How can I use event-wise weights in the TMVA factory?

[ChrisVer]

I have one question (simple I hope).
In the TMVA factory one can set global weights by typing:

// global event weights (see below for setting event-wise weights)
Double_t signalWeight = 1.0;
Double_t backgroundWeight = 1.0;

And use them afterwards by calling their names signalWeight, backgroundWeight.

If I want to use event-wise weights, the code below uses the factory->SetBackgroundWeightExpression etc:

factory->SetWeightExpression("EventWeight");
factory->SetSignalWeightExpression("EventWeight");
factory->SetBackgroundWeightExpression("EventWeight");

// This would set individual event weights (the variables defined in the
// expression need to exist in the original TTree)
// for signal : factory->SetSignalWeightExpression("weight1*weight2");
// for background: factory->SetBackgroundWeightExpression("weight1*weight2");

I was wondering, later on how can I use those weights? Should I put this whole command in position where "signalWeight" should be used?

Thanks. (https://hep.pa.msu.edu/wiki/pub/AtlasSingleTop/AnalysisVersion1405006Check/TMVAnalysis_test3.C)

 

[mfb]

Well, you use elements of the tree, so you should be able to access them like everything else in the tree if you have to. GetEntry, GetEvent or whatever you like as method, together with a branch address set somewhere.

 

[ChrisVer]

together with a branch address set somewhere.

What do you mean by that?

 

[ChrisVer]

In my tree I have the histogram of weights, let's say it's called: "weight".
If I try to write something like:

//BckgTraining is a string
factory->AddBackgroundTree(BckgTraining,
                           factory->SetBackgroundWeightExpression("weight") , "Training");

I am getting an error because the AddBackgroundTree expects to get a Double_t as a 2nd argument.
Or you mean that I could write:
weight->GetEntry ?

 

[mfb]

How do you access things in your tree in general?
There are multiple ways to do that.

I don't understand what your code example is supposed to do. The set method will probably return an integer as success message, and should be called exactly once to setup the readout.

 

[ChrisVer]

It's pretty much the same as in the example of the link (https://hep.pa.msu.edu/wiki/pub/AtlasSingleTop/AnalysisVersion1405006Check/TMVAnalysis_test3.C)

My tree is in a root file, I open the file and use the commands

TTree *signalTraining = (TTree*)inputTraining->Get("TMVAInputTree");
TTree *backgroundTraining = (TTree*)inputTraining->Get("TMVAInputTree");

TTree *signalTesting = (TTree*)inputTesting->Get("TMVAInputTree");
TTree *backgroundTesting = (TTree*)inputTesting->Get("TMVAInputTree");

The signal/bckg variable distributions are registered with AddVariable() or AddSpectator().

The thing is that when I read about weighting by event it gives that command with factory...

What I want to do... In the tree I have the signal and background events together with the weight histogram I want to use. I want to run the training/testing with weighting the background to the signal. In that case I could use a global signal weight:

Double_t sig_weight = 1.0;

But I cannot use the global background weight 1.0... Instead I want to use the weight histogram values.

factory->AddSignalTree( signalTraining, sig_weight, "Training" );
factory->AddSignalTree( signalTesting, sig_weight, "Test" );

factory->AddBackgroundTree( backgroundTraining, backgroundWeight, "Training" );
factory->AddBackgroundTree( backgroundTesting, backgroundWeight, "Test" );

 

[mfb]

Looking at the source, SetBackgroundWeightExpression sets some variable somewhere. There are so many AddTree methods that I don't see where something is actually done (probably in line 399, but where is this hidden?).

The background weight is optional - what happens if you just omit it?
What happens if you plug in a number, but call SetBackgroundWeightExpression before (compared to not calling it)?

 

[ChrisVer]

The background weight is optional - what happens if you just omit it?

I don't think I want to omit it... If I do then there will be some differencies between the variables because they are correlated to the momenta. In particular the weight I got, was by trying to make the background and signal Pt distributions look the same.
I tried doing that with a simple number (with that call and a background weight=1.0) and OK, my BDT ran normally , but didn't get what I started for to get (it didn't work). I know it didn't because if it did, the resulted Pt variable distributions would look similar which wasn't the case.

Looking at the source, SetBackgroundWeightExpression sets some variable somewhere. There are so many AddTree methods that I don't see where something is actually done (probably in line 399, but where is this hidden?).

The one leads to the other... The:
AddBackgroundTree() (line 431) leads to AddTree (line 367 ) and that to AddTree of line 384 . I am not sure what is that in line 399...

 

[mfb]

I don't think I want to omit it... If I do then there will be some differencies between the variables because they are correlated to the momenta. In particular the weight I got, was by trying to make the background and signal Pt distributions look the same.

Omit it in the call of the method, not in terms of physics. Ideally it takes the weight you declared before.

I tried doing that with a simple number (with that call and a background weight=1.0) and OK, my BDT ran normally , but didn't get what I started for to get (it didn't work). I know it didn't because if it did, the resulted Pt variable distributions would look similar which wasn't the case.

Was the output different from the code where you just used a weight of 1 without the other method before?

 

[ChrisVer]

Was the output different from the code where you just used a weight of 1 without the other method before?

Hmmm...I just tried it and no they are different... If I understood how that method works, then probably I'd figure out the error.