Neural Network-Based Classification of Higgs Boson Signal and Background Events in the H→τ τ Channel Using ATLAS Open Data

Abstract The simulated ATLAS H→ττ H → ττ open dataset, released through the CERN Open Data Portal, provides a useful benchmark for signal–background classification. However, model performance on labelled data can depend strongly on preprocessing, event-weight treatment, classification thresholds, and uncertainty estimation. This study presents a reproducible, leakage-controlled educational framework rather than a new classification algorithm. The framework combines gradient-verified NumPy implementations of logistic regression and a two-hidden-layer neural network with missingness-aware preprocessing fitted exclusively on training data. A frozen evaluation protocol is adopted in which models are selected using validation weighted ROC-AUC, the selected-event fraction is fixed at 13.0% for all models, and the internal holdout set is evaluated only after all modelling decisions are finalized. Raw-subset approximate median significance (AMS), with breg=10 b reg​=10, is used as the primary yield metric, while two alternative ×5 extrapolation conventions are reported separately as supplementary evaluations. Using 250,000 labelled simulated events, the five-seed neural-network ensemble achieves a holdout weighted ROC-AUC of 0.9335 and a raw AMS of 1.56, with a 95% event-bootstrap confidence interval of [1.45, 1.67] based on 2,000 resamples. Individual ensemble members achieve an average AUC of 0.9325±0.00040.9325±0.0004 and raw AMS of 1.51±0.021.51±0.02 across seeds. The ensemble significantly outperforms the tuned logistic-regression baseline, with a paired holdout AMS difference of 0.85 and an interval of [0.75, 0.96]. However, its performance advantage over tuned histogram gradient boosting, with an AMS difference of 0.040 and an interval of [−0.032, 0.124], is not distinguishable from evaluation-sample uncertainty; therefore, no statistically supported improvement over this model is claimed. Ablation studies are conducted using a single seed and are interpreted as directional evidence only. Since the dataset consists entirely of simulated events, detector-related systematic uncertainties are not modelled. The framework emphasizes reproducibility, controlled evaluation, and transparent reporting of classification performance in high-energy physics.