SVM contour of cross-validation accuracy

njasaj

Hi,
I am trying to solve a problem with Libsvm in Rapidminer. During  X-validation in ordrer to find the best combination of SVM parameters, I have logged the  C, Gamma, and accuracy measure and deviation. There are lot of parameters combination that leads to best accuracy and least deviation but after applying model on test data they will lead to very different answers. Is there any way to plot a contour of cross-validation accuracy like the one which is available in libsvm site (http://www.csie.ntu.edu.tw/~cjlin/libsvm/)? Or is there a way to select the best parameter combination?
thanks.

MariusHelf

Hi,

creating a contour plot that equals the one on the website is not possible, but you can plot C versus Gamma in a "scatter plot color" and put the accuracy onto the color. Since usually you optimize C and gamma on a logarithmic scale you should also set the axes of the plot to logarithmic.

Best regards,
Marius