Training One Class SVM using LibSVM

0
votes

I hope to use one-class SVM of LIBSVM to train a training samples so as to get a model. Then, I use the model to predict the new test data and the training data is same type or not. In the training process, I have some questions as follows:

  1. The training samples is all positive examples or not?
  2. Which kernel function can get better result,linear kernel or RBF kernel?
  3. What is the effect of nu's values to the model?
1
data-miningsvmlibsvm
This is not really a programming problem, so may be more appropriate to ask on Cross Validated (stats.stackexchange.com) – Bull

1 Answers

0
votes
  1. The class label is not used so training with negative examples isn't really a concept
  2. The best kernel will depend on the type of data you have. Easy enough to try two or more.
  3. According to Scholkopf's paper, nu is
    • "an upper bound on the fraction of outliers"
    • "a lower bound on the fraction of SVs".