LIBSVM Tools has two handy scripts easy.py and grid.py where grid.py does feature selection for you.You can use it to determine optimal C and Gamma.

You can also use Grid Search along with Cross Validation to optimize the parameters for the best model.

Good Luck

Intuitively, the C parameter trades off mis_classification of training examples against simplicity of the decision surface. Low  value C tends to make decision surface smooth, while a high C tries all training examples correctly by giving the model freedom to select more samples as support vectors.

Gamma parameter defines how far the influence of a single training example reaches, with low values connote far and high values connote the neighborhood. 

The values must be selected on trail and error basis and as such , to my knowledge, there is no specific method available for selection of values. It is rather problem specific.

You can tray WEKA, this tool allows you do grid search.

See WEKA tutorial.

Since you choose a linear kernel, you just need to optimize the C parameter. This can be done using line search on a pre-defined range, let's say that you seek your optimal C in a range of values 2^(-10), 2^(-9), ..., 2^0, ..., 2^9, 2^10. If you choose to use the RBF kernel, then you need to (jointly) optimize another parameter, namely the gamma parameter, which refers to the variance of the corresponding Gaussian bell around support vectors. In this case you need to use grid search. In any case, you need to perform cross-validation, e.g., 3-Fold, 10-Fold, etc. 

Hi,

You can perform a cross-validated grid search for C and gamma.

Hi Heitor Scalco Neto,

1. In order to find the optimum values of C and gamma parameters, you need to perform grid search. And for performing grid search, LIBSVM contains readymade python code (grid.py), just use that to get the best of parameters and thus best of accuracy. For details of using grid search check the attached paper.

2. With Linear Kernels, you need only C parameter. And to find the optimum value of C, LIBSVM contain options so read the 'readme' file from the LIBSVM.

3. If you are using only linear kernels then why not use the LIBLINEAR library? LIBLINEAR uses Linear Kernels and is more efficient.

4. Remember: Linear Kernels (LIBLINEAR) is suitable for problems with very large number of features like document classification.

5. Focus of Linear Kenels(LIBLINEAR) is on the training time, i.e., they are faster but less accurate than non-linear Kernels (LIBSVM). On the other hand focus of LIBSVM is on accuracy but it is slower than LIBLINEAR.

Thanks

Vinod

Article A Practical Guide to Support Vector Classification Chih-Wei ...

Article LIBLINEAR: a library for large linear classification

Hi Mr.

If you use linear kernel, there is no need to estimate GAMA parameter and this parameter is used for RBF(Radial Basic Function). In any way, I suggest three way to estimate these parameter.

1. by trial and error.

2. If there are some researches near your topic which have been used of SVM in them, you can find the proper C and GAMA in these articles.

3. finally, I suggest using of GA(Genetic Algorithm) for optimization these parameters.

Good Luck

LIBSVM Tools has two handy scripts easy.py and grid.py where grid.py does feature selection for you.You can use it to determine optimal C and Gamma.

You can also use Grid Search along with Cross Validation to optimize the parameters for the best model.

Good Luck

In R, one can do SVM parameters tuning by giving their range in tune function as   list(epsilon = seq(0,10,0.05), cost = 2^(2:9) and then you can decide your best SVM classifier.

https://sites.google.com/site/kittipat/libsvm_matlab

If you're using Linear Kernel, I don't think C and Gamma matter much. However, if your results are not good, I guess the data isn't linearly separable. So using a Non-linear kernel such as RBF or polynomial might be useful.

All the best.

With linear kernel, try tune.svm() from package e1071

The parameter C controls the trade off between errors of the SVM on training data and margin maximization (C = ∞ leads to hard margin SVM). Rychetsky (2001), page 82

"The factor C in (3.15) is a parameter that allows one to trade off training error vs. model complexity. A small value for C will increase the number of training errors, while a large C will lead to a behavior similar to that of a hard-margin SVM." Joachims (2002), page 40

First, you should determine that are the data linear separable or not? if they are not linear separable you need to chose a nonlinear kernel such as RBF, and/or e.g.

Take a look of this exemple : https://github.com/clemaitre58/RPiRDF : https://github.com/clemaitre58/RPiRDF/blob/e5caeeb5769d98ad99aefa31f3d16a44e225e46b/App/main.py#L127

This is a grid search for different gamma parameters.

Even the qs proposed two years ago, I am still motivated to present my prespectives regarding this qs as follows.

use best.svm() instead of svm() without giving a cost value