Weights in TMVA Training

TMVA
1

Hello,

I am using TMVA BDT Classification. I need to train and test the method on discrete samples of signal and backgrounds. I have some questions related to assigning weights to properly mix the different background processes in appropriate proportions. Suppose we consider that there is one signal process and 3 background processes for a specific final state products. Could you tell if I am telling this right way to ensure that in my sample signal and background samples be in 1:1 ratio?
Firstly, we need to ensure that all the events in signal and backgrounds are at same luminosity. To do this, we will weight each event in the signal and background by (cross-section)*luminosity/(total_events).
Then in the training sample, suppose in the background we have N1 events for B1 process, N2 for B2 background and N3 for B3 at the same luminosity. Then we weight each of these to 1/N1, 1/N2 and 1/N3 respectively. Then, we let the TMVA take care of the normalisation by:
dataloader->PrepareTrainingAndTestTree(" “,” ",“NormMode = EqualNumEvents”);

Is doing the above correct and sufficient to mix signal and background in 1:1 ratio?

I was also reading in some queries related to weights, that it is sufficient to have SignalWeight=1 and BackgroundWeight=1 in the Classification method. Could you please explain why is it so and if I am missing something here?

Thanks and regards,
Antara

2

Hi,
If you are doing a binary classification using BDT it is not relevant the total weight of the signal and background events. What is important, in case you have multiple backgrounds, that they are normalised to the same luminosity, in order to have a background sample closer to the real data situation.

Cheers

Lorenzo

3

Hi,

Thanks for the prompt reply!
I am able to understand that with multiple background processes, it is important that the events are at same luminosity.
However, could you elaborate a bit more why in a binary classification using BDT, the total weights of signal and backgrounds is irrelevant? And also, what happens in the multiclassification case?

Thanks,
Antara