• a new Boosting Decision Tree method with Systematical Uncertainties into training for High Energy Physics
• reference: https://arxiv.org/abs/1810.08387 (accepted for publication in NIM A)
• An example in High Energy Physics, search for Higgs -> tau tau gamma, under the directory tautaugamma
• To use it, just git clone. No complie. Only python and ROOT are required.

• Ligang Xia, [email protected], [email protected]

• run trainin: python runbdt.py trees0 0 0 10
• run testing: python testbdt.py trees0 (after training is done)
• I put the training and testing results in trees0/example/. You can have a comparison.

• run training: python runbdt.py trees1 1 1 10
• run testing: python testbdt.py trees1 1 (after training is done)
• I put the training and testing results in trees1/example/. You can have a comparison.

• command format: python runbdt.py dir Nsysts Switch Ntrees # see the explanation below
• dir: directory for storing training results
• Nsysts: number of systematics
• Switch: a boolean flag to switch on systematics or not in training. If Nsysts==0, Switch will be always 0.
• Ntrees: number of trees used for training, 100 by default if not specified.

• command format: python testbdt.py dir Nsysts Ntrees # see the explanation below
• dir: directory for storing training results
• Nsysts: number of systematics, 0 by default if not specified
• Ntrees: number of trees used for testing, 100 by default if not specified.

• qbdtmodule.py : define QBDT class (you do not need to touch it)
• runbdt.py : perform training
• testbdt.py : test and show performance
• root_dir : a directory storing root including nominal and systematic ntuples
• share : a directory storing other scripts, maybe useful, but you do not need to touch it
• AtlasStyle : a config script for plotting, borrowed from ATLAS

• We have to add a branch in the root file to tell the algorithm which events are used for training or testing. In the current example, this branch is "trainflag". It is generated randomly and uniformly from 0 to 1. Events with "trainflag<0.5" are used for training while the other events used for testing. I will try to split the events automatically in the future.

• Add a function to split events for training and testing automatically.
• Try to improve the training speed. I find python is slow. Maybe I should consider rewritting using C++.

I would like to thank my wife, who is always pushing me to publish PRL/Science/Nature papers and I always make her disappointed ...