"Cluster model for classification?"

harri678

Hello,

after reading some papers about Support Vector Clustering I tought this method is perfect for my type of work, but now it doesn't look like it.

I want to classify data as "normal" or "anormal", but only have "normal" labeled data for training. The thing i tried to do was to generate an SVC-model which contains one single cluster of normal data. After calculation i wanted to apply the generated clustermodel on the testing data which also contains anormal sets. I thought the model will calculate the affiliation to the normal-cluster, but it seems I am wrong. Is there any way to do this in RapidMiner beneath one class svm's?

Thx,
Harald

land

Hi Harald,
at least there's no such ready to go method I'm currently aware of. You might use simply a distance threshold to the nearest neighbor, but this isn't exactly the same. I could think of extending the SV Clustering in this way or implementing something like a normal distribution kernel based method, but if this would work depends on your data. Seems to be a rather complex setting

Greetings,
  Sebastian

harri678

Hi Sebastian,

thanks for the response!
It seems i have to look for other solutions of my problem.
I tried to use the libsvm's one-class in rapidminer-5.0-rc and there I have other problems (not shure if this is a bug).
When i use "Generate Data" to produce ExampleSets, the one-class libsvm fails with "numeric label not supported.". The same error is produced with other types of label. If I filter the label, it fails with "... does not have a label attribute.". In fact I wasn't able to get libsvm running for once. Do I have to do some XML tuning to prevent the label-madness or is there another way to train libsvm via rapidminer?

Greetings,
Harald