1
$\begingroup$

I am classifying webpages based on several features of HTML-structure. I get the best cross-validation results with the RBF-kernel.

With a linear kernel, if I understand correctly, it is possible to attribute importance to certain features in the feature vector (How does one interpret SVM feature weights?). This would help in correcting the model for certain specific misclassifications (finding the features that 'are to blame' for the mistake). Am I correct when I say that this is not possible for RBF kernels?

If so, is there a good way of correcting the model for specific documents that get wrong classifications?

Thanks!

$\endgroup$
  • $\begingroup$ I think you can consider about the "weighted SVM" which assign specific weight for each data. SVM: Weighted Sample provides code and example to do it. You can assign high weights for specific documents that get wrong classification. However, it may effect the classification of other docs. $\endgroup$ – Ben Dai Feb 18 '16 at 10:30
0
$\begingroup$

You are correct, it is not possible to assign weights to features using an RBF kernel since you work on the dual. Thus, it is not possible to 'put the blame' on specific words.

However, you can potentially blame specific documents that SVM with RBF kernel identified as support vectors and make some noise removal at that level, i.e. remove whole documents. An easier alternative might be train the linear classifier, use it to filter out some noise (if you have a way to spot it easily), and then retrain an RBF if that gives you better performance.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.