SVM参数意义

参考网站 http://blog.csdn.net/shenziheng1/article/details/54178685

1. >>help svmtrain

SVMSTRUCT = svmtrain(TRAINING, Y) 

trains a support vector machine (SVM)  classifier on data taken from two groups. TRAINING is a numeric matrix  of predictor data(TRAINING是预测数据的一个数阵). Rows of TRAINING correspond to observations(TRAINING的行数代表样本数); columns  correspond to features(列数代表特征的维数). Y is a column vector that contains the known  class labels for TRAINING(Y是列向量，里面存着TRAINING的分类标签). Y is a grouping variable, i.e., it can be a  categorical, numeric, or logical vector; a cell vector of strings; or a  character matrix with each row representing a class label (see help for  groupingvariable). Each element of Y specifies the group the  corresponding row of TRAINING belongs to. TRAINING and Y must have the  same number of rows. SVMSTRUCT contains information about the trained  classifier, including the support vectors, that is used by SVMCLASSIFY  for classification(SVMSTRUCT结构体中包含了训练好的分类器的所有参数，包括支持向量，这些支持向量也用于对测试集进行分类). svmtrain treats NaNs, empty strings or 'undefined'

values as missing values and ignores the corresponding rows in  TRAINING and Y.

2.分类机的参量选择svmtrain

********************************************************************************************

'kernel_function'  A string or a function handle specifying the  kernel function used to represent the dot  product  in a new space. The value can be one of  the following:

 'linear'  - Linear kernel or dot product  (default). In this case, svmtrain  finds the optimal separating plane  in the original space.

 'quadratic'  - Quadratic kernel（二次核函数）

 'polynomial' - Polynomial kernel with default  order 3. To specify another order,  use the 'polyorder' argument.（多项式核函数，默认是3阶，如果需要提升，在‘polyorder’进行参数设置）

 'rbf' - Gaussian Radial Basis Function  with default scaling factor 1. To   specify another scaling factor,   use the 'rbf_sigma' argument.（高斯径向核函数，默认核宽为1，在‘rbf_sigma’可以进行参数设置）

 'mlp' - Multilayer Perceptron kernel (MLP)  with default weight 1 and default  bias -1. To specify another weight  or bias, use the 'mlp_params'  argument.（多层感知核函数，默认权重1，偏好-1）

**************************************************************************************************************************

'rbf_sigma' A positive number specifying the scaling factor  in the Gaussian radial basis function kernel.  Default is 1.（设置高斯径向核函数的核宽）

**************************************************************************************************************************

'polyorder' A positive integer specifying the order of the  polynomial kernel. Default is 3.（设置多项式核函数的阶数）

**************************************************************************************************************************

'mlp_params'  A vector [P1 P2] specifying the parameters of MLP  kernel.  The MLP kernel takes the form:  K = tanh(P1*U*V' + P2),  where P1 > 0 and P2 < 0. Default is [1,-1]. （设置多层感知核函数的权重和偏好）  

***************************************************************************************************************************

'method' A string specifying the method used to find the  separating hyperplane. Choices are:（采用指定的方法寻找分类超平面）

'SMO' - Sequential Minimal Optimization (SMO)  method (default). It implements the L1  soft-margin SVM classifier.（序列最小优化算法）

'QP'  - Quadratic programming (requires an  Optimization Toolbox license). It  implements the L2 soft-margin SVM  classifier. Method 'QP' doesn't scale  well for TRAINING with large number of  observations.（二次规划）

'LS'  - Least-squares method. It implements the  L2 soft-margin SVM classifier.（最小平方方法）

***************************************************************************************************************************

  'options' Options structure created using either STATSET or  OPTIMSET. * When you set 'method' to 'SMO' (default), create the options structure using STATSET.  Applicable options:

'Display'  Level of display output.  Choices  are 'off' (the default), 'iter', and  'final'. Value 'iter' reports every  500 iterations.

'MaxIter'  A positive integer specifying the  maximum number of iterations allowed.  Default is 15000 for method 'SMO'.

* When you set method to 'QP', create the  options structure using OPTIMSET. For details  of applicable options choices, see QUADPROG  options. SVM uses a convex quadratic program, so you can choose the 'interior-point-convex'  algorithm in QUADPROG.

**************************************************************************************************************************

'tolkkt' A positive scalar that specifies the tolerance with which the Karush-Kuhn-Tucker (KKT) conditions  are checked for method 'SMO'. Default is  1.0000e-003. （‘method’='SMO'时；KKT迭代收敛条件） .

************************************************************************************************************************** 

'kktviolationlevel'   A scalar specifying the fraction of observations  that are allowed to violate the KKT conditions for  method 'SMO'. Setting this value to be positive  helps the algorithm to converge faster if it is   fluctuating near a good solution. Default is 0.

************************************************************************************************************************** 

'kernelcachelimit'    A positive scalar S specifying the size of the   kernel matrix cache for method 'SMO'. The

algorithm keeps a matrix with up to S * S  double-precision numbers in memory. Default is  5000. When the number of points in TRAINING  exceeds S, the SMO method slows down. It's  recommended to set S as large as your system  permits.(核函数内存空间设置).

 ************************************************************************************************************************** 

'boxconstraint'  The box constraint C for the soft margin. C can be  a positive numeric scalar or a vector of positive  numbers with the number of elements equal to the  number of rows in TRAINING.   Default is 1.

* If C is a scalar, it is automatically rescaled  by N/(2*N1) for the observations of group one,  and by N/(2*N2) for the observations of group  two, where N1 is the number of observations in  group one, N2 is the number of observations in  group two. The rescaling is done to take into  account unbalanced groups, i.e., when N1 and N2  are different.

 * If C is a vector, then each element of C  specifies the box constraint for the  corresponding observation.

************************************************************************************************************************** 

 'autoscale' A logical value specifying whether or not to  shift and scale the data points before training.   When the value is true, the columns of TRAINING  are shifted and scaled to have zero mean unit  variance. Default is true(数据归一化。默认是打开的).

************************************************************************************************************************** 

'showplot' A logical value specifying whether or not to show  a plot. When the value is true, svmtrain creates a  plot of the grouped data and the separating line  for the classifier, when using data with 2 features (columns). Default is false(分类结果显示，仅适用于特征空间是两维情况).

 ********************************************************************************************

3.SVMSTRUCT中包含哪些信息？

SupportVectors  Matrix of data points with each row corresponding to a support vector（支持向量）. 

Note: when 'autoscale' is false, this field  contains original support vectors in TRAINING.  When 'autoscale' is true, this field contains  shifted and scaled vectors from TRAINING.

Alpha  Vector of Lagrange multipliers for the support  vectors. The sign is positive for support vectors   belonging to the first group and negative for  support vectors belonging to the second group(拉格朗日向量).

Bias  Intercept of the hyperplane that separates  the two groups(超平面截距).

Note: when 'autoscale' is false, this field  corresponds to the original data points in  TRAINING. When 'autoscale' is true, this field  corresponds to shifted and scaled data points.

KernelFunction  The function handle of kernel function used（核函数）.

KernelFunctionArgs   Cell array containing the additional arguments  for the kernel function （核函数参量） .

GroupNames   A column vector that contains the known  class labels for TRAINING. Y is a grouping  variable (see help for groupingvariable).

SupportVectorIndices A column vector indicating the indices of support  vectors.

ScaleData   This field contains information about auto-scale. When 'autoscale' is false, it is empty. When   'autoscale' is set to true, it is a structure containing two fields:

                         shift       - A row vector containing the negative

                                       of the mean across all observations

                                       in TRAINING.

                         scaleFactor - A row vector whose value is

                                       1./STD(TRAINING).

FigureHandles   A vector of figure handles created by svmtrain   when 'showplot' argument is TRUE.

4.默认形式&参数设置对比

默认形式：

[cpp]  view plain  copy

1. <span style="font-size:18px;">svmStruct = svmtrain(data(train,:),groups(train),'showplot',true);  
2. classes = svmclassify(svmStruct,data(test,:),'showplot',true);  
3. classperf(cp,classes,test);  
4. cp.CorrectRate;</span>  

  

参数选择形式：

[cpp]  view plain  copy

1. <span style="font-size:18px;">svmStruct = svmtrain(data(train,:),groups(train),...  
2.   <span style="white-space:pre">    </span>    'Kernel_Function','rbf','RBF_Sigma',5,'showplot',true);  
3. classes = svmclassify(svmStruct,data(test,:),'showplot',true);  
4. classperf(cp,classes,test);  
5. cp.CorrectRate;</span>  

注意：参数名称和后面的修改值必须要一一对应！