Deep Learning + CWoLa for VBF vs. GGF Classification

This project applies deep learning to distinguish between VBF and GGF Higgs production modes using the Classification Without Labels (CWoLa) framework. The approach is inspired by the paper Classification without labels: Learning from mixed samples in high energy physics, which introduces CWoLa as a viable strategy for learning directly from mixed real data samples.

---

Environment Setup

1. Download miniconda through:

   # Assuming Linux system
   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
   
   # Install thorugh sh. Type 'yes' when asking automatically initialization.
   sh Miniconda3-latest-Linux-x86_64.sh

2. Create a virtual environment:

   # Initialization
   conda env create -f environment.yml
   
   # Update when the environment.yml changed
   conda env update -f environment.yml

3. Directly activate in Jupyter, or activate/exit with:

   conda activate cwola
   conda deactivate

4. (Optional) Create a .env such that VSCode can fetch the packages.

   PYTHONPATH=~/miniconda3/envs/cwola/lib/python3.12/site-packages
   

Models

Convolutional Neural Networks (CNN)

• CNN_Baseline: A reference implementation based on VBF vs. GGF Higgs with Full-Event Deep Learning: Towards a Decay-Agnostic Tagger. This model utilizes event-level CNNs to perform classification.

• CNN_Light: A simplified and more lightweight version of CNN_Baseline, designed to reduce model complexity and training time while retaining performance.

Particle Transformers (ParT)

• ParT_Baseline: A transformer-based architecture based on Particle Transformer for Jet Tagging. This model captures particle-level features using attention mechanisms tailored for jet tagging tasks.

• ParT_Light: A family of lighter variants derived from ParT_Baseline, offering faster training and inference with reduced computational cost.

Usage

• The training and inference scripts is given in scripts/.