Weights in TMVA

Newbie
1

I am doing a binary classification with BDT (AdaBoost) from TMVA.
I have S/B events in equal amount (say 1M each) for the training.
But in the simulation data, S/B are actually in the ratio of 1/9.
In order for the trained model to reflect this S/B distribution, do the following 3 methods work?

method 1: with equal amount of S/B, use weights
factory->AddBackgroundTree(SigTree, 1.0);
factory->AddBackgroundTree(BkgTree, 9.0);

method 2:
reselect the events from S/B tree so that they have the correct proportion
but this ends up with imbalanced data, doesnt it ?

method 3:
reselect the events from S/B tree so that they have the correct proportion.
then use weights
factory->AddBackgroundTree(SigTree, 1.0);
factory->AddBackgroundTree(BkgTree, 9.0);

btw, I am also not quite understand how to use these weights. could somebody explain to me ?

Thanks in advance

_ROOT Version: 6.24 (PyROOT via conda)
_Platform:Centos7
_Compiler: gcc9

2

Hi,
I am not sure I have understood correctly your questions. What do you mean with in the simulation data, S/B are actually in the ratio of 1/9 ? Is this the actual ration expected in the real data reflecting the different production cross section and selection ?

For the training you should use the maximum statistics available, you can use method1 but if you have a single background, it will not matter.
Then for finding the optimal selection efficiency, you should maximise a metric, for example S/sqrt(S+B), given the expected number of events in the data.

Cheers

Lorenzo