TMVA - Toolkit for Multivariate Data Analysis

---

Tutorial Preparation Local installation Using CERN afs Running an example training analysis Running the example analysis as a ROOT macro Running the example analysis as an executable The training phase Applying the trained classifiers to data The application phase using the *Reader* The application phase using *standalone C++ classes* Discussion Links TMVA Web Utilities

---

Tutorial

This tutorial treats a full TMVA training analysis and the application of the training results to the classification of data sets with unknown sample composition.

Preparation

The tutorial uses TMVA-v3.8.4 and has been tested for various ROOT 5 releases. Because remote ROOT graphics is slow it is preferred that the user runs the tutorial on a local computer (including a local installation of ROOT >=5 and TMVA-v3.8.4). Take care that the ROOTSYS environment variable is properly set. For non-local usage, it is assumed that the users have a CERN afs account (lxplus).

Local installation

It is assumed that ROOT >=5 has been installed (binaries are sufficient, they do include the necessary header files of all ROOT classes). If not, please consult the ROOT download page.

1. Download TMVA-v3.8.4 from Sourceforge.net
2. Unpack the tar.gz file:

   tar xfz TMVA-v3.8.4.tgz

1. Check the contents of the package:

   cd TMVA
   ls -l
   

   The top directory contains README and LICENSE text files as well as a setup script for bash (zsh) and (t)csh. Content of the subdirectories:
   • src: header and source files of all TMVA classes
   • include: symbolic link to directory containing include files (link created by setup script)
   • lib: contains shared library libTMVA.1.so after source code compilation
   • macros: example scripts for a training and application analysis, and all plotting scripts for the TMVA evaluation
   • examples: example executables for a training and application analysis
   • python: example python script for a training analysis
   • reader: obsolete
2. Setup the environment:

   source setup.[c]sh 

3. Create the TMVA shared library: libTMVA.1.so (this will take a few minutes)

   cd src
   make
   cd ../
   ls -l lib/
   

   Note that if you use ROOT >= 5.11/03, older versions of TMVA will also be part of the ROOT distribution. There should not be any conflict with the newly created package as long as libTMVA.1.so is properly loaded (or linked), as is done in the macros that are executed during this tutorial.

Using CERN afs

Use CERN afs on interactive lxplus with preinstalled TMVA version:

1. Login on lxplus.cern.ch
2. Setup ROOT (v5.14/00e, sys=amd64_linux26 on lxplus):

   cd /afs/cern.ch/sw/lcg/external/root/5.14.00e/slc4_amd64_gcc34/root 
   source bin/thisroot.[c]sh

3. Copy TMVA to a local area:

   cd ~/
   cp -rp /afs/cern.ch/atlas/offline/external/TMVA/TMVA-v3.8.4 .
   

4. Check the contents of the package:

   cd TMVA-v3.8.4
   ls -l
   

   ... and continue as above.

Running an example training analysis

The TMVA distribution comes with a idealised toy data sample consisting of four Gaussian-distributed, linearly-correlated input variables. The toy is sufficient to illustrate the salient features of a TMVA analysis, but it should be kept in mind that this data is by no means representative for a realistic HEP analysis. TMVA foresees three ways to execute an analysis: (i) as a macro using the CINT interpreter, (ii) as an executable, and (iii) as a python script via PyROOT. The first two options are run in the following.

Running the example analysis as a ROOT macro

Note that there is no significant loss in execution speed by running the macro compared to an executable, because all time-consuming operations in TMVA are done with compiled code accessed via the shared library.

• To quickly run the example macro for the training of a Fisher discriminant, type the following:

  cd macros
  root -l TMVAnalysis.C\(\"Fisher\"\)
  

  The macro should run through very quickly, and pop up a GUI with buttons for validation and evaluation plots. The buttons show the distributions of the input variables (also for preprocessed data, if requested), correlation profiles and summaries, the classifier response distributions (and probability and Rarity distributions, if requested), efficiencies of cuts on the classifier response values, the background rejection versus the signal efficiency, as well as classifier-specific validation plots. Try them all!
• We can now attempt to train more than one classifier, e.g., Fisher and the projective likelihood:

  root -l TMVAnalysis.C\(\"Fisher,Likelihood\"\) 

  Again the GUI will pop up, and it is interesting to check the background rejection versus the signal efficiency plot. Fisher performs better, which is due to the ignorance of correlations in the projective likelihood classifier. Plotting the likelihood response distributions indicates the problem: background slightly rises in the signal region, and vice versa.
• The likelihood performance on this toy data can be improved significantly by removing the correlations beforehand via linear transformation using a TMVA data preprocessing step:

  root -l TMVAnalysis.C\(\"Fisher,Likelihood,LikelihoodD\"\) 

  Plotting the background rejection versus the signal efficiency graph shows that the decorrelated likelihood classifier is now as performing as Fisher.
• Let's have a look at the latest log file.
• If you have time you could now run all preset classifiers, which may take O(10 mins):

  root -l TMVAnalysis.C

  After terminating, try how the the background rejection versus the signal efficiency plot looks like. Also, it is interesting to look into the response distributions of the classifiers.

Running the example analysis as an executable

• For the proof of principle, let's only run the training of a Fisher discriminant. For this we need to remove all other classifiers from the list:

  cd ../examples
  emacs -nw TMVAnalysis.cxx
  # now set all "Bool_t Use_<Classifier> = 0;" 
  # except for Fisher, and leave emacs (ctrl-x-c)
  

• Build the executable:

  make TMVAnalysis

• And execute it:

  ./TMVAnalysis

• Start the GUI and plot:

  root -l ../macros/TMVAGui.C

The training phase

Let's look at the analysis script:
emacs TMVAnalysis.C
Discussion of main analysis steps:

1. Instantiation of the Factory
2. Registration of the signal and background trees
3. Registration of the input variables
   ( note that the types 'F' or 'I' given indicate whether a variable takes floating point or integer values, respectively: 'F' stands for both float and double, 'I' stands for int, short, char, and unsigned integers)
4. Preparation of independent training and test trees
5. Booking of the classifiers (discussion of the classifier labels and the configuration options)
6. Train, test and evaluate all booked classifiers
7. Created outputs: weight files, standalone C++ classes, target file

Applying the trained classifiers to data

Once the training phase (including training, testing, validation and evaluation) has been finalised, selected classifiers can be used to classify data samples with unknown composition of signal and background events:

root -l TMVApplication.C\(\"Fisher,Likelihood,LikelihoodD\"\) 

The macro runs over signal events only. Plot the produced histograms:

.ls
MVA_Fisher.Draw()
MVA_Likelihood.Draw()
MVA_LikelihoodD.Draw()

The application phase using the *Reader*

Let's look at the application script:
emacs TMVApplication.C
Discussion of main application steps:

1. Instantiation of the Reader
2. Registration of the input variables
3. Booking of the classifiers via weight files
4. Run the event loop and request classifier response for each event (special case: the Rectangular Cut classifier)

The application phase using *standalone C++ classes*

These classes are automatically created during the classifier training, and are fully self-contained. The classes are ROOT-independent. Let's look at the application script:

emacs ClassApplication.C

Discussion of main application steps:

1. Create vector with input variable names
2. Load C++ class for classifier response
3. Instantiate classifier response class object
4. Run the event loop and request classifier response for each event

Note that C++ standalone classes are not available for the classifiers: Rectangular Cuts, PDERS, k-NN, TMlpANN and CFMlpANN

Discussion

A reminder of useful discussion topics

• Strengths and weaknesses of the classifiers
• How good are the default configuration settings of the classifiers ?
• How much training statistics is required ?
• How to avoid overtraining - is independent validation required ?
• Is data preprocessing, such as decorrelation always useful ?
• How to treat systematic uncertainties (in the training, in the application) ?
• Better use the Reader or the standalone C++ classes for the application ?

---

Links

• TMVA Home
• TMVA Download
• TMVA Users Guide
• TMVA Sourceforge Project Page

---

TMVA Web Utilities

•   - advanced search
• WebTopicList - all topics in alphabetical order
• WebChanges - recent topic changes in this web
• WebNotify - subscribe to an e-mail alert sent when topics change
• WebRss, WebAtom - RSS and ATOM news feeds of topic changes
• WebStatistics - listing popular topics and top contributors
• WebPreferences - preferences of this web

---

-- AndreasHoecker - 18 Jun 2007