About the TMVA: How to give weight factor to Signal and Background tree

Dear all,

We are using the TMVA code namely: TMVAClassification.C with BDT method.

We have created two root TTrees samples with 2000 events for each Signal and Background. These TTrees included a few input variables (cuts) to distinguish between signal and background.

We want to know if there is a way to give weights to signal and background trees. In the above mentioned code: TMVAClassification.C, we found the functions for example:
Double_t signalWeight = 1;
Double_t backgroundWeight = 1;
dataloader->AddSignalTree ( signalTree, signalWeight );
dataloader->AddBackgroundTree( background, backgroundWeight );

but we don’t see any differenece if we change values of weights like:
Double_t signalWeight = 10000;
Double_t backgroundWeight = 1;
dataloader->AddSignalTree ( signalTree, signalWeight );
dataloader->AddBackgroundTree( background, backgroundWeight );
This doesn’t have any effect on results. Why?

Is there a way to give weights to signal and background trees?
and also could you please let us know how these weight factors are being used in the TMVA program?

What we like to have is to simulated a cross-section for signal and background.

Thanks,
Vipul Bairathi

Hello,

Let’s ping @moneta , he’ll be able to give you an answer (please bear in mind that people might be slower to answer during holidays).

Hi,

You can give an overall weight for the signal and the background tree. This is useful if you have multiple datasets (trees), for example multiple background contributions, and you need to apply a weight for them to consider the relative cross sections.
If you have a single signal and background tree, applying that weight it will not have any effect on the result

Lorenzo

Dear Lorenzo,

Thank for the your reply.

We have only one tree for each signal and background, which contain e.g 1000 entries for our particle of interest (let say phi meson) in both trees.

What we need is to tell TMVA to use appropriate weightage for signal and background candidate.

Since we have generated the signal and background trees from simulation assuming same production (e.g. 1 particle per event), but in real experiment data the signal and background cross-section is different (more sources of background then the signal).

Is there a way to apply weightage to signal and background using TMVA or alternatively if you can suggest that we change our input trees somehow to able to put weightage.

Thanks,
With regards,
Vipul Bairathi

Hi,
TMVA will train an algorithm to distinguish between signal and background. You will obtain from the simulated events an efficiency for signal and background events. Then to find the optimal working point you will need to consider the real composition of signal and background in your data, but this is not relevant when training the algorithms

Lorenzo