Boosted Decision Tree algo - notes

Updated on Thu, 2012-06-14 09:50. Originally created by balewski on 2012-06-14 09:20.

My Notes on Boosted Decision Tree

Boosting a simple learner results

"supervised learnning" algorithms

Alternating_decision_tree - are simpler to interpret than boosted decision tree

See PDF below

Adaptiv Boost (AdaBoost): weights of previously misclassied events are multiplied by a common boost weight 'alpha'.

MiniBon article, 4 pages

http://www.physics.ox.ac.uk/phystat05/proceedings/files/phystat05-proc.pdf

Nice long presentation about both: Boosted and Alternating

http://jboost.sourceforge.net/presentations/BoostingLightIntro.pdf

“Margin” is a measure of how correct an example is

ROOT: TMVA package contains some BDT code

http://root.cern.ch/drupal/content/tmva

Package description is at:

http://tmva.sourceforge.net/#exec_summary

User guide:

http://tmva.sourceforge.net/docu/TMVAUsersGuide.pdf

CERNT tutorial:

https://twiki.cern.ch/twiki/bin/view/TMVA

RCF code contains not-functioning TMVA but it was trivial to install the TMVA root package on a SL5.6 VM (I used DarkLight VM)

My test example consist of

a) tree generator wrJanTree.C produced Signal & background files: htreeB.root htreeS.root

b) adopted traing macro TMVAClassificationJ.C

Groups:

The STAR experiment