CN109447178A - A kind of svm classifier method based on mixed kernel function - Google Patents
A kind of svm classifier method based on mixed kernel function Download PDFInfo
- Publication number
- CN109447178A CN109447178A CN201811343420.4A CN201811343420A CN109447178A CN 109447178 A CN109447178 A CN 109447178A CN 201811343420 A CN201811343420 A CN 201811343420A CN 109447178 A CN109447178 A CN 109447178A
- Authority
- CN
- China
- Prior art keywords
- kernel function
- mixed
- support vector
- sample
- function
- 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
Abstract
The svm classifier method based on mixed kernel function that the invention discloses a kind of, one, collection data set, analyzes each sample in the data set record of collection, distinguishes sample different attribute, determine input and output sample;Two, kernel function is selected and constructed, exponential distribution kernel function is mixed with Radial basis kernel function;Three, the parameter in mixed kernel function is optimized;Four, C-SVC model is selected, the support vector cassification model based on novel mixed kernel function is established;Five, classification prediction is carried out by the support vector cassification model established.The present invention makes full use of the global performance of exponential function and the Local Property of radial basis function, and using the particle swarm optimization algorithm Optimized model parameter with Gaussian mutation, improves the overall performance of support vector machines;The study of exponential distribution kernel function of overall importance and Generalization Capability are above other single kernel functions, and the support vector machines performance of novel mixed kernel function is substantially better than the support vector machines performance of other mixed kernel functions.
Description
Technical field
The present invention relates to a kind of svm classifier methods, and in particular to a kind of svm classifier method based on mixed kernel function.
Background technique
SVM (Support Vector Machine) refers to support vector machines, is a kind of common method of discrimination.In machine
Device learning areas is the learning model for having supervision, commonly used to carry out pattern-recognition, classification and regression analysis.The side SVM
Method is that sample space is mapped in a higher-dimension or even infinite dimensional feature space by a Nonlinear Mapping p
(space Hilbert), so that being converted into the problem of Nonlinear separability in original sample space linear in feature space
The problem of can dividing.Briefly, peacekeeping linearisation is exactly risen.Dimension is risen, sample is done to higher dimensional space exactly and is mapped, ordinary circumstance
Lower this will increase the complexity of calculating, or even can cause " dimension disaster ".But SVM is come as the problems such as classification, recurrence
Say, it is likely that make low-dimensional sample space can not linear process sample set, can but pass through a line in high-dimensional feature space
Property hyperplane realize linear partition (or return).The general dimension that rises can all bring the complication of calculating, and SVM method dexterously solves
This problem: kernel function is introduced, and requires no knowledge about the explicit expression of Nonlinear Mapping;Due to being in high-dimensional feature space
In establish linear learning machine, so not only hardly increasing the complexity of calculating, and to a certain degree compared with linear model
On avoid " dimension disaster ".
At this stage, SVM has been applied to Face datection, and the fault diagnosis of Turbo-generator Set is classified, and returns, cluster, when
Between sequence prediction, System Discrimination, Financial Engineering, biological medicine signal processing, data mining, biological information, text mining is adaptive
Induction signal processing, splice site identification, the database learning algorithm based on support vector machines, the identification of handwritten form similar character are supported
Application of the vector machine Function Fitting in fractal interpolation, the Initial Alignment of Inertial Navigation Systems based on support vector machines, Prediction for Rock Burst
Support vector machines, defect recognition, computer keyboard subscriber authentication, video caption are automatically positioned on extraction, and speaker is really
The multiple fields such as recognize.
Original sample can be projected higher dimensional space from lower dimensional space by kernel function, be that support vector machines (SVM) processing is non-thread
One of the important method of the inseparable problem of property, it make sample from former Space Nonlinear can not isloation state be converted to higher dimensional space
Approximately linear can isloation state.The type of kernel function has determined the size of mapping inner product and higher dimensional space, to change sample number
According to the complexity being distributed in higher dimensional space.If the space dimensionality of data projection is larger, obtained model is more complex, experience
Risk is small but fiducial range is big, is easy to appear overfitting;Vice versa.
For the selection of kernel function or construction, there are no unified rules at present, generally select core letter using empirical method
Number, as long as the function for meeting Mercer condition is theoretically all chosen as kernel function.Therefore, it is necessary to design a kind of efficient core letter
Number, by data projection to suitable higher dimensional space, can make classifying quality more excellent.
Summary of the invention
It is an object of the invention to overcome defect of the existing technology, provide the SVM based on mixed kernel function points a kind of
Class method.
In order to achieve the above objectives, the technical scheme adopted by the invention is that: a kind of svm classifier side based on mixed kernel function
Method, steps are as follows:
One, data set is collected, each sample in the data set record of collection is analyzed, sample is distinguished and does not belong to
Property, determine input and output sample;
Two, kernel function is selected and constructed, the exponential distribution global kernels and radial base core of Mercer condition will be met
Function is mixed, and novel mixed kernel function is obtained:
In formula, parameter t is ratio shared by two kinds of kernel functions, and σ is the core width of Radial basis kernel function, and γ is exponential distribution
The parameter of kernel function, x, y are input and output sample pair in sample set;
Three, the parameter in novel mixed kernel function is optimized;
Four, C-SVC model is selected, the support vector cassification model based on novel mixed kernel function is established;
Five, classification prediction is carried out by the support vector cassification model established.
Further, the optimization is using the particle swarm optimization algorithm for introducing Gaussian mutation, in novel mixed kernel function
Three parameters σ, γ, t optimize.
Further, the step of particle swarm optimization algorithm for introducing Gaussian mutation is as follows:
1, parameter, population size value, maximum number of iterations value are set, and local search ability C1, ability of searching optimum C2 are used to
Property weights omega value, limit particle search area and population flying speed;
2, the position X of particle (σ, γ, t) is initializediWith speed Vi;
3, the fitness value of each particle is calculated;
4, global extremum and individual extreme value are determined;
5, iteration optimizing judges whether to reach maximum classification accuracy, if so, turning to step 8, otherwise continues downward
It executes;
6, basisWithThe speed of more new particle and position;
7, a particle is randomly choosed, Gaussian mutation is carried out to it, population diversity is kept, goes back to step 3;
8, optimal solution is exported.
The method have the benefit that: using the exponential distribution global kernels and radial direction for meeting Mercer condition
The novel mixed kernel function of base kernel function building, that is, have preferable extrapolability, the data point of wide apart can be made to core letter
Numerical value has an impact, and has stronger learning ability, also has an impact to neighbouring local message;Introduce the population of Gaussian mutation
Optimization algorithm optimizes the parameter of the support vector cassification model based on novel mixed kernel function simultaneously, more common than other
Mixed kernel function has higher classification accuracy.
Detailed description of the invention
The present invention is further elaborated with embodiment with reference to the accompanying drawing.
Fig. 1 is the output characteristic curve figure of RBF of the present invention;
Fig. 2 is the output characteristic curve figure of exponential kernel functions of the present invention;
Fig. 3 is curve graph (γ=0.2) of the compound kernel function of the present invention at test point;
Fig. 4 is curve graph (σ=0.2) of the compound kernel function of the present invention at test point.
Specific embodiment
Embodiment 1
A kind of svm classifier method based on mixed kernel function, steps are as follows:
One, data set is collected, each sample in the data set record of collection is analyzed, sample is distinguished and does not belong to
Property, determine input and output sample;
Two, kernel function is selected and constructed, exponential distribution kernel function is mixed with Radial basis kernel function;
Three, the parameter in mixed kernel function is optimized, using the particle swarm optimization algorithm for introducing Gaussian mutation;
Four, C-SVC model is selected, the support vector cassification model based on novel mixed kernel function is established;
Five, classification prediction is carried out by the support vector cassification model established.
Wherein the generation of novel mixed kernel function is according to as follows:
The concept of kernel function, it is assumed that give m training sample { x1, x2..., xm, each xiA corresponding feature to
Amount.So, there is a function K (xi, xj) meet K (xi, xj)=Φ (xi)TΦ(xj) and meet Mercer condition, then K (xi, xj)
Referred to as kernel function.Wherein, Φ (x) is some nonlinear mapping function.Mercer condition refers to for arbitrary g (x) ∈ RnAnd
≠ 0, there is ∫ ∫ K (xi, xj)g(xi)g(xj)dxidxj≥0.I.e. for all training sample (x1, x2..., xm, K (xi, xj)
It is effective kernel function, then its corresponding nuclear matrix is symmetrical and positive semi-definite.
Kernel function can be divided into two major classes according to the characteristic of kernel function by the design of mixed kernel function: global kernel function drawn game
Portion's kernel function.Global kernel function has global property, is good at extracting sample global property, such as line its main feature is that interpolation ability is weaker
Property, Polynomial kernel function etc..Local kernel function has locality, and local learning ability is higher than global learning ability.This method
The local kernel function of use is Radial basis kernel function (Radial Basis Function, RBF), and formula is as follows:
Shown in Fig. 1, RBF kernel function output curve diagram when test point takes 0.2, σ to take 0.1,0.2,0.5,0.8,1 is exactly
One typical local kernel function.It can be seen from the figure that the region near test point, kernel function influences it very greatly, and from
Test point is remoter, and output valve levels off to 0.
Using the global kernels of an exponential distribution, it meets Mercer condition, is an effective kernel function.
It proves: for m training sample { x1, x2..., xm, by any two xiAnd xjIt brings into above formula (3), can obtain:
As it can be seen that corresponding m m matrix K is a symmetrical matrix.
Assuming that each training sample has k attributive character, Φk(x) it is right to indicate that original space sample x is mapped to higher dimensional space institute
The kth dimension attribute value of the mapping function Φ (x) answered.So for any vector z, obtain:
Therefore, the kernel matrix K obtained on training set is positive semi-definite (K >=0).
To sum up, index nuclear matrix K is symmetrical positive semi-definite to meet Mercer condition.Therefore, exponential kernel functions are effective
Kernel function, have been used for core study.Exponential kernel functions when test point equally takes 0.2, γ to take 0.1,0.2,0.5,0.8,1 respectively
Curve of output is as shown in Figure 2.It can be seen from the figure that the kernel function has preferable extrapolability, wide apart may make
Data point all has an impact to kernel function value.Although the output characteristics of exponential kernel functions is similar with Polynomial kernel function, from letter
It is seen in number prototype, only one γ of the parameter of the function, and Polynomial kernel function has tri- parameters of γ, r, d to need to select, because
This reduces the dependence of parameter.
By RBF kernel function in conjunction with above-mentioned exponential kernel functions, for the advantages of both giving full play to, formed of the invention novel
Mixed kernel function.
According to the definition of kernel function, if K1、K2It is kernel function, then there is α > 0, β > 0 is for arbitrary vector z, zTα
K1Z > 0, zTβK2Z > 0.Therefore, zTαK1z+zTβk2Z=zT(αK1+βK2) z > 0, (α, β > 0) to get arrive α K1+βK2Nuclear moment
Battle array is positive definite namely α K1+βK2It is kernel function.Formula is as follows:
In formula, parameter t is ratio shared by two kinds of kernel functions, generally takes 0≤t≤1.Combination nova kernel function is in test point
Output when 0.2 is illustrated in fig. 3 shown below.
Kernel functional parameter based on PSO optimizes, and there are three parameters (in the σ, new kernel function in RBF in novel kernel function
γ, kernel function proportion t) needs optimize, their value has a direct impact the result of SVM model.Using structure
Simply, the particle swarm optimization algorithm (Particle Swarm Optimization, PSO) being easily achieved carries out model parameter
Optimization.
The present invention uses the support vector machines (C-SVC) with punishment parameter C, therefore parameter C is also required to optimize.
Specific step is as follows for algorithm:
Step1: parameter setting, population size are set as 20, and maximum number of iterations 200 represents local search ability
c1=1.5, in order to make particle evolve to globe optimum, setting represents the c of ability of searching optimum2=1.7, inertia weight ω's
Value is 1, particle search area C ∈ [0.1,100], σ ∈ [0.1,100], γ ∈ [0.01,10], t ∈ [0.01,1];Particle
Group's flying speed range [- 0.6*Cmax, 0.6*Cmax], [- 0.6* σmax, 0.6* σmax], [- 0.6* γmax, 0.6* γmax], [-
0.6*tmax, 0.6*tmax];
Step2: the position and speed of initialization particle (C, σ, γ, t), the position of each particle: Xi=(Xmax-Xmin)*
rand+Xmin;Speed: Vi=Vmax*rands;
Step3: calculating the fitness value of each particle, using the classification accuracy of training set as fitness function:
Step4: retain each individual and run to current time obtained desired positionsWith it is best when former generation group
Position
Step5: iteration optimizing judges whether to reach maximum classification accuracy, if so, turning to Step 8, otherwise execute
Step 6~7;
Step6: according toWith
The speed of more new particle and position;
Step7: one particle of random selection carries out Gaussian mutation X to iti=Xi× g, g ∈ N (0,1) keep population more
Sample goes back to Step3;
Step8: output optimal solution.
Embodiment 2
The chemical composition analysis for recording the grape wine of three kinds of different cultivars on Italian the same area forms Wine data set,
Data set shares 178 samples, and each sample contains 13 attributes.
The validity of the novel mixed kernel function of the verifying present invention establishes base by Wine data set by 10 random groupings
In the support vector cassification model of mixed kernel function, involved parameter is optimized using PSO, carries out classification prognostic experiment.With
Machine is grouped the adaptability for reflecting novel mixed kernel function.
The classification results for the support vector cassification model that obtained result and other kernel functions are constructed compare, as follows
Shown in table 1.
1 Wine data set of table is grouped the classification accuracy of different kernel functions at random
As can be known from Table 1, generally for the random division training set and test set of Wine data set, novel mixed nucleus letter
Several training set classification accuracies and test set classification accuracy are above other kernel functions;Therefore, novel mixed kernel function fills
Distribute the of overall importance of the locality and index core for having waved RBF, learning ability and generalization ability are all compared with other kernel functions height, it was demonstrated that
The validity of the novel mixed kernel function of the present invention.
Embodiment 3
The present embodiment be Iris data set verify example, in order to more fully more all kinds of kernel functions it is single or mixing
SVM classifying quality, experimental subjects are the Iris data set that UCI database website provides.
Iris data set includes 150 samples, is divided into 3 classifications, and each sample contains 4 attributes.The present invention will
The 50% of every class sample of the data set is used as training set, and 50% in addition is used as test set.Parameter is optimized through PSO respectively,
Svm classifier model is established, the classification accuracy of test set is as shown in table 2 below.
Classification results under 2 Iris data set difference kernel function of table
As shown in Table 2, for using single kernel function, kernel function proposed by the invention than linear, multinomial,
The classification accuracy of RBF kernel function wants high;For mixed kernel function, the classification of novel mixed kernel function of the invention is imitated
Fruit will get well compared with other mixed kernel functions, illustrate the support vector machines performance of novel mixed kernel function construction of the invention better than other
The support vector machines of mixed kernel function.
This method constructs a kind of new mixed kernel function that can take into account study and Generalization Ability, the experimental results showed that,
The classification accuracy of data set is all higher than single kernel function and other mixed kernel functions.Therefore, with the novel mixing of the present invention
The SVM model of Kernel is classified, and preferably classification prediction result can be obtained.
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited to above-mentioned
Particular implementation, those skilled in the art can make various deformations or amendments within the scope of the claims, this not shadow
Ring substantive content of the invention.
Claims (3)
1. a kind of svm classifier method based on mixed kernel function, steps are as follows:
One, data set is collected, each sample in the data set record of collection is analyzed, distinguishes sample different attribute, really
Determine input and output sample;
Two, kernel function is selected and constructed, the exponential distribution global kernels and Radial basis kernel function of Mercer condition will be met
It is mixed, obtains novel mixed kernel function:
In formula, parameter t is ratio shared by two kinds of kernel functions, and σ is the core width of Radial basis kernel function, and γ is exponential distribution core letter
Several parameters, x, y are input and output sample pair in sample set;
Three, the parameter in novel mixed kernel function is optimized;
Four, C-SVC model is selected, the support vector cassification model based on novel mixed kernel function is established;
Five, classification prediction is carried out by the support vector cassification model established.
2. the svm classifier method according to claim 1 based on mixed kernel function, it is characterised in that: the optimization uses
The particle swarm optimization algorithm for introducing Gaussian mutation, optimizes three parameters σ, γ, t in novel mixed kernel function.
3. the svm classifier method according to claim 2 based on mixed kernel function, it is characterised in that: the introducing Gauss
The step of particle swarm optimization algorithm of variation, is as follows:
Step 1: setting parameter, population size value, maximum number of iterations value, local search ability C1, ability of searching optimum C2,
Inertia weight ω value limits particle search area and population flying speed;
Step 2: the position X of initialization particle (σ, γ, t)iWith speed Vi;
Step 3: calculating the fitness value of each particle;
Step 4: determining global extremum and individual extreme value;
Step 5: iteration optimizing, judges whether to reach maximum classification accuracy, if so, turning to step 8, otherwise continue downward
It executes;
Step 6: according toWithThe speed of more new particle and position;
Step 7: one particle of random selection, Gaussian mutation is carried out to it, population diversity is kept, step 3 is gone back to;
Step 8: output optimal solution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811343420.4A CN109447178A (en) | 2018-11-13 | 2018-11-13 | A kind of svm classifier method based on mixed kernel function |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811343420.4A CN109447178A (en) | 2018-11-13 | 2018-11-13 | A kind of svm classifier method based on mixed kernel function |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109447178A true CN109447178A (en) | 2019-03-08 |
Family
ID=65552081
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811343420.4A Pending CN109447178A (en) | 2018-11-13 | 2018-11-13 | A kind of svm classifier method based on mixed kernel function |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109447178A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111192453A (en) * | 2019-12-30 | 2020-05-22 | 深圳市麦谷科技有限公司 | Short-term traffic flow prediction method and system based on Bayesian optimization |
CN111340061A (en) * | 2020-01-18 | 2020-06-26 | 中国人民解放军国防科技大学 | Multi-mode data fusion and classification method based on SVM model parameter optimization |
CN112861242A (en) * | 2021-03-17 | 2021-05-28 | 辽宁工程技术大学 | non-Gaussian fluctuating wind speed prediction method based on hybrid intelligent algorithm |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106874935A (en) * | 2017-01-16 | 2017-06-20 | 衢州学院 | SVMs parameter selection method based on the fusion of multi-kernel function self adaptation |
CN108615555A (en) * | 2018-04-26 | 2018-10-02 | 山东师范大学 | Colorectal cancer prediction technique and device based on marker gene and mixed kernel function SVM |
-
2018
- 2018-11-13 CN CN201811343420.4A patent/CN109447178A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106874935A (en) * | 2017-01-16 | 2017-06-20 | 衢州学院 | SVMs parameter selection method based on the fusion of multi-kernel function self adaptation |
CN108615555A (en) * | 2018-04-26 | 2018-10-02 | 山东师范大学 | Colorectal cancer prediction technique and device based on marker gene and mixed kernel function SVM |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111192453A (en) * | 2019-12-30 | 2020-05-22 | 深圳市麦谷科技有限公司 | Short-term traffic flow prediction method and system based on Bayesian optimization |
CN111340061A (en) * | 2020-01-18 | 2020-06-26 | 中国人民解放军国防科技大学 | Multi-mode data fusion and classification method based on SVM model parameter optimization |
CN112861242A (en) * | 2021-03-17 | 2021-05-28 | 辽宁工程技术大学 | non-Gaussian fluctuating wind speed prediction method based on hybrid intelligent algorithm |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wu et al. | Robust latent factor analysis for precise representation of high-dimensional and sparse data | |
Aburomman et al. | A novel SVM-kNN-PSO ensemble method for intrusion detection system | |
Mukhopadhyay et al. | Survey of multiobjective evolutionary algorithms for data mining: Part II | |
Vens et al. | Random forest based feature induction | |
Patel et al. | Efficient classification of data using decision tree | |
Liu et al. | Mining adversarial patterns via regularized loss minimization | |
WO2021109464A1 (en) | Personalized teaching resource recommendation method for large-scale users | |
CN109447178A (en) | A kind of svm classifier method based on mixed kernel function | |
CN112199536A (en) | Cross-modality-based rapid multi-label image classification method and system | |
Campos et al. | An unsupervised boosting strategy for outlier detection ensembles | |
CN103258210A (en) | High-definition image classification method based on dictionary learning | |
Chen et al. | Fuzzy kernel alignment with application to attribute reduction of heterogeneous data | |
CN110059756A (en) | A kind of multi-tag categorizing system based on multiple-objection optimization | |
Zhu et al. | Collaborative decision-reinforced self-supervision for attributed graph clustering | |
Xiong et al. | Diagnose like a pathologist: Transformer-enabled hierarchical attention-guided multiple instance learning for whole slide image classification | |
CN110008899A (en) | A kind of visible remote sensing image candidate target extracts and classification method | |
Yang et al. | RE-GCN: relation enhanced graph convolutional network for entity alignment in heterogeneous knowledge graphs | |
Chen et al. | Learning to segment object candidates via recursive neural networks | |
Chen et al. | Parameter selection algorithm of DBSCAN based on K‐means two classification algorithm | |
Biswas et al. | ECKM: An improved K-means clustering based on computational geometry | |
Coppersmith | Vertex nomination | |
Liu et al. | A weight-incorporated similarity-based clustering ensemble method | |
Xia et al. | GRRS: Accurate and efficient neighborhood rough set for feature selection | |
Cao et al. | A bootstrapping framework with interactive information modeling for network alignment | |
Manoju et al. | Conductivity based agglomerative spectral clustering for community detection |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190308 |