Prepare

In order to run this tutorial a working installation of conda is needed. Once you have this, follow the instructions to setup the environm ent.

1. Clone this repository

$ git clone https://github.com/maxgalli/UsefulHEPScripts
$ cd UsefulHEPScripts/flashgg_investigate

2. Create a conda environment

$ conda env create -f environment.yml

Content

This folder includes useful studies performed to find an optimal structure for the bew flashgg framework.

XGBoost -> RDataFrame

Here we investigate the possibility of efficiently apply tags and systematics in an RDataFrame analysis flow. The application would be the new flashgg framework. The idea consists in training a model with XGBoost for some variables and then apply it, in the context of an RDataFrame flow, to a new dataset which has the same variables the model was trained with plus two "modified" versions (Up and Down, which represent the systematics).

Workflow

• prepare_data_with_sys.py: this program fetches two different remote datasets, SMHiggsToZZTo4L.root and ZZTo2e2mu.root, which will be respectively our signal and background. Data is processed to flatten the variable of interest (Muon_pt_1, Muon_pt_2, Electron_pt_1, Electron_pt_2) and Define is applied to produce fake branches of systematic variations (variables + _Up and variables + _Down), for a total amount of 12 branches in each dataset (signal and background). Then, both datasets are split into training and test, ending up with train_sys_signal.root, train_sys_background.root, test_sys_signal.root and test_sys_background.root.
• train_and_save_models.py: here we use XGBoost to train a classifiers that separates signal events from background events. The classifier is saved into two different formats: plain XGBoost in classifier.pkl and in TMVA digestible format as myBDT inside classifier.root.
• test_jitting.py: here we set-up a trick (necessary due to the fact that RDataFrame.Define accepts strings of C++ code as arguments) to conveniently make signal-bkg predictions for the events stored in test_sys_signal.root, ending up in the branches y, y_up and y_down. The bdt is set up only once for the three predictions. The predictions for the first 200 events are printed (run with python test_jitting.py > test_jitting.log to store it in a log file).

Double-check

Run python test_pure_xgb.py > test_pure_xgb.log to perform the test for signal dataset using only XGBoost from Python and compare the results.