Rarity distribution for BDT is non-uniform

TMVA
1

Hi,

I’m training a BDT Classifier to reject background events of my data sample to increase the purity. I was reading the TMVA Users Guide and in section 3.1.13 it says:

Another useful aspect of the probability integral transformation (rarity) is the possibility to directly
visualise deviations of a test background (which could be physics data) from the training
sample, by exhibition of non-uniformity

So, I habilitate the “CreateMVAPds” flag and plotted Rarity to check. The result can be seen below.

image

What does it mean? And, how can I avoid it?

I’m passing as signal an MC sample with a cut in the signal region, and for background, I’m passing the wings of mass distribution far from the signal region.

Thanks!

2

Hi,
It looks then that the distribution from training and test BDT output is very different. You can easily check that by plotting the output distribution train and test ( case 4b in TMVAGui)

Lorenzo

3

Hi Lorenzo,

Looking TMVAGui 4b option I don’t see a big difference.

image

4

Hi,
This is very strange. Can you please share the macro and the input data file so I can try to reproduce this problem ?

Cheers

Lorenzo