search

holub008 / xrf

eXtreme RuleFit (sparse linear models on XGBoost ensembles)
Other
43 stars 13 forks source link

addition of pmax and weights parameters, plus parallel option on glmnet #1

Closed yama1968 closed 5 years ago

yama1968 commented 5 years ago

Great, thanks very much!

Now with weights it works pretty well on unbalanced datas. You only have to set weights to 1 for the majority class, to 1/mean(y) to the minority class, for the training.

Best Regards, Yannick

Le lun. 25 mars 2019 à 21:37, Karl Holub notifications@github.com a écrit :

Merged #1 https://github.com/holub008/xrf/pull/1 into master.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/holub008/xrf/pull/1#event-2227812480, or mute the thread https://github.com/notifications/unsubscribe-auth/AHG3n63u0PPK6cMH8Rn2y_DOKLVHeJMPks5vaTObgaJpZM4cJFu_ .

holub008 commented 5 years ago

Great! I tend to work with cases where posterior calibration is important, so don't use weighting. Glad to have it nonetheless.

FYI I recently found a bug in xgboost split value reporting (see https://github.com/holub008/xrf/issues/2) which may be weakening the power of any rulesets and models you build. A quick fix for now is to set sparse=FALSE in xrf()

yama1968 commented 5 years ago

Thanks for the fix, I'll try that! Yannick

Le mar. 26 mars 2019 à 07:43, Karl Holub notifications@github.com a écrit :

Great! I tend to work with cases where posterior calibration is important, so don't use weighting. Glad to have it nonetheless.

FYI I recently found a bug in xgboost split value reporting (see #2 https://github.com/holub008/xrf/issues/2) which may be weakening the power of any rulesets and models you build. A quick fix for now is to set sparse=FALSE in xrf()

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/holub008/xrf/pull/1#issuecomment-476495782, or mute the thread https://github.com/notifications/unsubscribe-auth/AHG3n7pCvu0R_xh_sCPk-tj5OVnZ096Jks5vacGogaJpZM4cJFu_ .