How to select the C parameter for linear SVMs?
Hi,
I want to use the GridOptimization operator for parameter optimization but in what range should I search the C value for a SVM with linear kernel. Does the C depend on the count of the examples? I sometimes see values between 0 and 30000.
Thanks
I want to use the GridOptimization operator for parameter optimization but in what range should I search the C value for a SVM with linear kernel. Does the C depend on the count of the examples? I sometimes see values between 0 and 30000.
Thanks
Find more posts tagged with
Sort by:
1 - 3 of
31
Thanks for your answer. Since I want to use EqualLabelWeigting because of the class skew I need to use the JMySVM operator instead of LibSVM. Are there any differences concerning the performance of the learner and the application of the model?
Hi,
no, there are no differences in principle. The LibSVM, however, is already able to work on more than 2 classes in a classification setting. For the JMySVM, you have to wrap a Polynominal2Binominal meta learner around the SVM. Another thing: the kernel parameter sigma (JMySVM) corresponds to the parameter gamma (LibSVM). Please note that sigma is 1 / gamma.
Cheers,
Ingo
no, there are no differences in principle. The LibSVM, however, is already able to work on more than 2 classes in a classification setting. For the JMySVM, you have to wrap a Polynominal2Binominal meta learner around the SVM. Another thing: the kernel parameter sigma (JMySVM) corresponds to the parameter gamma (LibSVM). Please note that sigma is 1 / gamma.
Cheers,
Ingo
no general answer here but trying, sorry. It totally depends on your data. You could have a look into my Gecco 2007 paper to see how much C sometimes might vary for different data sets. as a rule of thumb, you could try (without parameter optimization first) a low value like and 0.0001 and higher values like 1000, 10000, and even 100000. Check if the runtime is ok for you and have a look at the predictive power. Then start with a rough grid parameter optimization to find a good range and start (at least) a second one to further optimize it.
Cheers,
Ingo