Does TMVA transform variables for application stage?

Hi all,

I found that BDT response shape is distorted severely in application stage ( reader.BookMVA() and then reader.EvaluateMVA() ) compared to the output histograms from TMVAGui, once I introduce several variable transformations. For example, tmva gui gives two gaussian-like overlapped histograms with median value of -0.3 and 0.3 for bkg and sig respectively, while the evaluated histograms are highly populated near +1.

This seems very strange because the user guide says
"Any preprocessing performed for the training is automatically also performed in the application through the Reader class".

In case of Keras, situation goes similarly. This can be a huge problem when I want to apply the classifier to the real data and set limit since the wierd shape can affect the limit values in case of shape analysis.

Does anyone have an idea about this problem?

@moneta maybe you can help here?

Hi Jiwon,

TMVA should apply the variable transformations also in the application phase. Did you try running the application with the train/test data that you used initially?

If you used a different data set, it might be that this second set has a different distribution than your training data. In this case you must manually compensate for the difference.