CN108846437A - The method of raising TWSVM algorithm robustness based on capped-l1 norm - Google Patents
The method of raising TWSVM algorithm robustness based on capped-l1 norm Download PDFInfo
- Publication number
- CN108846437A CN108846437A CN201810622213.6A CN201810622213A CN108846437A CN 108846437 A CN108846437 A CN 108846437A CN 201810622213 A CN201810622213 A CN 201810622213A CN 108846437 A CN108846437 A CN 108846437A
- Authority
- CN
- China
- Prior art keywords
- matrix
- data
- twsvm
- capped
- norm
- 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.)
- Withdrawn
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
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computing Systems (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to data processing fields, disclose a kind of method of raising TWSVM algorithm robustness based on capped-l1 norm, including:Input raw data matrix M and parameter C1、C2、ε1And ε2;M points are positive and negative data matrix H and G;Two diagonal matrix matrix Fs and D are initialized as unit matrix;According to H, G, C1, C2, ε1、ε2, F and D, the parameter w and b of classification plane is calculated;According to w and b, calculate all data points in H and D to classification plane distance, if the distance for having data point to plane of classifying in H and D is respectively greater than ε1And ε2, then the point is judged for outlier, sets one close to zero value smallval for that entry value in corresponding F and D;To update diagonal matrix F and D, target value is calculated.The present invention can greatly improve the robustness of TWSVM algorithm, and also have preferable precision on original data set.
Description
Technical field
The present invention relates to algorithm improvement and data processing field, in particular to a kind of raising based on capped-l1 norm
The method of TWSVM algorithm robustness.
Background technique
Support vector machines has been widely applied to data classification and regression problem up as a good tool
?.Including every field such as biological information, text classification, image procossings.In recent years, Mangasarian and Wild is proposed
More plane approximation support vector machines based on generalized eigenvalue (Proximal SVM based 0n Generalized
Eigenvalues, GEPSVM).The algorithm solves two non-parallel hyperplane by solving two generalized eigenvalue problems.
GEPSVM is while guaranteeing computational efficiency better compared with SVM, also available preferable classification performance.Derived from the think of of GEPSVM
Think, 2007, Javadeva et al. propose twin support vector machines (Twin SVM, TSVM) .TSVM seek two it is non-parallel
Optimal classification surface so that each classifying face is suitable for cross division far from other class sample .TSVM close to a kind of sample
The classification of face data collection, and solve two relatively smaller quadratic programming problem (Quadratic Programming
Problem, QPP), after this makes the speed of TSVM be significantly faster than that the support vector machines of standard so far, many researchers exist
Many algorithms are proposed on the basis of TWSVM.
But many already present TWSVM models often do not account for noise data in training classifying face, in reality
In, many data can all have data noise, if not removing these noise datas of pipe, can be easy to cause the classification trained and obtained
Plane has deviation, and the accuracy of classification reduces, and the robust performance of algorithm is bad.Why noise data for classification plane certainly
Plan has an impact because that many models based on TWSVM use or L2 norm, L2 norm can be expanded by square operation
The influence of big noise data.If a data do not have noise, the performance of these algorithms can be very good, still, reality
In life, not having noisy data to be impossible existing.Therefore, in algorithm for design, we are contemplated that making an uproar for data
Sound problem.It will be apparent that if many of data noise outlier, L2 norm is very not applicable, the TWSVM based on L2 norm
Model algorithm is also just very not applicable.In order to improve algorithm for the robustness of outlier, we have proposed capped-l1 norms
TWSVM has abandoned the disadvantage of L2, greatly improves the robustness of algorithm.
Summary of the invention
Goal of the invention:Aiming at the problems existing in the prior art, the present invention provides a kind of mentioning based on capped-l1 norm
The method of high TWSVM algorithm robustness can greatly improve the robustness of TWSVM algorithm, and on original data set
There is preferable precision.
Technical solution:The side for the raising TWSVM algorithm robustness that the present invention provides a kind of based on capped-l1 norm
Method includes the following steps:Step 1:Input raw data matrix M and parameter C1、C2、ε1And ε2;The data matrix M points is just
Data matrix H and negative data matrix G, wherein H ∈ Rm1×(n+1), G∈Rm2×(n+1), wherein m1And m2For the data matrix M
In data amount check, n be the data matrix M data dimension;Step 2:Two diagonal matrix matrix Fs and D are initialized as list
Bit matrix;Step 3:According to the correction data matrix H, negative data matrix G, parameter C1, C2, ε1、ε2And diagonal matrix F and D,
The parameter w and b of classification plane is calculated;Step 4:According to the w and b, all numbers in the correction data matrix H are calculated
Strong point to classification plane distance, if having data point be greater than ε1, then the point is judged for outlier, by the corresponding diagonal matrix
That entry value in F is set as smallval, smallval be one close to zero value;According to the w and b, calculate described negative
Data point in data matrix G to the classification plane distance, if there is data point to be greater than ε2, then judge the point for outlier,
The smallval is set by that entry value in the corresponding diagonal matrix D;In this approach come update diagonal matrix F and
D;Step 5:Calculate target value obj.
Preferably, it in the step 4, is set that entry value in corresponding diagonal matrix F to according to outlier
The calculation method of smallval is as follows:;It will be corresponded to according to outlier
Diagonal matrix D in that entry value be set as smallval calculation method it is as follows:
。
Preferably, in the step 5, the target value;Its
In, z=(w, b)T, e is a m1× 1 column vector, each element are 1.
Further, further comprising the steps of after the step 5:Step 6:Step 3-step 5 described in iteration
Step is until the target value obj restrains;Step 7 determines the parameter w and b of optimal classification plane.Step 3 and step 4
Iteration update setting so that the present invention hair can obtain it is optimal classification plane parameter w and b, further increase TWSVM algorithm
Robustness and precision.
Preferably, the smallval=1e-5.
Beneficial effect:
Different from traditional TWSVM algorithm, capped-l1 normal form is applied in objective function by this algorithm, enables the algorithm
Greatly improve the robustness of TWSVM algorithm;Its principle is the influence for removing those noise outliers to the decision of classification plane,
Decision is carried out again after removing these outliers.In addition, the present invention is because the loss part in objective function applies
Capped-l1 norm, therefore regardless of how serious a point is by the of mistake minute, the value of the loss of this function will not change too
Greatly, the robustness of algorithm is improved with this;Finally, the present invention also simply effectively iteratively solves method by a set of, solve
A local optimum out.
Compared to more traditional TWSVM algorithm, the present invention is based on the TWSVM methods of capped-l1 normal form distance metric to have
Better robustness, and also have preferable precision on original data set;The present invention has selected 12 UCI data, in phase
With on data set, by this method and TWSVM, WLTSVM, L1-GEPSVM, L1-NPSVM and this 5 algorithms progress of pTWSVM are smart
Comparison in difference is spent, comparison result shows that method of the invention is concentrated in 12 data, there is the better than other algorithms of 8 performances, and
And in the case where identical noise situations are added, what is still showed on 8 data sets is better than other algorithms.
Detailed description of the invention
Fig. 1 is target value iteration convergence mistake of the TWSVM algorithm based on capped-l1 norm on data set Haberman
The schematic diagram of journey;
Fig. 2 is target value iterative convergent process signal of the TWSVM algorithm based on capped-l1 norm on data set Sonar
Figure.
Fig. 3 is that in the case where noise spot artificially is added into the data of crossing plane, capped-l1 model is based in the present invention
The precision of several TWSVM algorithms and other five algorithms changes comparison schematic diagram;
Fig. 4 is the TWSVM algorithm and other five based on capped-l1 norm in the present invention in different noise factors
The precision of a algorithm changes comparison schematic diagram.
Specific embodiment
The present invention is described in detail with reference to the accompanying drawing.
The method for improving TWSVM algorithm robustness based on capped-l1 norm that present embodiments provide for a kind of, including
Following steps:
Step 1:Input raw data matrix M and parameter C1、C2、ε1And ε2;Data matrix M points are correction data matrix H and negative
According to matrix G, wherein H ∈ Rm1×(n+1), G∈Rm2×(n+1), wherein m1And m2For the data amount check in data matrix M, n is data
The data dimension of matrix M;
Step 2:Two diagonal matrix matrix Fs and D are initialized as unit matrix;
Step 3:According to correction data matrix H, negative data matrix G, parameter C1, C2, ε1、ε2And diagonal matrix F and D, it calculates
To the parameter w and b of classification plane;
Step 4:According to w and b, calculate all data points in correction data matrix H to plane of classifying distance, if there is data point
Greater than ε1, then judge that the point for outlier, passes through formulaIt will correspond to
Diagonal matrix F in that entry value be set as 1e-5;
According to w and b, calculate the data point in negative data matrix G to classification plane distance, if there is data point greater than ε2, then sentence
The point break as outlier, passes through formula
1e-5 is set by that entry value in corresponding diagonal matrix D;
Update diagonal matrix F and D in this approach;
Step 5:Target value obj is calculated,, wherein z=(w, b)T, e is
One m1× 1 column vector, each element are 1.
Step 6:Iteration three-step 5 of above-mentioned steps step is until target value obj restrains;
Step 7 determines the parameter w and b of optimal classification plane.
In order to intuitively show that this method can quickly restrain, two are listed in present embodiment based on capped-
Convergence sex expression of the TWSVM algorithm of l1 norm on two UCI data sets.Such as Fig. 1 and 2, from Fig. 1 and 2 it can be found that this
Method can with cracking iteration convergence to a stable value, illustrate the TWSVM algorithm based on capped-l1 norm calculate and
It is all feasible on time complexity.
In addition, artificially joined obvious noise spot, such as Fig. 3 in present embodiment into the data of crossing plane
Show,
With this data instance, TWSVM, WLTSVM, L1-GEPSVM are compared, L1-NPSVM, pTWSVM and be based on
Precision of the TWSVM algorithm of capped-l1 norm in above data, is 55.26%, 95.12%, 97.60% respectively,
54.64%, 55.36% and 98.07%.It will be apparent that this method(TWSVM algorithm based on capped-l1 norm)Precision is higher than
Other 5 algorithms, robust performance are more preferable.
In addition, present embodiment compared the TWSVM in different noise proportionals, based on capped-l1 norm
The variation of the precision of algorithm and other five algorithms.As Fig. 4 is based on capped- as can be seen from Figure 4 under different noise situations
The TWSVM algorithm of l1 norm is not only got well than other arithmetic accuracies, but also more steady than other algorithms.Although in noise factor
In the case where 0.25, the precision of TWSVM and the precision of the TWSVM algorithm based on capped-l1 norm will as, still
TWSVM is showed very unstable.In addition, WLTSVM, L1-GEPSVM, although L1-NPSVM and pTWSVM is in different noises
Under the influence of show very steady, but their mean accuracy is 78.90%, 83.66%, 80.23% and 77.67%, and base
In capped-l1 norm TWSVM algorithm be 86.00%.So finally or this method(Based on capped-l1 norm
TWSVM algorithm)It is more advantageous.
The technical concepts and features of above embodiment only to illustrate the invention, its object is to allow be familiar with technique
People cans understand the content of the present invention and implement it accordingly, and it is not intended to limit the scope of the present invention.It is all according to the present invention
The equivalent transformation or modification that Spirit Essence is done, should be covered by the protection scope of the present invention.
Claims (5)
1. a kind of method of the raising TWSVM algorithm robustness based on capped-l1 norm, which is characterized in that including following step
Suddenly:
Step 1:Input raw data matrix M and parameter C1、C2、ε1And ε2;The data matrix M point for correction data matrix H and
Negative data matrix G, wherein H ∈ Rm1×(n+1), G∈Rm2×(n+1), wherein m1And m2For the data in the data matrix M
Number, n are the data dimension of the data matrix M;
Step 2:Two diagonal matrix matrix Fs and D are initialized as unit matrix;
Step 3:According to the correction data matrix H, negative data matrix G, parameter C1, C2, ε1、ε2And diagonal matrix F and D, meter
It calculates and obtains the parameter w and b of classification plane;
Step 4:According to the w and b, calculate all data points in the correction data matrix H to classification plane distance, if
There is data point to be greater than ε1, then judge that the point for outlier, sets that entry value in the corresponding diagonal matrix F to
Smallval, smallval be one close to zero value;
According to the w and b, calculate data point in the negative data matrix G to the classification plane distance, if there is data point
Greater than ε2, then the point is judged for outlier, sets described for that entry value in the corresponding diagonal matrix D
smallval;
Update diagonal matrix F and D in this approach;
Step 5:Calculate target value obj.
2. the method for the raising TWSVM algorithm robustness according to claim 1 based on capped-l1 norm, feature
It is, in the step 4, sets that entry value in corresponding diagonal matrix F to according to outlier the meter of smallval
Calculation method is as follows:
;
The calculation method for setting smallval for that entry value in corresponding diagonal matrix D according to outlier is as follows:
。
3. the method for the raising TWSVM algorithm robustness according to claim 1 based on capped-l1 norm, feature
It is, in the step 5,
The target value;
Wherein, z=(w, b)T, e is a m1× 1 column vector, each element are 1.
4. the raising TWSVM algorithm robustness according to any one of claim 1 to 3 based on capped-l1 norm
Method, which is characterized in that further comprising the steps of after the step 5:
Step 6:Step 3 described in iteration-step 5 step is until the target value obj restrains;
Step 7 determines the parameter w and b of optimal classification plane.
5. the raising TWSVM algorithm robustness according to any one of claim 1 to 3 based on capped-l1 norm
Method, which is characterized in that the smallval=1e-5.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810622213.6A CN108846437A (en) | 2018-06-15 | 2018-06-15 | The method of raising TWSVM algorithm robustness based on capped-l1 norm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810622213.6A CN108846437A (en) | 2018-06-15 | 2018-06-15 | The method of raising TWSVM algorithm robustness based on capped-l1 norm |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108846437A true CN108846437A (en) | 2018-11-20 |
Family
ID=64202083
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810622213.6A Withdrawn CN108846437A (en) | 2018-06-15 | 2018-06-15 | The method of raising TWSVM algorithm robustness based on capped-l1 norm |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108846437A (en) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105335615A (en) * | 2015-10-31 | 2016-02-17 | 电子科技大学 | Low-complexity two-dimensional angle and polarization parameter joint estimation method |
CN106847248A (en) * | 2017-01-05 | 2017-06-13 | 天津大学 | Chord recognition methods based on robustness scale contour feature and vector machine |
-
2018
- 2018-06-15 CN CN201810622213.6A patent/CN108846437A/en not_active Withdrawn
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105335615A (en) * | 2015-10-31 | 2016-02-17 | 电子科技大学 | Low-complexity two-dimensional angle and polarization parameter joint estimation method |
CN106847248A (en) * | 2017-01-05 | 2017-06-13 | 天津大学 | Chord recognition methods based on robustness scale contour feature and vector machine |
Non-Patent Citations (2)
Title |
---|
WENHAO JIANG: ""Robust Dictionary Learning with Capped `1-Norm"", 《PROCEEDINGS OF THE TWENTY-FOURTH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
ZHOUYUAN HUO: ""Video Recovery via Learning Variation and Consistency of Images"", 《PROCEEDINGS OF THE THIRTY-FIRST AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP6922387B2 (en) | Recognition devices, training devices and methods based on deep neural networks | |
JP2021077377A (en) | Method and apparatus for learning object recognition model | |
CN105210115B (en) | Performing gesture recognition using 2D image data | |
CN107292902B (en) | Two-dimensional Otsu image segmentation method combined with drosophila optimization algorithm | |
CN105551015A (en) | Scattered-point cloud image registering method | |
CN105046694A (en) | Quick point cloud registration method based on curved surface fitting coefficient features | |
JP2010266983A (en) | Information processing apparatus and method, learning device and method, program, and information processing system | |
JP2005242808A (en) | Reference data optimization learning method and pattern recognition system | |
Riaz et al. | Fouriernet: Compact mask representation for instance segmentation using differentiable shape decoders | |
WO2015025472A1 (en) | Feature conversion learning device, feature conversion learning method, and program storage medium | |
CN112633413B (en) | Underwater target identification method based on improved PSO-TSNE feature selection | |
CN114936518A (en) | Method for solving design parameters of tension/compression spring | |
JP6942203B2 (en) | Data processing system and data processing method | |
CN108846437A (en) | The method of raising TWSVM algorithm robustness based on capped-l1 norm | |
CN106022212A (en) | Gyroscope temperature drift modeling method | |
LI et al. | Training restricted boltzmann machine using gradient fixing based algorithm | |
JP6121187B2 (en) | Acoustic model correction parameter estimation apparatus, method and program thereof | |
CN114818203A (en) | Reducer design method based on SWA algorithm | |
JP7364047B2 (en) | Learning devices, learning methods, and programs | |
CN112070127A (en) | Intelligent analysis-based mass data sample increment analysis method | |
JP5130934B2 (en) | Recognition system, information processing apparatus, design apparatus, and program | |
WO2020087254A1 (en) | Optimization method for convolutional neural network, and related product | |
Lu et al. | Improved SVM classifier incorporating adaptive condensed instances based on hybrid continuous-discrete particle swarm optimization | |
CN116152316B (en) | Image registration method based on self-adaptive parameter particle swarm algorithm | |
Wang et al. | A novel visual tracking system with adaptive incremental extreme learning machine |
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 | ||
WW01 | Invention patent application withdrawn after publication | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20181120 |