set, result under 20% noise ratio
Table 4-breast data set, result under 30% noise ratio
Table 5-breast data set, result under 40% noise ratio
Table 6-wdbc data set, result under 10% noise ratio
Table 7-wdbc data set, result under 20% noise ratio
Table 8-wdbc data set, result under 30% noise ratio
Table 9-wdbc data set, result under 40% noise ratio
More than show to show in 2-9, in two data of experiment, based on different noise ratios, the algorithmic stability that the present invention proposes be better than two kinds of traditional algorithms.
In sum, above embodiment only in order to technical scheme of the present invention to be described, is not intended to limit protection scope of the present invention.Within the spirit and principles in the present invention all, any amendment made, equivalent replacement, improvement etc., it all should be encompassed in the middle of right of the present invention.
Claims (6)
1., based on an iterative label noise identification algorithm of repeatedly voting, it is characterized in that, comprise the following steps:
Step 1) determine algorithm input variable, comprise pending sample set D, maximum iteration time maxIter, repeatedly vote frequency n umVote, minimum ballot frequency n umFinalPass needed for final noise identification, random block count numCross, single ballot sorter frequency n umClassifier, minimum ballot frequency n umPass needed for single noise identification, initialization is repeatedly voted iterations t=1, peripheral iterations m=1, the pending sample set E=D of initialization;
Step 2) E is divided into numCross subset of the same size at random
initiation parameter i=1;
Step 3) use
in set, sample does training data, selects the sorting algorithm that numClassifier different, the sorter H that training numClassifier is different
1, H
2..., H
numClassifier;
Step 4) use H
1, H
2..., H
numClassifierto sample set
middle sample classification, adds up each sample by the frequency n umWrong of mis-classification, if numWrong is more than or equal to specify threshold value numPass, then this sample is classified as suspicious noise by this ballot;
Step 5) iteration execution step 2) to 4), after each iteration, i value adds 1, until i value equals numCross, stops iteration, generates suspicious noise set;
Step 6) iteration execution step 2) to 5), after each iteration, t value adds 1, until t=numVote, generates numVote suspicious noise set;
Step 7) comprehensive analysis numVote suspicious noise set, if the frequency n umExist of sample appearance in numVote set is more than or equal to specify threshold value numFinalPass, then basis repeatedly voting results, assert that this sample is noise, if based on the m time iteration, the noise set of generation is
Step 8)
iteration performs step 2) to 7), after each iteration, m value adds 1, until
or till m=maxIter;
Step 9) return E value, E is the purified sample collection after erased noise, and algorithm terminates.
2. the iterative label noise identification algorithm based on repeatedly voting according to claim 1, is characterized in that: described step 3) in, numClassifier is chosen to be odd number.
3. the iterative label noise identification algorithm based on repeatedly voting according to claim 2, is characterized in that: arrange described numClassifier=3.
4. the iterative label noise identification algorithm based on repeatedly voting according to claim 1, is characterized in that: described step 4) in, described numPass value is chosen to be numClassifier/2 or numClassifier.
5. the iterative label noise identification algorithm based on repeatedly voting according to claim 1, is characterized in that: described step 7) selected numVote/2 or numVote of middle numFinalPass value.
6. the iterative label noise identification algorithm based on repeatedly voting according to claim 1, is characterized in that: described step 7) in numFinalPass value by independently verifying sample, calculation optimization; Concrete steps comprise: noise ratio a) estimating pending noise data according to priori, b) in verification sample, random noise is added, c) traversal institute likely numFinalPass numerical value under calculating this numerical value, this algorithm is to the recognition accuracy of noise in verification sample, d) selection has the numFinalPass of the highest recognition accuracy.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510490699.9A CN105046236A (en) | 2015-08-11 | 2015-08-11 | Iterative tag noise recognition algorithm based on multiple voting |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510490699.9A CN105046236A (en) | 2015-08-11 | 2015-08-11 | Iterative tag noise recognition algorithm based on multiple voting |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN105046236A true CN105046236A (en) | 2015-11-11 |
Family
ID=54452765
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510490699.9A Pending CN105046236A (en) | 2015-08-11 | 2015-08-11 | Iterative tag noise recognition algorithm based on multiple voting |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105046236A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107292330A (en) * | 2017-05-02 | 2017-10-24 | 南京航空航天大学 | A kind of iterative label Noise Identification algorithm based on supervised learning and semi-supervised learning double-point information |
| CN108509969A (en) * | 2017-09-06 | 2018-09-07 | 腾讯科技（深圳）有限公司 | Data mask method and terminal |
| CN110060247A (en) * | 2019-04-18 | 2019-07-26 | 深圳市深视创新科技有限公司 | Cope with the robust deep neural network learning method of sample marking error |
| CN110163376A (en) * | 2018-06-04 | 2019-08-23 | 腾讯科技（深圳）有限公司 | Sample testing method, the recognition methods of media object, device, terminal and medium |
| CN111352966A (en) * | 2020-02-24 | 2020-06-30 | 交通运输部水运科学研究所 | Data tag calibration method in autonomous navigation |
| CN112562730A (en) * | 2020-11-24 | 2021-03-26 | 北京华捷艾米科技有限公司 | Sound source analysis method and system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040083270A1 (en) * | 2002-10-23 | 2004-04-29 | David Heckerman | Method and system for identifying junk e-mail |
| CN101330476A (en) * | 2008-07-02 | 2008-12-24 | 北京大学 | Method for dynamically detecting junk mail |
-
2015
- 2015-08-11 CN CN201510490699.9A patent/CN105046236A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20040083270A1 (en) * | 2002-10-23 | 2004-04-29 | David Heckerman | Method and system for identifying junk e-mail |
| CN101330476A (en) * | 2008-07-02 | 2008-12-24 | 北京大学 | Method for dynamically detecting junk mail |
Non-Patent Citations (1)
| Title |
|---|
| DONGHAI GUAN等: "Class Noise Detection by Multiple Voting", 《2013 NINTH INTERNATIONAL CONFERENCE ON NATURAL COMPUTATION》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107292330A (en) * | 2017-05-02 | 2017-10-24 | 南京航空航天大学 | A kind of iterative label Noise Identification algorithm based on supervised learning and semi-supervised learning double-point information |
| CN107292330B (en) * | 2017-05-02 | 2021-08-06 | 南京航空航天大学 | Iterative label noise identification algorithm based on double information of supervised learning and semi-supervised learning |
| CN108509969A (en) * | 2017-09-06 | 2018-09-07 | 腾讯科技（深圳）有限公司 | Data mask method and terminal |
| CN110163376A (en) * | 2018-06-04 | 2019-08-23 | 腾讯科技（深圳）有限公司 | Sample testing method, the recognition methods of media object, device, terminal and medium |
| CN110163376B (en) * | 2018-06-04 | 2023-11-03 | 腾讯科技（深圳）有限公司 | Sample detection method, media object identification method, device, terminal and medium |
| CN110060247A (en) * | 2019-04-18 | 2019-07-26 | 深圳市深视创新科技有限公司 | Cope with the robust deep neural network learning method of sample marking error |
| CN111352966A (en) * | 2020-02-24 | 2020-06-30 | 交通运输部水运科学研究所 | Data tag calibration method in autonomous navigation |
| CN112562730A (en) * | 2020-11-24 | 2021-03-26 | 北京华捷艾米科技有限公司 | Sound source analysis method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105046236A (en) | Iterative tag noise recognition algorithm based on multiple voting | |
| CN107292330B (en) | Iterative label noise identification algorithm based on double information of supervised learning and semi-supervised learning | |
| Chung et al. | Slice finder: Automated data slicing for model validation | |
| Hassan et al. | Detecting prohibited items in X-ray images: A contour proposal learning approach | |
| CN110969166A (en) | Small target identification method and system in inspection scene | |
| CN107229942A (en) | A kind of convolutional neural networks rapid classification method based on multiple graders | |
| CN102324046A (en) | Four-classifier cooperative training method combining active learning | |
| CN102346829A (en) | Virus detection method based on ensemble classification | |
| CN107830996B (en) | Fault diagnosis method for aircraft control surface system | |
| CN102254193A (en) | Relevance vector machine-based multi-class data classifying method | |
| CN104657574B (en) | The method for building up and device of a kind of medical diagnosismode | |
| CN106326913A (en) | Money laundering account determination method and device | |
| CN117033912B (en) | Equipment fault prediction method and device, readable storage medium and electronic equipment | |
| CN103020643A (en) | Classification method based on kernel feature extraction early prediction multivariate time series category | |
| CN109255029A (en) | A method of automatic Bug report distribution is enhanced using weighted optimization training set | |
| CN108416373A (en) | A kind of unbalanced data categorizing system based on regularization Fisher threshold value selection strategies | |
| CN108509996A (en) | Feature selection approach based on Filter and Wrapper selection algorithms | |
| CN112307860A (en) | Image recognition model training method and device and image recognition method and device | |
| CN104615789A (en) | Data classifying method and device | |
| CN107480441B (en) | Modeling method and system for children septic shock prognosis prediction | |
| CN106951728B (en) | Tumor key gene identification method based on particle swarm optimization and scoring criterion | |
| CN114254146A (en) | Image data classification method, device and system | |
| CN102945238A (en) | Fuzzy ISODATA (interactive self-organizing data) based feature selection method | |
| CN111209939A (en) | SVM classification prediction method with intelligent parameter optimization module | |
| CN103810210B (en) | Search result display methods 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: 20151111 |
|
| WD01 | Invention patent application deemed withdrawn after publication |