Machine Learning-Based Higgs Boson Event Classification Using CERN ATLAS Data

Abstract The decay of the Higgs boson into a tau-lepton pair (H → ττ) is challenging to isolate because each tau decay produces at least one neutrino, which prevents complete event reconstruction, while several background processes yield similar visible final states. This study evaluates machine-learning classifiers for simulated H → ττ events using the ATLAS Higgs Machine Learning Challenge dataset released through CERN Open Data. The analysis uses 818,238 events, 30 original input features, and the official training, validation, and test partitions. Performance is assessed using weighted ROC-AUC and the Approximate Median Significance (AMS), with the operating threshold selected exclusively on the validation set. Nonlinear classifiers outperform logistic regression: random forests, multilayer perceptrons, and gradient-boosted trees achieve test AUC values close to 0.93, compared with 0.8997 for logistic regression. The engineered-feature gradient-boosting model reaches a test AMS of 3.579, with a 95% bootstrap interval of approximately 3.52–3.66. Paired bootstrap comparisons indicate that the observed gains from hyperparameter tuning and feature engineering are not statistically distinguishable from zero, whereas per-jet specialised models perform significantly worse. A four-model ensemble obtains an AMS of 3.648 in a reduced-computation experiment; this result is promising but requires confirmation with full training configurations. SHAP analysis identifies the estimated parent mass as the dominant predictor, with its contribution largest in the vicinity of the Higgs-boson mass. The results support a reproducible and statistically cautious baseline rather than a claim of state-of-the-art performance.