SVM Learning Notes (1) libsvm in MATLAB under the use of installation _ machine learning

1 from LIBSVM's official website http://www.csie.ntu.edu.tw/~cjlin/libsvm/Download the latest version of the LIBSVM, the current version of the Libsvm-3.16.zip

2 decompression and compression package to the computer on a location, such as: C:\Program files\libsvm-3.16

3 assume that you are using a 64-bit operating system and MATLAB. At this point, add the Windows directory under the Libsvm-3.16 folder to the MATLAB directory. The MATLAB directory is added: C:\ProgramFiles\libsvm-3.16\windows. Because of Windows contains MATLAB executable binary file Libsvmread.mexw64/libsvmwrite.mexw64/svmpredict.mexw64/svmtrain.mexw64

4 If you are using a 32-bit operating system and MATLAB, you need to compile the corresponding binaries.

In the MATLAB command window input

>>mex–setup

At this time Matlab will prompt you to choose to compile the Mex file of C + + compiler.

Select one of the C + + compilers installed on your computer, such as Microsoft Visual C + + 2010

The current directory of MATLAB set to: CD ' C:\Program Files\libsvm-3.16\matlab '

Enter command make

At this point you will see that the current directory generates binary file LIBSVMREAD.MEXW32/LIBSVMWRITE.MEXW32/SVMPREDICT.MEXW32/SVMTRAIN.MEXW32. Add the current directory to the MATLAB path.

5 Example a test

Note that there is a data file under Libsvm-3.16, named Heart_scale. This is a LIBSVM format data file. You can use the Libsvmread function to convert it to MATLAB format.

You can use the following code to test:

[Heart_scale_label,heart_scale_inst]=libsvmread (' Heart_scale ');

Model = Svmtrain (Heart_scale_label,heart_scale_inst, '-C 1-g 0.07 ');

[Predict_label, accuracy, dec_values] =svmpredict (Heart_scale_label, heart_scale_inst, model); % test the Trainingdata

6) Svmtrain function Related parameter description

The model returned by the Svmtrain function can be used to predict the test dataset. This is a structural variable, consisting mainly of the following domains. [Parameters, Nr_class, TOTALSV, Rho, Label, ProbA, Probb, Nsv,sv_coef, SVs]. The English description is as follows:

-parameters:parameters

-nr_class:number of classes; = 2 for Regression/one-class SVM

-totalsv:total #SV

-rho:-B of the decision function (s) wx+b

-label:label of each class; Empty for Regression/one-class SVM

-proba:pairwise probability information; Empty if-b 0 or in ONE-CLASSSVM

-probb:pairwise probability information; Empty if-b 0 or in ONE-CLASSSVM

-nsv:number of SVs for each class; Empty for Regression/one-class SVM

-sv_coef:coefficients for SVs in decision functions

-svs:support vectors

If the '-B 1 ' option is not specified, the Proba and Probb are empty matrices. In addition, when the '-V ' option is specified, the returned model is a numeric value, which is the accuracy rate of the cross-validation.

Where the model.paramter is a 5x1 vector, the parameter meaning is:

Model. The parameters parameter meaning is from top to bottom:

-S SVM Type: SVM settings type (default 0)

-T kernel function type: kernel function setting type (default 2)

-D Degree: degree setting in kernel function (for polynomial kernel function) (default 3)

-G R (GAMA): Gamma function setting in kernel function (for polynomial/rbf/sigmoid kernel function) (Reciprocal of default category number)

-R COEF0: COEF0 setting in kernel function (for polynomial/sigmoid kernel function) (default 0)

7) svmpredict function parameter description

The Svmpredict function returns three values, Predict_label, which is the label vector that the training set predicts. The second output is accuracy, is a 3-D vector, from top to bottom is: Classification rate (the parameters used in classification problems), mean square error (MSE (mean squared error)) (regression problem used in the parameter index); the square correlation coefficient (R2 ( Squared correlation coefficient)) (parameter indices used in regression problems). The third output is a matrix that contains the decision value or probability estimate (when ' B 1 ' is specified). When the training data has a K class, the decision value matrix is an n row k* (k-1)/2 column matrix (n is the number of test datasets, K is the category), and the output of each row is the result of the k* (k-1)/2 two classifier. When '-B 1 ' is specified, the probability estimate matrix is a n-row K-Class matrix (n is the number of test datasets, K is a category), and the output of each row is the probability that the test data belongs to each class.