1
$\begingroup$

Karatzoglou (2006) reviews SVM in R, but I see no mention of Cascade SVM, though obviously the article is dated.

My google searching hasn't led anywhere constructive. Is there an implementation of cascade SVM in R? If not what are recommendations for speedy SVM for multiclass problems for high dimensional data?

I've currently been building SVMs on a reduced dataset (~25k x 100p) but would like to work with datasets closer to (300-500k x 100p). When attempting to use radial basis kernels via kernlab::ksvm, I wasn't seeing convergence after a week for a dataset with roughly ~200k observations.

Any suggestions would be appreciated!

$\endgroup$
1
$\begingroup$

Google released sofia-ml, which allows for pairwise stochastic gradient descent SVMs, and it has been implemented in R (RSofia). SGD SVM should be very fast in most cases, as you can see in some of the references, such as Large Scale Learning to Rank, with improvements being on the order of many magnitudes.

Note that RSofia is not as full-featured as many of the other available SVM packages, so you'll have to do things like centering and scaling on your own.

$\endgroup$
  • $\begingroup$ Thanks--I was aware of Sophia, but not RSofia! $\endgroup$ – Alex W Dec 15 '15 at 20:35
  • $\begingroup$ In reading the documentation, it's unclear that Sofia/RSophia supports unranked multiclass classification problems. Can you confirm/deny? $\endgroup$ – Alex W Dec 16 '15 at 1:14
  • $\begingroup$ I don't specifically recall which paper, but I remember reading that Google deployed this in a multi-class environment. That being said, it wouldn't surprise me if the way this was accomplished was by manually coding a one-vs-all SVM for each class and then taking the highest-ranked prediction. Given the speed gained through using Sofia, that should solve your problem. $\endgroup$ – Tchotchke Dec 16 '15 at 16:20
  • $\begingroup$ Noted- that is one way that multiclass is solved after all, though the manual coding would be annoying. Please provide a reference / llink if you find the paper $\endgroup$ – Alex W Dec 16 '15 at 16:42

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.