Wrong TMVA response on the TMVA Rectangular Cuts Method

TMVA
1

I am using TMVA Rectangular Cuts Method.
So here it is my xml file after training.
screenshot-20171206-23:06:33
I want to get tmva response from TMVA::Reader when testing trained method. So I write (inside a loop over events). NOTE: I test a BACKGROUND tree, the same as I used for training. So I am interested in the background efficiency.

if ( ! reader->EvaluateMVA( "Cuts", 9.8e-1 ) ) { continue; }

Here I use another method PrintCuts to confirm parameters of this ROC point:
screenshot-20171206-23:18:20
Then I have actual efficiency which is equal to (surprise, surprise) 0.4267. Why not something about 2.36e-1? But if instead of 9.8e-1 I write 9.799999 I get (surprise, surprise) 0.231969. So what happens here? Is it a bug?

2

Hi,

I’ve been looking into this issue a bit. My conclusion so far is that there is an of by one error somewhere in the histogram generation. Can you work around this for the time being?

One approach to getting the correct efficiency estimation would be to run the “Test Tree” through the classifier and calculating the corresponding score. The resulting value is what should be in the printout of PrintCuts.

Sorry for the inconvenience :confused:

Cheers,
Kim

3

Selbstverständlich, no problem.