One-Class SVM Rho Value

[robin-tukl]

Hey Everyone,

I'm currently working on One-Class SVMs in Java and I came across the following problem.

When I finish training and for whatever reason try to classify my training data, it appears that only half of the samples become classified correctly.

`
wekaSVM = new LibSVM();

wekaSVM.setKernelType(new SelectedTag(LibSVM.KERNELTYPE_RBF, LibSVM.TAGS_KERNELTYPE));

wekaSVM.setSVMType(new SelectedTag(LibSVM.SVMTYPE_ONE_CLASS_SVM, LibSVM.TAGS_SVMTYPE));

wekaSVM.buildClassifier(dataInst);

double[][] normData = wekaSVM.distributionsForInstances(dataInst);
`

I have tested the same data with my Matlab Code(all parameter(nu = 0.5, cache size, epsilon) the same) and It seems that the model is similar (same SVs and coefficients) expect the bias. Here they differ from -10.52 (Matlab) to -42.89 (Java). Furthermore, I have checked the output of the Kernels and they also seem to match perfectly, but because of the different biases, I have a 100% accuracy on my training data for Matlab, but only 50% in Java.

Therefore my questions, is there an error in my setup and have anyone encountered similar issues?
Does the libsvm code finish as soon as it reaches the 50% mark (nu)?

Best regards
Yannick