I know one successful attempt one-class SVM but I'm rather interested in neural approaches over the problem like reconstruction networks or other fundamental neural methods. Are there any other functional examples?
I would recomend you to read the following PhD thesis and/or contact the author:
http://www.sc.ehu.es/ccwbayes/isg/administrator/components/com_jresearch/files/theses/borja.calvo.phd.dissertation.2008.pdf
Here is a toolbox with several classifiers: http://prlab.tudelft.nl/david-tax/dd_tools.html
OCSVM or One Class SVM is the good choice for your problem.
I've experienced good results with OCSVM for anomaly detection. In my case I had only positive data for training.
For working in a production environment I strongly recommend lib library instead of libsvm.
libsvm is poor spaghetti code and not thread safe.
See :http://dlib.net/ml.html
The only problem for finding best parameters in SVM is cross-validation which is not supported for one-class. The only way I can recommend is brute-force.
Cheers,
Hassan H. Monfared
- Badges
- Science topic