Plot final signal/background distributions, weight problem

TMVA
1

Dear Sir,

In my TMVA study, I have ~10 background, all background with different weighting (cross section size).

After TMVA, I wish to plot the signal and background distributions for the input variables, to see the TMVA CUT effect on the variables. I followed this : TMVA: Macro to plot final signal/background distributions?

My question is :
for all the background is added together, it seems the final distribution shown with i.e. " TestTree->Draw(“myvar1”,“LD>0.3”,"");" is plotted with the un-weighted # of events.

Because each contributed BG have different shape of distribution and weight (cross section),
I wish to know, how can I plot with background distribution with correct contribution from each background?

Thanks!

Best,
Jung

2

Hi,

See section 12.20.1 of the root users guide :slight_smile:

https://root.cern.ch/root/htmldoc/guides/users-guide/ROOTUsersGuide.html#simple-analysis-using-ttreedraw

Basically TestTree->Draw(“myvar1”,“weight*(LD>0.3)”); should work if the event weight is stored in a branch/leaf name weight.

Cheers,
Kim

3

Hi Kim,

Thanks for the reply. I had tried this. But the weight branch gave by TMVA is strange, i.e. I have totally 9000 events, I want total weight as 20.0, but in the weight branch, it shown as x axis is 22, with y axis 9000.

If I used TestTree->Draw(“myvar1”,“weight*(LD>0.3)”); it gave each event times 9000.

My main problem is, it used # of events for each background, not the weight I input.

I have more than 10 background. Each have different weights, if it used # of events for each background, it will give a wrong BG distribution. And because all BG are added in a BG branch. I cannot re-weight it again.

So I wish to know how to add weight to each BG, which gave a correct total BG distribution.

Thanks,

Jung

4

Hi,

Ok, thanks for the clarification. If all events have the same weight, then I suspect they are not propagated correctly from your input to TMVA.

Could you explain how you do it?

If you have a branch containing the weights for each event regardless of class using

dataloader->SetWeightExpression("branch_name_in_input_tree");

should do the trick, and the correct weight should end up in the final output file.

Cheers,
Kim

5

Hi Kim,

Thanks for the reply.

This is part of my BG and its weight :

Double_t bbjaWeight = 367000.0;

Double_t bbaaWeight = 140.0;

Double_t bbjjWeight = 434000000.0;

Double_t bbhWeight = 1.26;

Double_t tthWeight = 1.37;

Double_t ccjaWeight = 1050000.0;

dataloader->AddBackgroundTree( bbja_BG, bbjaWeight*3000.0*5.0/10000.0 *73273.0/3000000.0);

dataloader->AddBackgroundTree( bbaa_BG, bbaaWeight*3000.0*60253.0/3000000.0 );

dataloader->AddBackgroundTree( bbjj_BG, bbjjWeight*3000.0*56126.0/5000000.0*5.0/10000.0*5.0/10000.0);

dataloader->AddBackgroundTree( bbh_BG, bbhWeight*3000.0*666.0/1000000.0);

dataloader->AddBackgroundTree( tth_BG, tthWeight*3000.0*38478.0/1000000.0);

dataloader->AddBackgroundTree( ccja_BG, ccjaWeight *3000.0*10173.0/1000000.0*5.0/10000.0*1/8.0*1/8.0);

I’m thinking is it related to TMVA setting : NormMode, previously I set it as NumEvents. Maybe I should use NormMode=None?

(Actually I had tried with None, seems the output is no difference. I will delete previous output.root file, and try again.)

Thanks!

Best,

Jung

6

Hi,

That is how you should set the weights for the trees (and I verified that it works in the TMVAClassification example irrespective of the setting of NormMode).

I think some deeper debugging is needed to resolve your issue. Would it be possible for you to reduce your script to a minimal reproducer? (I forsee this will either help you find a bug in your code, or provide us with a concrete test case to debug!)

Cheers,
Kim

7

Hi Kim,

Thanks for the mail. I just used the default TMVAClassification file provided in ROOT/tutorials/tmva, and add my tree and weighting.

I attached the file. Thanks!

Best,

Jung

TMVAClassification.C (15.7 KB)

8

Hi,

I can’t run your script since I lack access to the input file but it seems fine from a TMVA perspective. Could you post a plot of the weight branch in the output file?

Cheers,
Kim

9

Hi Kim,

Thanks for the reply. I had already add two new reply, but it did not shown here.

Sorry I found I made a very stupid mistake. I thought the weight given at " dataloader->AddBackgroundTree( BG, Weight);", the Weight is the total weight. When I read the manual more carefully, I found its the weight of each event.

I modified the weight to each events, now everything is ok.

I have one naive question, if I want to use the weight I set to do the TMVA study, all I need is set NormMode=None?

And if NormMode= NumEvents, it means it will rescale BG and signal to Ns and Nb I set in “nTrain_Signal=0:nTrain_Background=0”, if I set it as 0, then the signal and BG will be weighted as 1/2 of total Ns and Nb.

If I set NormMode= EqualNumEvents, it means the signal and BG will be weighted equally.

Is my understanding of these NormMode setting correct?

And another question is can I have TMVA variables given in different root file?
i.e.
variable A from input1.root
variable B from input2.root

Sorry and Thanks!!

Best,
Jung

10

Hi,

Glad you figured it out! For you questions:

Cheers,
Kim

11

Hi Kim,

Thanks a lot for the reply! I got it!

Best,

Jung