What is coolest about svmlight in comparison with libsvm

kudkudak commented 9 years ago

http://svmlight.joachims.org/ - here is a list of awesome features. Should we include in the API any of those?

ktalik commented 9 years ago

I think this issue covers the same topic as #329

kudkudak commented 9 years ago

Not exactly, @lejlot would just list features, so it is just a part

ktalik commented 9 years ago

Of course -- mentioned issue is my task. Just added here for reference :)

lejlot commented 9 years ago

Ok, so let us make a short list:

svmlight should be faster, however we need to empirically test it, similarly to this document https://www.google.pl/url?sa=t&rct=j&q=&esrc=s&source=web&cd=2&ved=0CCgQFjAB&url=http%3A%2F%2Fwww.cs.upc.edu%2F~lluism%2Fseminaris%2Fsvmimpl.ps.gz&ei=mpD4VJDvBc7baNOKgMgN&usg=AFQjCNGjI6hvyJpEfY-f5tXmnjbyuqwqzw&sig2=v6oParTYRBuTUk1StW_ekA in general svmperf is one of the fastest methods [for linear svm]
existance of internal sample weighting, which in particular gives us class-weighting schemes (oh yeah, we should add "auto class weighting") so we can maximize balanced measures (not just accuracy)
I told about it while ago, but maybe it is a time to get back to details... svmlight performs nice estimation of error called XiAlpha-estimate and returns them from training. It would be nice if we could include them in the summary. It also estimates the Vapnik-Chervonenkis dimension of the model (VCdim field from the output) - it would be nice to gather it too.
I would include the optimizer options in our API -q, -n, -m, -e, -h, -f, -#, as this makes this code much more tunable than libsvm
It would also be nice to be able to use "Transductive SVM" feature, as this is truly something unique

gmum / gmum.r