[MRG] FIX ignore null weight when computing estimator error in AdaBoostRegressor#14294
Conversation
|
BTW, this seems related to #14286 which investigates a similar issue with SVM. |
ede568e to
6902eff
Compare
|
ping @rth @jeremiedbb I think that you are the best to review this bug fix. |
jeremiedbb
left a comment
There was a problem hiding this comment.
Not using points with zero weight to compute the error seems fair.
Below are some comments.
2f27d59 to
810c637
Compare
jeremiedbb
left a comment
There was a problem hiding this comment.
besides minor comments, LGTM.
Co-Authored-By: jeremiedbb <34657725+jeremiedbb@users.noreply.github.com>
|
@adrinjalali I'm pinging you aggressively ;) |
adrinjalali
left a comment
There was a problem hiding this comment.
I need to double check a few things before I can be sure about its correctness and completeness. But I was wondering if it'd be easier to mask the samples and the sample weight right at the beginning of the fit, and continue with all of what's left.
P.S. thanks for the very aggressive ping :P
Nop because the bootstrapping will lead to different results since the input data will be different due to masking. |
|
@adrinjalali Any other comments? |
adrinjalali
left a comment
There was a problem hiding this comment.
Other than the small nit, LGTM, thanks @glemaitre
|
Ready to be merged @adrinjalali @jeremiedbb |
|
Thanks @glemaitre :) |
AdaBoostRegressor suffers from a bug where the error was normalized using the max of the absolute error on all prediction even if they were corresponding to a null-weight in
sample_weigth