tZq_bdt_example.yaml

# Directory containing input files for the classifier.
# The input directory is expected to contain a ROOT file for each process named
# histofile_$PROCESS.root. This should contain a Ttree named
# Ttree_$PROCESS. Additional trees named with the schema
# Ttree_$PROCESS__$SYSTEMATIC__(plus/minus) should be included for each
# relevant shape nuisance. The Ttrees should contain a branch for each
# observable and a branch containing event weights.
input_dir: /scratch/data/TopPhysics/mvaDirs/inputs/2016/all/mz50mw50/

# Random number generation seed
seed: 52

# ROOT selection string specifying the cuts that should be made before
# classifier training takes place.
selection: >-
    zMass > 71.2 && zMass < 111.2 &&
    wPairMass > 60.4 && wPairMass < 100.4 &&
    chi2 < 40 &&
    Channel == 1

# Name of the channel, only used to name output and does not apply any selection
channel: ee

# List of channels which should be considered signal
signals:
    - tZq

# List of channels which should be considered background
backgrounds:
    - DYToLL_M10To50_aMCatNLO
    - DYToLL_M50_aMCatNLO
    - FakeEG
    - FakeMu
    - TbartChan
    - TbartW
    - THQ
    - TsChan
    - TT
    - TtChan
    - ttH
    - TTW
    - TtW
    - TTZ
    - TWZ
    - Wjets
    - WW
    - WWW
    - WWZ
    - WZ
    - WZZ
    - ZZ
    - ZZZ

# Name of process containing collision data
# The location of the true data needs to be known when combine/THETA output is
# generated
data_process: DataEG

# Directories plots, root files, and trained classifiers should be output into
plot_dir: plots/
root_dir: root/
mva_dir: mva/

# Fraction of data to be reserved in test sample
test_fraction: 0.25

# If true, the weights of the signal channels are linearly scaled so that the
# overall normalisation for both the signal and background channels is the same
equalise_signal: true

# How negative event weights should be treated
#   passthrough: negative weights are unaltered
#   abs: the absolute value of all negative weights is taken
#   reweight: The absolute value of all negative weights is taken. The original
#             normalisation for each process is then restored by linearly
#             scaling the resulting weights down. This will fail if any
#             processes have an overall negative weight.
#   zero: negative weights are set to 0
negative_weight_treatment: passthrough

# Classifier selection
#   bdt_grad: Gradient Boosted Decision Tree (scikit-learn)
#   bdt_xgb: Gradient Boosted Decision Tree (XGBoost)
#   bdt_lgbm: Gradient Boosted Decision Tree (LightGBM)
#   random_forest: Random Forest
#   mlp: Multi-Layer Perceptron (Keras)
#   load: load classfier specfied by classifer_path option
classifier: bdt_grad

# BDT configuration. Passed to scikit-learn's GradientBoostingClassifier()
# See scikit-learn documentation for more information
bdt_grad:
    n_estimators: 100
    verbose: 1
    min_samples_split: 0.1
    subsample: 0.75
    learning_rate: 0.02
    max_depth: 5

# Options governing the root file output
root_out:
    # Whether output should be in the format for combine (true) or THETA (false)
    combine: true

    # What form the (pseudo)-data in the files should take
    # empty: Empty histograms
    # poisson: Sum the Monte Carlo histograms, and perform a Poisson jump on
    #          each bin
    # real: Use the real data
    data: empty

    # The strategy used to bin the MVA response in the resulting root files
    #   equal: specified number of equal-width bins in the (0, 1) range
    #          (default).
    #   quantile: specified number of equally-populated bins, achieved by
    #             placing bin edges at quantiles. Bin population does not take
    #             event weight into account.
    #   recursive_median: response is recursively bisected at the median
    strategy: equal

    # Set the number of bins for the equal or quantile binning stategies
    bins: 20

    # The recursive binning strategies will stop splitting once these limits
    # are reached
    min_signal_events: 0
    min_background_events: 1
    max_signal_error: 0.3
    max_background_error: 0.3

# Features to be included in the classifier training
features:
    - bTagDisc
    - fourthJetPt
    - jetMass
    - jjdelR
    - leadJetEta