Multivariate classification for ggH → ZZ* → 4ℓ using Boosted Decision Trees

Abstract

This analysis is performed using the publicly available ATLAS Open Data 2015–2016 (Run 2), corresponding to proton–proton collision data at a centre-of-mass energy of \(\sqrt{s}=13\,\text{TeV}\). We study the Higgs boson decay channel \(gg\mathrm{H}\to H\to ZZ^{*}\to4\ell\) using open collision datasets together with Monte Carlo simulation samples. The four-lepton final state offers a clean signature for Higgs boson reconstruction, while the main background comes from irreducible non-resonant \(ZZ^*\) continuum production. We design a conventional cut-based event selection scheme and compare it systematically with a multivariate analysis built upon the Boosted Decision Tree (BDT) classifier. Six kinematic observables characterizing the four-lepton system and two reconstructed Z boson candidates are used as input features for BDT training.Signal and background samples show distinguishable kinematic distributions and feature correlation patterns, which deliver strong separation capability. We assess BDT performance via three metrics: signal efficiency, background rejection rate and signal purity. Statistical and systematic uncertainties are calculated following the official ATLAS collaboration methodology for the \(H\to ZZ^*\to4\ell\) channel, with a full systematic uncertainty budget tabulated. In the invariant-mass window \(80<m_{4\ell}<160\,\mathrm{GeV}\), the multivariate BDT approach lifts signal purity from \(25\%\) (cut-based baseline) to \(37\%\), which clearly proves the strengths of machine-learning methods for Higgs-physics analyses in the four-lepton final state.