machine learning - Liblinear vs Pegasos -

machine learning - Liblinear vs Pegasos -

i have been looking fast linear svm library , came across 2 of of import ones liblinear , pegasos , paper presented of liblinear looks liblinaer outperforms pegasos. pegasos claims if info sparse works fast. pegasos came before there no comparing in it;s documentation.

so sparse info should take ?

as far know, sparse info handled fine both. question more on number of info points. liblinear has solvers both primal , dual, , these solve problem high precision without need tune parameters. pegasos or similar subgradient descent solvers (if want 1 of these, i'd recommend leon bottou's sgd) result dependant on initial learning rate , learning rate schedule, can tricky tune.

as rule of thumb, if have less 10k info points, i'd utilize liblinear (with primal solver), maybe 100k. above that, i'd consider using sgd if sense liblinear slow. if liblinear slower, prefer using means don't have think learning rate, learning rate decay , number of epochs.

btw, can compare these different solvers using framework scikit-learn, includes sgd, liblinear , libsvm solvers, or lightning, includes lot of solvers.

machine-learning svm liblinear