Tau The Difference

Separate Higgs to Tau Tau decays from Genuine and Fake Backgrounds.

Channels currently supported:

• Fully hadronic ($\tau_h\tau_h$)

Models currently supported:

• XGBClassifier (optimal so far)
• Simple DNN

Production

The production of training tuples starts with parquet files that are NanoAOD skims processed by the HiggsDNA framework for H $\to\tau\tau$ analyses.

The code necessary to producte samples for training (and validating/evaluating) models is in the Production directory.

Configuration

In the config subdirectory, there are yaml files detailing the datasets to be used for each era, as well as relevant run parameters.

Currently available:

• Run3_2022
• Run3_2022EE
• Run3_2023
• Run3_2023BPix

The central production is controlled by config.yaml. Here the input/output paths, eras to be used, and selection to apply is specified.

Execution

Production of the ShuffleMerge files used for training can be done in one command:

./scripts/run_prod.sh

This will execute python/PreProcess.py and python/ShuffleMerge.py.

The PreProcess.py script applies selections to the input datasets, adds class and era labels, and any relevant physics weihts (lumi*xs, filter efficiencies, LHE etc...)

The ShuffleMerge.py mixed all samples, and creates a class_weight where all three categories (genuine tau (0), higgs signal (1) and background (2)) are balanced. This ensures that the classes are considered equally during training.

Output

Running these steps will produce several files:

• ShuffleMerge_ALL.parquet which contains all events (useful for analysis histograms etc)
• ShuffleMerge_EVENmodel_Train and ShuffleMerge_EVENmodel_Val to train and validate models that will be applied to EVEN events (these files contain ODD events only)
• ShuffleMerge_ODDmodel_Train and ShuffleMerge_ODDmodel_Val to train and validate models that will be applied to ODD events (these files contain EVEN events only)

These datasets can be found in the output directory specified in teh central config.yaml file.

Analysis

Plotting and analytics for dataframes.

Training

Model training and validation.

Evaluation

Inference (applying training), plotting tools (confusion matrices, score distributions, ROC)