Dear Researchers,
i trained one-class-classification (svmtrain) with 1400 features (H0). I want to test this model (H0) with two-class examples (H0 and H1) that the first class is H0 and the second class is H1. For testing, I have 300 examples for H0 and 300 for H1. When my test is finished my outputs predict labels are approximately are 1? Why? I think must be H1 negative in output because of the features different from H0. I used libsvm library for training and testing. I select -n with parameter 0.01.
My implementation code is attached for your information and the examples of H0 and H1 are uploaded in this link https://ufile.io/qh88t
thanks
https://ufile.io/qh88t
Mr. Ehsan Nowroozi,
Your question was a bit difficult to follow. What is meant by features?. Basically, we will train and test a model based on the number of examples but not on features. Features represent examples with a fewer a number of samples which are capable of distinguishing examples. In your case, there are two types of examples belongs to either class H0 or H1. First, a pattern recognition model was developed by 1400 examples (consisting both H0 and H1 examples) using SVMtrain. Later, conducted testing with remaining examples. Finally, most of the examples predicted class H1. If this is the case, there are a lot of reasons influences final prediction. One such case is data imbalance. For further reading follow the given link.
Article Learning from imbalanced data: Open challenges and future directions
Andreas Theissler
thank you so much for answering my question. well, I describe more about my problem.
I have to H0 (legitimate) and H1 (Malicious) that the dimension is 686 and 2000 rows. To clarify more, first I choose (H0) 1400 for training and 300 for cross-validation and 300 for testing. In training part, I trained 1400 of examples (H0) with one-class-classification. well, my model (svmtrain) has only one-class of H0. For the testing part, I test 300 examples that I have for testing and choose 300 of samples of H1 separately, but the output of (svmpredict) in predicted values are similar and label too. why? my data H0 and H1 are different but why in testing part are equal (predict values and labels)?
Note: I used -s2 and -n 0.01 and use RBF kernel (-t2) and I use Libsvm 3-17.
What is your suggestion? I think you told me trained all of 2000 but how to test it by H0 and H1 separately because all of the examples of H0 are trained.
another question is that please recommend me how many examples of H0 must be I choose for training and testing? the cross-validation for one-class is important of omitting the part of cross validation? how about scaling?
My implementation code is attached for your information and the examples of H0 and H1 are uploaded in this link https://ufile.io/qh88t
Hi,
Is the data have zero mean?
The one class svm is very sensitive to data distribution.
You must try to create a zero mean data either for training nor testing.
Hope it's help.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training