[achompas]

Sibling comment is impressed, but I'm skeptical that you've improved libsvm. Sorry, can't help it. :)

Can you speak a bit about how you optimized it? I took libsvm to be fairly optimized as far as SVM libraries go...

[waterside81]

Yeah, in terms of performance "improved" is the wrong term. We've tailored things more to our use cases, especially since we need feature weighting and multi classification, rather than simply binary classification.

[achompas]

Very cool! I geek out over SVM optimizations and tweaks--have you collected your work anywhere (post? HN comment?) by any chance?

[waterside81]

Not really, sorry, but we're adding a new API to be able to setup and configure your own svms. Stay tuned or follow us on twitter @repustate

[the_cat_kittles]

its possible he is using a specialized kernel or something, or some other edge use case that would make optimizations more feasible