I’m wondering how I should use TMVA with more than one background, especially when cross-sections for them differ significantly (few orders of magnitude). To complicate more big backgrounds are not a problem (I can eliminate them with few simple cuts) but those with small cross-sections are very problematic. Should I use weights to scale their relative proportions to correct values or just leave them as they are. I would greatly appreciate some explanation on this topic.
If you are using a 1 node output tree (or whatever)
Make the simple event selection you want.
Make the background sample you are using for training have the same background that you expect.
You can also train a net that has more than one output node (e.g., first node is signal versus background 1, second node is signal versus background 2, etc.).