---

CERNsight — Higgs Boson Event Classifier

CERNsight is a heterogeneous ensemble classifier that distinguishes Higgs boson signal events from background noise in particle collision data. It combines gradient-boosted trees (XGBoost) with a deep neural network, blending their outputs via soft voting (probability averaging) rather than a gated mixture-of-experts routing scheme.

The model is trained on a 500,000-row stratified sample of the HIGGS benchmark dataset (11M total collision events, 28 kinematic features per event), originally introduced by Baldi, Sadowski & Whiteson (2014) to test whether deep learning could automatically discover the high-level derived physics features that experts had previously hand-engineered.

Predictions are paired with SHAP-based explainability, so each classification can be traced back to the specific detector measurements that drove it — important for a domain where physicists need to audit, not just trust, a model's output.

This repository holds the trained baseline models from the two-week club submission track. A longer-term scaled version — stacking meta-learner, cross-validated training, MC Dropout uncertainty estimates, and deeper physics-alignment analysis via SHAP — is in progress, targeting a workshop/journal submission.

Model Description

• Architecture: XGBoost (gradient boosted trees) + Dense NN (512→256→128→1), combined via probability averaging ...