Dear fellow ROOT-ers and TMVA-users,
I am using the TMVA package to perform a classification problem and in particular, have been focusing on the various neural-net based methods to try and accomplish this.
Poking around the internet, I have heard several mentions of features that are not present in the TMVA users’ guide , such as “deep”-neural networks (DNN ), and additional cross-validation techniques besides the default train/test split (e.g. mentions of k-fold cross validation). Now, granted, the users’ guide is somewhat old (at least, the first version I find with a simple google search), but I do see in later ROOT versions (6.x.y onwards it seems) that there is the existence of “new” TMVA features like the MethodDNN (root/tmva/tmva/inc/TMVA/MethodDNN.h). It’s also worth noting that I didn’t see the mention of some of these features in the ROOT release notes, either.
Is there a current document detailing all of the new features that have been added to TMVA over the past few years?
Sorry… you are very right, the Users guide is somewhat ‘outdated’ and doesn’t mention the new Neural Net implementation nor the ‘cross validation’ that seems to have been added. Unfortunatly, as I wasn’t part of any of those two developments, I also wouldn’t be able to ‘fix that’ but I’ll try to encourage the corresponding authors to do so
Thank you very much for your rapid reply!
Yes, it would be very good for the sub-community of ROOT users who use TMVA to have access to up-to-date information; thank you (in advance) for getting in contact with the respective authors of these new features.
On that note, for fellow TMVA users who are following along with this thread, I have found a presentation  that discusses the current (as of May 4, 2016) status of TMVA. It is not explicit documentation for these new features, but knowing about their existence alone can be half the battle . Hopefully it is useful for people.
 indico.cern.ch/event/496146/con … rkshop.pdf