Marco Stolpe recently publisheda nice process on the myExperiment platform showing how you can calculate the performance of a one-class SVM.
RapidMiner includes the One-Class SVM as a part of the LibSVM operator. The basic idea is to partition the data set as usual and to train the one class classifier only on one class inside the cross validation, but to test it on both classes for the part that was left out for testing. So for training, we have to remove all examples with the label we don't want to train the classifier on (Filter). As the SVM operator expects the nominal label values to consist of only one value, we need to map the nominal value "Mine" to "Rock". When we apply the one class model to our test data, we get "inside/outside" as a prediction. These values have to be mapped back to the original corresponding nominal values "Rock" and "Mine". Afterwards, we can use the standard performance operator.
Nice idea! Please note, however, that the data set (Sonar) is just a toy data set chosen for demonstration. More information can be found on the process web page at myExperiment .
The complete process can be downloaded with our Community Extension . The name of the process is "X-Validation with One-Class SVM".