CN107480690A - A kind of more sorting techniques for including unknown classification based on SVMs - Google Patents
A kind of more sorting techniques for including unknown classification based on SVMs Download PDFInfo
- Publication number
- CN107480690A CN107480690A CN201710536516.1A CN201710536516A CN107480690A CN 107480690 A CN107480690 A CN 107480690A CN 201710536516 A CN201710536516 A CN 201710536516A CN 107480690 A CN107480690 A CN 107480690A
- Authority
- CN
- China
- Prior art keywords
- classification
- sample
- svms
- sorting techniques
- belongs
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
-
- 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
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention provides a kind of more sorting techniques for including unknown classification based on SVMs, including:1) the i-th category classifier after each training is based respectively on, identifies whether sample to be identified belongs to the i-th classification;I-th category classifier is with the known sample structure positive sample collection for belonging to the i-th classification, negative sample collection is built with the known all samples for belonging to remaining classification of N 1, the two classification graders for identifying whether inputted sample belongs to the i-th classification are used for obtained from being trained based on SVM models;2) when by each two classification grader of sample to be identified input, when acquired results are no, the current sample to be identified of identification belongs to unknown classification in step 1);When in step 1), when being, then to assert, currently sample to be identified belongs to this two classification classified corresponding to grader to the output result of one and only one two classification grader.The present invention has the detectability of unknown classification, has higher recall ratio and precision ratio.
Description
Technical field
The present invention relates to machine learning field, and specifically, the present invention relates to a kind of including not based on SVMs
Know more sorting techniques of classification.
Background technology
SVMs (SVM) is used as a kind of typical machine learning algorithm, due to its complete theories integration, various
It is widely applied in classification problem.SVM assumes sample dataIt is generally not linear separability in luv space, can incites somebody to action
Sample data is mapped to new feature space (new feature Spatial General 6 R is more higher-dimension), sample data from luv spaceNew
The corresponding points of feature space areTwo classification SVM target can be described as finding in feature space using sample data
One linear hyperplaneWhereinIt is the normal vector of linear hyperplane, b is offset, if f
(xi) the then y of > 0i=+1 represents that the sample data belongs to positive classification, if f (xi) the then y of < 0i=-1 represents the sample data category
In anti-classification.Optimal linear hyperplane should cause sample data from the Maximizing Minimum Distance of the linear hyperplane, mathematics
Form is expressed asMeet constraints Wherein C is experience system
Number, ξiFor slack variable, to a small number of sample data relaxed constraints for being unsatisfactory for firm constraints condition, (cost is that object function is punished
Penalize).
For more classification SVM, appoint and take the other sample data of two species to be trained using aforesaid way, respectively obtain classification
Model (linear hyperplane), which classification is new samples data finally belong to is produced by the ballot of all disaggregated models.Such as N classification
SVM can be decomposed into N (N-1)/2 two classification SVM, if N (N-1)/2 two classification SVM chooses identical kernel function (feature
Space reflection function) and relevant parameter, equivalent to N (N-1)/2 linear hyperplane whole feature space is divided, such as
Shown in Fig. 1.
Current almost all of machine learning algorithm, its judged result is known class, such as digital 0-9's is hand-written
Which kind of data no matter input identification, input, and machine learning algorithm can be chosen one of as identification from this 10 numerals
As a result, for invalid stochastic inputs data, more preferably result is one unknown classification of output to represent invalid output.
In field of human-computer interaction, the problem of defect causes unobvious, for invalid recognition result, the mankind can simply differentiate simultaneously
Ignore the output.But in some full-automatic fields, due to the presence of exceptional condition, then must automatic decision input data whether just
Often or effectively.
For SVM, can solve the problem using following three kinds of modes at present.
First way is to obtain the sample data of unknown classification in the training stage, is as gathered for numeral identification various
The data arbitrarily inputted, such as wave (~~), the data that (√), fork (×) etc. are not belonging to digital 0-9 classifications are hooked, so as to
To comprising the other sample data of 11 species, then whole sample data set is trained to obtain disaggregated model using traditional SVM.This
The shortcomings that kind of mode is it will be evident that the class number of unknown classification is inherently infinite, therefore the sample data gathered can not
Covering is distributed.
The second way is the form using probability, and train to obtain is the disaggregated model containing probability, and forecast period is defeated
What is gone out is that each new samples data belong to various types of other probability rather than single category result.This mode can be solved partly
During preceding several maximum probability sizableness of the certainly problem, especially forecast period output, substantially it can be assumed that new samples data
It is unknown classification, because should be that the probable value for corresponding to the category is far longer than the general of other corresponding classifications when belonging to known class
Rate value.If a certain subclass of a certain known class and unknown classification with respect to " close ", which will obvious error in judgement,
Such as digital 1-9 identification, when actually entering " 0 ", the probable value that disaggregated model also will be considered to corresponding classification " 6 " is far longer than
The probable value of other corresponding classifications.Therefore which is an adequate condition rather than necessary condition.
The third mode is that SVM mutation algorithm is used in combination --- Support Vector data description (SVDD) algorithm.Instructing
Practice the stage, first using the sample data of all known class as normal category, secondly utilize the sample number of this normal category
SVDD models are obtained according to training, then train whole sample data set to obtain SVM models using traditional SVM.In forecast period,
Judge whether new samples data are normal according to SVDD models first, if normally, judging new samples data using SVM models
Which classification particularly belonged to.The major defect of this mode is present in the SVDD stages, due to there was only a kind of sample number of classification
According to the SVDD models for training to obtain are general more coarse, while lack the limitation and correction of other classification sample datas, SVDD moulds
Type also easily produces over-fitting, therefore whether normally judges that precision is inherently poor for new samples data, that is, to new sample
The judgement precision whether notebook data belongs to unknown classification is poor.
The content of the invention
Therefore, task of the invention be to provide it is a kind of be more suitable for the application scenarios containing unknown classification based on support to
More classification solutions of amount machine.
A kind of according to an aspect of the invention, there is provided more classification sides for including unknown classification based on SVMs
Method, comprise the following steps:
1) for sample to be identified, the i-th category classifier after each training is based respectively on, identifies the sample to be identified
Whether i-th classification is belonged to;Wherein, i=1,2 ... N;I-th category classifier is with the known sample for belonging to the i-th classification
This structure positive sample collection, negative sample collection is built with the known all samples for belonging to remaining N-1 classification, is carried out based on SVM models
Two classification grader obtained from training, the two classification grader are used to identify whether inputted sample belongs to the i-th classification;
2) when by each two classification grader of sample to be identified input, when acquired results are no, assert ought in step 1)
Preceding sample to be identified belongs to unknown classification;When in step 1), output results of one and only one two classification grader for when being,
Then assert that current sample to be identified belongs to the classification corresponding to this two classification grader.
Wherein, the step 2) also includes:When the output result in step 1), having multiple two classification graders is is,
Directly assert that current sample to be identified belongs to unknown classification;Or corresponding to these the two classification graders for being yes in output result
Classification in the range of, further differentiate which classification current sample to be identified belongs to using other sorting techniques.
Wherein, in the step 1), i-th category classifier is trained as follows:
11) taking the i-th classification, remaining N-1 kinds classification is as anti-classification as positive classification;
12) it is trained based on SVM models and obtains the two classification graders as the i-th category classifier.
Wherein, in the step 12), when being trained, first by known sampleNew feature is mapped to from luv space
Space, obtain the sample after corresponding Feature Space TransformationTo cause the classification line of demarcation of SVM models more regular.
Wherein, in the step 12),
The object function of SVM models is configured as:A hypersphere is found in feature space using sample dataWhereinIt is the hyperspherical centre of sphere, R is hyperspherical radius, and y is exported if f (x) < R
=+1 represents that the sample belongs to positive classification, exports y=-1 if f (x) > R, represents that the sample belongs to anti-classification;
Find optimal hypersphere so that the hyperspherical surface area minimizes and known sample data are from the hypersphere
Maximizing Minimum Distance;Then the optimized parameter of SVM models is obtained according to described optimal hypersphere.
Wherein, in the step 12) so that known sample data are from the hyperspherical Maximizing Minimum Distance:Will just
Anti- classification sample data forms two concentric hyperspheres respectively from former hyperspherical minimum range so that concentric hypersphere composition
The thickness of the shell of hypersphere shell maximizes.
Wherein, the object function is:
Constraints isWherein C1And C2It is empirical coefficient, d is suitable
In meeting the sample data of constraints from hyperspherical minimum range, ξiFor slack variable;
The optimized parameter of described SVM models is:
Wherein, in the step 12), the optimized parameter method for solving of described SVM models is as follows:
121) Lagrangian of object function is obtained by method of Lagrange multipliers;
122) local derviation for making Partial Variable in Lagrangian is zero, obtains solving the dual problem of object function, should
The preferred Gaussian kernel of kernel function in dual problem;
123) dual problem is solved by Novel Algorithm, trains to obtain parameter according to known sample dataWith
R.It is to be appreciated that in step 122), although being optimal in most of situation Gaussian kernel, Gaussian kernel is not uniquely to select
Select, in certain embodiments, polynomial kernel (being used as kernel function by the use of polynomial kernel) is also effective to part scene.
Wherein, in the step 12), the Feature Space Transformation enables to the classification linear separability or poly- of SVM models
Collection can divide.
Compared with prior art, the present invention has following technique effect:
(1) generic sample data can cluster, and non-diverging, meet the actual distribution of data, therefore this hair
It is bright closer to real data.Judge that the precision of unknown classification is higher.
(2) disaggregated model has the detectability of unknown classification, not only have the recall ratio suitable with traditional SVM but also
With higher precision ratio.
(3) disaggregated model is simple, and similar to traditional SVM, final disaggregated model only and is located at hypersphere shell interface on a small quantity
Upper or wrong side sample data (i.e. supporting vector) is relevant.
(4) forecast period algorithm is simple, and implementation complexity is low, once trains particularly suitable for model is metastable, more
The occasion of secondary prediction.
Brief description of the drawings
Hereinafter, embodiments of the invention are described in detail with reference to accompanying drawing, wherein:
Fig. 1 is traditional SVM of the prior art three classification schemes schematic diagrames;
Fig. 2 is the schematic diagram of two classification schemes in one embodiment of the invention;
Fig. 3 is the schematic diagram of three classification schemes in one embodiment of the invention.
Embodiment
The present invention is illustrated with reference to the accompanying drawings and detailed description.
According to one embodiment of present invention, there is provided a kind of three sorting techniques for including unknown classification.The three classification side
Method can be decomposed into three two sorting techniques, and each two sorting technique includes training and two stages of prediction.
(1) training stage:
Step 1), one of which classification is chosen as positive classification, remaining two kinds of classification is as anti-classification.
Step 2), by sample dataNew feature space, sample data are mapped to from luv spaceIn new feature space
Corresponding points beAs in Fig. 2 and Fig. 3 ▲, ■ and ● it is shown.Feature space mapping is generally adopted in svm classifier algorithm
A kind of data prediction mode taken, the purpose is to cause sample data, in new feature space, (feature space dimension is more after mapping
It is high) in be more readily separated, for example linear separability or aggregation can divide.That is, in this step, feature space mapping can allow not
It is more more regular with the line of demarcation between sample class.In the present embodiment, if it is possible to can be by two class samples with a linear function
This is completely separable, just claims these sample linear separabilities.If it is complete by two class samples to correspond to hyperspherical function with one
It is complete to separate, just claim the aggregation of these samples to divide.In specific implementation, feature space mapping can be from x-y feature spaces
To the mapping of r- θ feature spaces, its Feature Space Transformation formula is:
θ=arctan (y, x)
Feature space mapping can also be the mapping from x-y feature spaces to u-v feature spaces, and its Feature Space Transformation is public
Formula is:
U=x-1
V=y-1
Above two feature space mapping mode is merely exemplary, in other embodiments, can also use other
Feature space mapping mode, as long as the classification line of demarcation of known sample can be made more regular.
Step 3), the object function of two sorting techniques can be described as:One is found in feature space using sample data
Individual hypersphereWhereinIt is the hyperspherical centre of sphere, R is hyperspherical radius, if f (x) < R
yi=+1 represents that the sample belongs to positive classification, the y if f (x) > Ri=-1 represents that the sample belongs to anti-classification.Optimal hypersphere
Face should make it that hyperspherical surface area minimizes and sample data is from the hyperspherical Maximizing Minimum Distance, the i.e. hypersphere
The surface area in face is as far as possible small, and positive and negative classification sample data forms two concentric hyperspheres respectively from former hyperspherical minimum range,
The thickness of the shell of the hypersphere shell of concentric hypersphere composition is as far as possible big, as shown in Fig. 2 mathematical form is expressed as
Meet constraintsWherein C1And C2It is empirical coefficient, d is suitable
In meeting the sample data of constraints from hyperspherical minimum range, ξiFor slack variable, firm constraints are unsatisfactory for minority
The sample data relaxed constraints of condition (cost is that object function is penalized).
Step 4), the Lagrangian of formula 1 is obtained by method of Lagrange multipliers
Wherein αi>=0, μi>=0 is Lagrange multiplier.
Step 5), orderIt is rightR, d andLocal derviation be zero, can obtain the dual problem of formula (1)
WhereinKernel function is represented, meets constraintsIt is preferred that
Ground, kernel function can select Gaussian kernel
Step 6), formula (3) are a typical quadratic programming problems, are solved using general Novel Algorithm.It is excellent
Choosing, sequence can be used to minimize optimization (SMO) algorithm formula (3) is solved, train to obtain parameter by sample dataWith
R, that is, obtain a disaggregated model.
(2) forecast period
For each new sample data, predicted according to disaggregated modelThe table if f (x) < R
Show that the sample belongs to positive classification, represent that the sample belongs to anti-classification if f (x) > R.
For three sorting techniques of the present embodiment, choose each of which classification and make as positive classification, remaining all categories
For anti-classification, it is trained using above-mentioned two mode classification, respectively obtains disaggregated model, new samples data are according to each classification mould
Type is predicted respectively, may finally be judged whether to belong to known class and be belonged to which known class.
If three two classification problems choose identical kernel function and relevant parameter, equivalent to three hyperspheres will be whole
Feature space is divided into four parts, if new samples data are located at all three hyperspherical outsides, then it is assumed that the sample category
In unknown classification, if new samples data are located at more than one hyperspherical inside (multiple hyperspheres intersect), it can be used
His mode judges the classification of the sample, such as traditional SVM, can also directly think that the sample belongs to unknown classification, such as Fig. 3 institutes
Show.
The effect of the present invention can pass through following description of test.
The more traditional SVM of this experiment and the present invention classifying quality.Traditional SVM parameter configuration is:Linear kernel function (nothing
Feature space maps), empirical parameter C=1;The present invention parameter configuration be:Linear kernel function (no feature space mapping), experience
Parameter C1=10, C2=1.
Infrared data of the data source used in this experiment in daily behavior data set, refer to " Jiang X, Chen Y,
Liu J,et al.AIR:recognizing activity through IR-based distance sensing on
feet[C]//Proceedings of the 2016ACM International Joint Conference on
Pervasive and Ubiquitous Computing:Adjunct.ACM,2016:97-100”.The data set is total to comprising 6 classes
3372 sample datas, be respectively walk (walk), downstairs (downstair), upstairs (upstair), run (run), stand (stay)
Walked (walk_in_place) with original place, each sample data is 38 dimensions.
This experiment is by the use of the preceding 5 class data in above-mentioned data set as training data, and all 6 class data are as test number
According to.Experimental situation is WIN7, Matlab2016b, and optimized algorithm (SMO) uses CVX instruments.Table 1 shows traditional SVM point
Class result.
Table 1
Table 2 shows the svm classifier result of one embodiment of the invention.
Table 2
Test result indicates that the present invention makes disaggregated model have the detectability of unknown classification, not only have and tradition is propped up
Hold the suitable recall ratio of vector machine and there is higher precision ratio, so that the overall precision of classification is also improved.
It should be noted last that the above embodiments are merely illustrative of the technical solutions of the present invention and it is unrestricted.Although ginseng
The present invention is described in detail according to embodiment, it will be understood by those within the art that, to the technical side of the present invention
Case is modified or equivalent substitution, and without departure from the spirit and scope of technical solution of the present invention, it all should cover in the present invention
Right among.
Claims (9)
1. a kind of more sorting techniques for including unknown classification based on SVMs, comprise the following steps:
1) for sample to be identified, the i-th category classifier after each training is based respectively on, whether identifies the sample to be identified
Belong to the i-th classification;Wherein, i=1,2 ... N;I-th category classifier is with the known sample structure for belonging to the i-th classification
Positive sample collection is built, negative sample collection is built with the known all samples for belonging to remaining N-1 classification, is trained based on SVM models
Obtained from two classification graders, the two classification grader is used to identify whether inputted sample belongs to the i-th classification;
2) when by each two classification grader of sample to be identified input, when acquired results are no, identification is currently treated in step 1)
Identification sample belongs to unknown classification;When in step 1), the output result of one and only one two classification grader is when being, then to recognize
Settled preceding sample to be identified belongs to the classification corresponding to this two classification grader.
2. more sorting techniques according to claim 1 that include unknown classification based on SVMs, it is characterised in that
The step 2) also includes:When the output result in step 1), there are multiple two classification graders is is, directly assert and currently treat
Identification sample belongs to unknown classification;Or the scope of the classifications corresponding to these the two classification graders for being yes in output result
It is interior, further differentiate which classification current sample to be identified belongs to using other sorting techniques.
3. more sorting techniques according to claim 1 that include unknown classification based on SVMs, it is characterised in that
In the step 1), i-th category classifier is trained as follows:
11) taking the i-th classification, remaining N-1 kinds classification is as anti-classification as positive classification;
12) it is trained based on SVM models and obtains the two classification graders as the i-th category classifier.
4. more sorting techniques according to claim 3 that include unknown classification based on SVMs, it is characterised in that
In the step 12), when being trained, first by known sampleNew feature space is mapped to from luv space, is obtained pair
Sample after the Feature Space Transformation answeredTo cause the classification line of demarcation of known sample more regular.
5. more sorting techniques according to claim 4 that include unknown classification based on SVMs, it is characterised in that
In the step 12),
The object function of SVM models is configured as:A hypersphere is found in feature space using sample dataWhereinIt is the hyperspherical centre of sphere, R is hyperspherical radius, and y is exported if f (x) < R
=+1 represents that the sample belongs to positive classification, exports y=-1 if f (x) > R, represents that the sample belongs to anti-classification;
Find optimal hypersphere so that the hyperspherical surface area minimizes and known sample data are hyperspherical most from this
Small distance maximizes;Then the optimized parameter of SVM models is obtained according to described optimal hypersphere.
6. more sorting techniques according to claim 5 that include unknown classification based on SVMs, it is characterised in that
In the step 12) so that known sample data are from the hyperspherical Maximizing Minimum Distance:By positive and negative classification sample data
Two concentric hyperspheres are formed respectively from former hyperspherical minimum range so that the thickness of the shell of the hypersphere shell of concentric hypersphere composition
Maximize.
7. more sorting techniques according to claim 5 that include unknown classification based on SVMs, it is characterised in that
The object function is:
Constraints isWherein C1And C2It is empirical coefficient, d is equivalent to satisfaction
The sample data of constraints is from hyperspherical minimum range, ξiFor slack variable;
The optimized parameter of described SVM models is:
8. more sorting techniques according to claim 7 that include unknown classification based on SVMs, it is characterised in that
In the step 12), the optimized parameter method for solving of described SVM models is as follows:
121) Lagrangian of object function is obtained by method of Lagrange multipliers;
122) local derviation for making Partial Variable in Lagrangian is zero, obtains solving the dual problem of object function, the antithesis
Selection of kernel function Gaussian kernel or polynomial kernel in problem;
123) dual problem is solved by Novel Algorithm, trains to obtain parameter according to known sample dataAnd R.
9. more sorting techniques according to claim 7 that include unknown classification based on SVMs, it is characterised in that
In the step 12), the Feature Space Transformation enables to the classification linear separability of SVM models or aggregation to divide.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710536516.1A CN107480690A (en) | 2017-07-04 | 2017-07-04 | A kind of more sorting techniques for including unknown classification based on SVMs |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710536516.1A CN107480690A (en) | 2017-07-04 | 2017-07-04 | A kind of more sorting techniques for including unknown classification based on SVMs |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107480690A true CN107480690A (en) | 2017-12-15 |
Family
ID=60595341
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710536516.1A Pending CN107480690A (en) | 2017-07-04 | 2017-07-04 | A kind of more sorting techniques for including unknown classification based on SVMs |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107480690A (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108551709A (en) * | 2018-04-26 | 2018-09-18 | 南昌航空大学 | A kind of street light modulating method that multinuclear directed acyclic graph support vector machines controls under complex environment |
CN108922583A (en) * | 2018-08-10 | 2018-11-30 | 青岛大学附属医院 | It is a kind of for the liquid handling of medical test, transfer device and its method |
CN109492096A (en) * | 2018-10-23 | 2019-03-19 | 华东理工大学 | A kind of unbalanced data categorizing system integrated based on geometry |
CN109857849A (en) * | 2019-01-18 | 2019-06-07 | 三角兽(北京)科技有限公司 | Answering method, return mechanism, information processing unit and storage medium |
CN109961032A (en) * | 2019-03-18 | 2019-07-02 | 北京字节跳动网络技术有限公司 | Method and apparatus for generating disaggregated model |
CN110503160A (en) * | 2019-08-28 | 2019-11-26 | 北京达佳互联信息技术有限公司 | Image-recognizing method, device, electronic equipment and storage medium |
CN111724353A (en) * | 2020-05-30 | 2020-09-29 | 同济大学 | Surface mount LED flaw labeling method based on One-Class SVM |
CN111753583A (en) * | 2019-03-28 | 2020-10-09 | 阿里巴巴集团控股有限公司 | Identification method and device |
CN111932035A (en) * | 2020-09-22 | 2020-11-13 | 南京福佑在线电子商务有限公司 | Data processing method and device based on multiple models and classified modeling method |
JP2021179944A (en) * | 2020-05-14 | 2021-11-18 | 株式会社東芝 | Classification system, program, and learning system |
CN113989603A (en) * | 2018-01-26 | 2022-01-28 | 唯亚威通讯技术有限公司 | Reduced false positive identification for spectral classification |
CN114241526A (en) * | 2022-02-28 | 2022-03-25 | 南京甄视智能科技有限公司 | Classification model, training method, classification method, electronic device, and storage medium |
CN114881110A (en) * | 2022-04-02 | 2022-08-09 | 西安交通大学 | Real-time detection method for total pressure change mode in on-orbit spacecraft cabin |
-
2017
- 2017-07-04 CN CN201710536516.1A patent/CN107480690A/en active Pending
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113989603A (en) * | 2018-01-26 | 2022-01-28 | 唯亚威通讯技术有限公司 | Reduced false positive identification for spectral classification |
CN108551709A (en) * | 2018-04-26 | 2018-09-18 | 南昌航空大学 | A kind of street light modulating method that multinuclear directed acyclic graph support vector machines controls under complex environment |
CN108922583A (en) * | 2018-08-10 | 2018-11-30 | 青岛大学附属医院 | It is a kind of for the liquid handling of medical test, transfer device and its method |
CN109492096A (en) * | 2018-10-23 | 2019-03-19 | 华东理工大学 | A kind of unbalanced data categorizing system integrated based on geometry |
CN109857849A (en) * | 2019-01-18 | 2019-06-07 | 三角兽(北京)科技有限公司 | Answering method, return mechanism, information processing unit and storage medium |
CN109961032B (en) * | 2019-03-18 | 2022-03-29 | 北京字节跳动网络技术有限公司 | Method and apparatus for generating classification model |
CN109961032A (en) * | 2019-03-18 | 2019-07-02 | 北京字节跳动网络技术有限公司 | Method and apparatus for generating disaggregated model |
CN111753583A (en) * | 2019-03-28 | 2020-10-09 | 阿里巴巴集团控股有限公司 | Identification method and device |
CN110503160A (en) * | 2019-08-28 | 2019-11-26 | 北京达佳互联信息技术有限公司 | Image-recognizing method, device, electronic equipment and storage medium |
US12008799B2 (en) | 2020-05-14 | 2024-06-11 | Kabushiki Kaisha Toshiba | Classification system and learning system |
JP2021179944A (en) * | 2020-05-14 | 2021-11-18 | 株式会社東芝 | Classification system, program, and learning system |
JP7354063B2 (en) | 2020-05-14 | 2023-10-02 | 株式会社東芝 | Classification systems, programs and learning systems |
CN111724353A (en) * | 2020-05-30 | 2020-09-29 | 同济大学 | Surface mount LED flaw labeling method based on One-Class SVM |
CN111932035A (en) * | 2020-09-22 | 2020-11-13 | 南京福佑在线电子商务有限公司 | Data processing method and device based on multiple models and classified modeling method |
CN111932035B (en) * | 2020-09-22 | 2021-01-08 | 南京福佑在线电子商务有限公司 | Data processing method and device based on multiple models and classified modeling method |
CN114241526A (en) * | 2022-02-28 | 2022-03-25 | 南京甄视智能科技有限公司 | Classification model, training method, classification method, electronic device, and storage medium |
CN114881110A (en) * | 2022-04-02 | 2022-08-09 | 西安交通大学 | Real-time detection method for total pressure change mode in on-orbit spacecraft cabin |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107480690A (en) | A kind of more sorting techniques for including unknown classification based on SVMs | |
CN111144490A (en) | Fine granularity identification method based on alternative knowledge distillation strategy | |
US11837329B2 (en) | Method for classifying multi-granularity breast cancer genes based on double self-adaptive neighborhood radius | |
CN110717554A (en) | Image recognition method, electronic device, and storage medium | |
CN111506729B (en) | Information processing method, device and computer readable storage medium | |
Song et al. | Classifier calibration: a survey on how to assess and improve predicted class probabilities | |
CN102034107A (en) | Unhealthy image differentiating method based on robust visual attention feature and sparse representation | |
WO2021114818A1 (en) | Method, system, and device for oct image quality evaluation based on fourier transform | |
CN111241992A (en) | Face recognition model construction method, recognition method, device, equipment and storage medium | |
CN111160538A (en) | Method and system for updating margin parameter value in loss function | |
Wang et al. | Mushroom toxicity recognition based on multigrained cascade forest | |
Owsiński et al. | Reverse clustering | |
CN113724195B (en) | Quantitative analysis model and establishment method of protein based on immunofluorescence image | |
Elwahsh et al. | A new approach for cancer prediction based on deep neural learning | |
CN114511759A (en) | Method and system for identifying categories and determining characteristics of skin state images | |
Vivona et al. | Automated approach for indirect immunofluorescence images classification based on unsupervised clustering method | |
Mat Jizat et al. | Evaluation of the transfer learning models in wafer defects classification | |
Kim et al. | Multiple instance neural networks based on sparse attention for cancer detection using T-cell receptor sequences | |
CN110879821A (en) | Method, device, equipment and storage medium for generating rating card model derivative label | |
CN115587884A (en) | User loan default prediction method based on improved extreme learning machine | |
CN115472179A (en) | Automatic detection method and system for digital audio deletion and insertion tampering operation | |
Han et al. | Automatic used mobile phone color determination: Enhancing the used mobile phone recycling in China | |
Zhu et al. | A Three‐step Method for Three‐way Clustering by Similarity‐based Sample’s Stability | |
Liu | The alexnet-resnet-inception network for classifying fruit images | |
Cai et al. | Application and research progress of machine learning in Bioinformatics |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171215 |
|
RJ01 | Rejection of invention patent application after publication |