What question did this study set out to answer?

This research aims to provide an interpretable and reproducible method for classifying Higgs boson events from background noise.

July 10, 2026Open Access

Interpretable Machine Learning for Higgs Boson Event Classication A Reproducible Gradient-Boosting Baseline with SHAP Analysis

Key Points

This research aims to provide an interpretable and reproducible method for classifying Higgs boson events from background noise.
Developed a reproducible pipeline for event classification on a large ATLAS-format dataset with 2.5 x 10^5 labeled events.
Applied preprocessing techniques including sentinel-value imputation, standardization, and leakage-checked stratified splitting.
Trained both Random Forest and gradient-boosted decision-tree models, with performance evaluated using AUC and F1 scores.
Random Forest model achieved AUC = 0.8912 and F1 = 0.7486, while LightGBM model had AUC = 0.8895 and F1 = 0.7430.
LightGBM model interpretation highlighted the importance of DER_mass_MMC as the primary feature.
No improvement was found with the untuned LightGBM model compared to the baseline Random Forest.

Abstract

Separating Higgs-boson signal events from large Standard-Model backgrounds is a standardsupervised-learning task on high-level collider observables, and the classiers used for it aremost useful when their decisions can be inspected and reproduced rather than merely scored. We present a compact, fully reproducible pipeline for Higgs event classication on the ATLAS-format dataset popularised by the Higgs Machine Learning Challenge (HiggsML; 2. 5 × 105labelled training events and 5. 5×105 unlabelled test events). The pipeline applies a transparentpreprocessing protocol (sentinel-value imputation by training-set medians, standardisation, andleakage-checked stratied splitting), augments twelve high-level observables with six kinemat-ically motivated composite features, and trains two tree-ensemble learners: a Random Forest (RF) baseline and a gradient-boosted decision-tree model (LightGBM). On a held-out validationfold of 50, 000 events, the RF baseline attains AUC = 0. 8912 and F1 = 0. 7486, while the Light-GBM model attains AUC = 0. 8895 and F1 = 0. 7430; both operate in the 0. 880. 92 AUC rangereported for high-level tabular solutions to this dataset. We interpret the LightGBM model withgain-based feature importance and Shapley Additive Explanations (SHAP). Both attributionsplace the approximate di-tau mass DERₘassMMC as the dominant discriminant, followed bythe transverse-mass surrogate DERₘassₜransverseₘetₗep and missing-transverse-energyrelated quantities, and the SHAP dependence of DERₘassMMC is smooth and single-peaked, consistent with resonance-like behaviour. The contribution of this work is not a new state of theart but a readable, auditable, and reproducible baseline: we release the preprocessing, feature-construction, training, and evaluation scripts together with all predictions, metrics, and guresso that every reported number can be regenerated. We also report, without embellishment, thatthe untuned LightGBM model does not improve on the RF baseline here, and we discuss thecalibration, robustness, and systematic-uncertainty studies that a downstream physics analysiswould additionally require.

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper