When we train a BDT model using some variables, the TMVA output gives the correlation matrix. What should be the cut-off value of the percentage correlation below which the variable should be accepted in the training? What metric should I use to find that cut-off value, and how do I use it in ROOT TMVA?
There is no cut-off in TMVA on the input variable correlation. The BDT method is designed to deal with input variable correlations. However, it is better if you have 100% correlation, to remove in this case one of the input variable.