ML-Based Fake Factor Evaluation

The framework supports the evaluation of fake factors based on machine learning models in the ONNX format instead of the classic approach of measuring them (see Fake Factor Calculation). ONNXRuntime is utilized to to evaluate the models within ROOT RDataFrames.

ML Training

Not include yet. Files/models have to be provided externally.

Configuration

To use ML-based fake factors, the configuration must specify the ONNX model details for each process. These details are defined in a YAML configuration file, which should be located in the workdir/TAG directory. The configuration includes:

• Model Path: The path to the ONNX model file.
• Model Inputs: A list of input variable names required by the model.
• Define Columns: Optional expressions to define missing columns dynamically.

Example YAML configuration:

target_processes:
  QCD:
    model_path: "models/qcd_fake_factors.onnx"
    model_input:
      - "event_parity"
      - "pt_1"
      - "njets"
      ...
    define_columns:
      event_parity: event % 2
      ...
  Wjets:
    model_path: "models/wjets_fake_factors.onnx"
    model_input:
      - "pt_2"
      - "nbtag"
      ...
  ...

Notes

• Ensure that the ONNX model is compatible with ONNXRuntime and that the input variable names match those in the configuration.