Higgs Boson Classification Project

Machine learning-based search for Higgs boson events using binary and multiclass classification to optimise binned likelihood fits.

Project Structure

higgs-boson-classification/
├── README.md                               # Project documentation
├── requirements.txt                        # Python dependencies
├── MResFindingTheHiggsAssignment.ipynb     # Original assignment specification
├── src/
│   ├── data/
│   │   ├── loader.py                       # Data loading utilities
│   │   ├── preprocessor.py                 # Data preprocessing and scaling
│   │   └── binning.py                      # ML-based histogram creation
│   └── models/
│       ├── xgboost_classifier.py           # XGBoost classifier (auto-detects binary/multiclass)
│       ├── random_forest.py                # Random Forest classifier
│       └── neural_network.py               # PyTorch neural network
├── notebooks/
│   ├── analysis.ipynb                      # Written summary of complete analysis
│   ├── exploratory_data_analysis.ipynb     # EDA of Higgs dataset
│   ├── binary_classification.ipynb         # Binary classification experiments
│   └── multiclass_classification.ipynb     # Multiclass classification experiments
└── configs/
    ├── xgboost_config.yaml                 # XGBoost hyperparameters
    ├── nn_config.yaml                      # Neural network architecture
    └── rf_config.yaml                      # Random Forest parameters

Installation

cd higgs-boson-classification
pip install -r requirements.txt

Acknowledgements

AI assistance used for:

• Formatting tables and data for visual clarity
• Markdown cells and headers for clearer separation of sections
• Setting up workspace structure (configs, notebooks, src directories)
• Initial README documentation
• requirements.txt generation

Quick Start

Option 1: Read Project Summary

Open notebooks/analysis.ipynb for a written summary of the complete project:

• Overview of exploratory data analysis, binary and multiclass classification approaches
• Explanation of methods: data cleaning, model training, probability binning, likelihood fitting
• Summary of results and performance across all notebooks
• No code execution required - just a narrative walkthrough of the analysis

Option 2: Detailed Exploration

Explore individual notebooks:

• exploratory_data_analysis.ipynb - Understand the dataset
• binary_classification.ipynb - Signal vs background approach
• multiclass_classification.ipynb - Full multiclass analysis with optimisation

Assignment Goal

Optimise a Higgs boson search by creating ML-based bins that provide the most precise binned likelihood fit for signal rate measurements.

Target Precision (CMS Published):

• μ (combined): 6%
• μ_ggH: 9%
• μ_VBF: 18%

Approach

Binary Classification

• Separate signal (ggH + VBF) from background (Z)
• Use classifier scores to create 1D histogram bins
• Run likelihood fits for μ_ggH and μ_VBF

Multiclass Classification (Primary Approach)

• Train separate XGBoost models per channel (et, mt, tt)
• Classify events as Z (background), ggH (signal), or VBF (signal)
• Create 2D histograms using multiclass probabilities:
  • x-axis: P(ggH) / [P(ggH) + P(VBF)] — separates production modes
  • y-axis: P(ggH) + P(VBF) — separates signal from background
• Optimised binning: 4×15 grid (60 bins total)
• Combine histograms across channels for final likelihood fit
• Provides superior separation for measuring independent signal strengths

Models

• XGBoost: Gradient boosted trees (primary classifier)
• Random Forest: Ensemble of decision trees
• Neural Network: PyTorch-based MLP

Data

Dataset from CMS H→ττ analysis:

• Channels: et (electron-tau), mt (muon-tau), tt (tau-tau)
• Processes: Z (background), ggH (signal), VBF (signal)
• Features: 27 kinematic variables (pt, eta, mass, etc.)

Configuration

Hyperparameters can be adjusted in YAML files in configs/:

• xgboost_config.yaml - Learning rate, max depth, n_estimators, etc.
• nn_config.yaml - Hidden layers, activation functions, optimizer
• rf_config.yaml - Number of trees, max features, max depth

Results

Precision achieved on signal strength measurements (lower is better):

Method	μ (%)	μ_ggH (%)	μ_VBF (%)
CMS Published	6.0	9.0	18.0
Multiclass XGBoost	5.78	8.47	17.22

The multiclass approach with per-channel XGBoost models meets all CMS precision benchmarks.

Key Features

• Per-channel training: Separate models preserve channel-specific kinematics
• Auto-detection: XGBoost classifier automatically detects binary vs multiclass mode
• Missing value handling: Converts -9999 indicators to median-filled values
• Histogram scaling: Applies physics cross-section scaling (Z×8.4, ggH×0.034, VBF×0.011)
• Binning optimisation: Configurable grid search for optimal precision
• Likelihood fitting: Poisson NLL fits using iminuit

References

• CMS Collaboration H→ττ analysis
• ML Assessment 3 dataset: https://github.com/gputtley/ML-Assessment-3

License

Academic project for ML course assignment.