[TMVA-users] TMVA BDT problem with multiple backgrounds

[Shabnam J. <ja...@fn...>]

Dear TMVA experts and users,

I am using TMVA BDT application for an analysis at CMS. I have come across a little bit strange behavior. and I need your help.


I am using tmva-V04-01-00 and I am training with multiple backgrounds:

factory->AddSignalTree    ( signal,     signalWeight );
      factory->AddBackgroundTree( bkgndwjets, bkgndwjetsWeight );
      factory->AddBackgroundTree( bkgndTtbar, bkgndTtbarWeight);
      factory->AddBackgroundTree( bkgndTchan, bkgndTchanWeight);
      factory->AddBackgroundTree( bkgndTchanb, bkgndTchanWeightb );
      factory->AddBackgroundTree( bkgndSchan, bkgndSchanWeight);
      factory->AddBackgroundTree( bkgndSchanb, bkgndSchanWeightb);
      factory->AddBackgroundTree( bkgndTWchan, bkgndTWchanWeight);
      factory->AddBackgroundTree( bkgndTWchanb, bkgndTWchanWeightb);
      factory->AddBackgroundTree( bkgndZJets, bkgndzjetsWeight);

When preparing Training tree

factory->PrepareTrainingAndTestTree( mycut, "nTrain_Signal=686:nTest_Signal=686:nTrain_Background=686:nTest_B
ackground=686:SplitMode=Random:NormMode=EqualNumEvents:!V" );

With splitmode random option, I see a small bias (within 2 sigma) in the final output. I see the same thing if I use more events in the background training and testing  (where I have lot more statistics) than signal events. In principle these two things should work but they dont, and I see a bias in the output distributions, which in the end becomes a 2 sigma difference between expected and observed limits as the data distribution seems to be shifted a little bit with respect to the background distribution.

What I want to do is to train  signal against multiple backgrounds. I have multiple background trees stacked in one root file and I want make sure that I pick right proportion of events from each background for testing and training, and that training and testing sets are completely independent of one another.

Can you please tell me what options I should use and what is the best way to do this.


Many thanks for your help.
shabnam

-- 






--------------------------------
Shabnam Jabeen
Brown University, RI
http://www-d0.fnal.gov/~jabeen/
-----------------------------------