A variable with high separate power doesn't improve ROC curve

TMVA
1

Dear expert,

I’m currently use BDTG to for classification.
we use 8 variables to discriminate signal and background, I try to add a new variable.
The new variable has the highest separation power and highest variable importance among these 9 variables, but the ROC curve doesn’t improve at all, what could be the issue?
From the Gui, the smirnov test shows BDTG is not overtrained.

Thanks in advance,
Binghuan

2

Dear Binghuan,

There are various possible reasons for this, among which is that the individual separation power of a single variable is not an indicator of the overall classifier performance made from a collection of variables. It’s possible that the information provided by other variables standalone makes up for the individual performance of this variable.

Best,

Sergei

3

Dear Sergei,

Thank you for a prompt reply.
I also checked the correlation between the new variable and other 8 variables, they all are less than 20 %.
So I doubt the information provided by other variables can make up for this variable.
Are there any other reasons to explain this or any checks I can do?

Thanks,
Binghuan