2
$\begingroup$

I am having difficulties in understanding the difference between bagging and downbagging. I understand that:

  • Downbagging is an extension of bagging where downsampling is used.
  • In downbagging, the whole minority class will be used as an input, together with a randomly selected subsample of the majority class.

So what's the difference between downbagging and bagging, especially in cases of imbalanced data sets? Will bagging choose the same number of samples from each class or it will choose a larger number of sample from the majority and a smaller from the minority?

$\endgroup$

1 Answer 1

1
$\begingroup$

, as such, does not care about class membership. Thus, you would expect "straightforward" bagging to use samples from each class at roughly the same frequency as the class' originally observed frequency.

Downbagging does not do random sampling. It thus is similar to the stratified bootstrap, which is somewhat more common. Reading up a bit on the stratified bootstrap may give you a better intuition for downbagging.

$\endgroup$
8
  • $\begingroup$ So the error resulted from having imbalanced data sets is expected to affect the performance as well when using bagging alone because it hasn't been addressed properly in this case.. right ? $\endgroup$
    – Ophilia
    May 18, 2016 at 13:10
  • $\begingroup$ Yes. "Affected" is a good word, because I don't think it's clear whether downbagging will improve your estimation/prediction in any given case. $\endgroup$ May 18, 2016 at 13:15
  • $\begingroup$ Sorry.. I am not sure I understand your last comment! it seems that downbagging is addressing the issue of imblanced datasets .. so in theory the performance should be improved when replacing bagging with downbagging right ? $\endgroup$
    – Ophilia
    May 18, 2016 at 13:24
  • $\begingroup$ @Ophilia if you wanted to have a larger sample from majority class and smaller from minority you talk about stratified sampling. So I think you understand the difference between those sampling methods right. But if you wanted to improve performance on imbalanced datasets you have to do some research. Sampling itself isn't just everything. For example, one common method is EasyEnsemble approach( cse.seu.edu.cn/people/xyliu/publication/tsmcb09.pdf ), in which we build an ensemble on various bags from majority class and a whole minority $\endgroup$ May 18, 2016 at 14:22
  • $\begingroup$ Stratified sampling "should" improve the model in unbalanced datasets, true. However, it will not necessarily do so. Plus, whether the "stratified" model is better than an unstratified one will depend on your notion of "better", which is not always clear-cut. E.g., in detecting rare illnesses, false positives or false negatives may be more problematic for different reasons. $\endgroup$ May 18, 2016 at 14:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.