A Representation-Based Sum-Product Neural Network for Higgs Boson Event Classification
Year: 2026
Type: book-chapter
Abstract: The discrimination of Higgs boson signal events from Standard Model backgrounds is one of the key tasks for modern particle physics. In H→ττ channel, the signal refers to simulated events with Higgs boson mass of 125 GeV decaying into two tau leptons, whereas the background includes known processes such as Z→ττ, top quark pair production, and W boson, which mimic the same observable final state. Here, we present an interpretation of the performance of two neural network models based on physical insight in the context of the ATLAS Higgs Boson Machine Learning Challenge dataset. The direct sum-product model considers a combination of additive and multiplicative terms calculated directly from scaled event level variables. In contrast, the representation-based model learns dense latent representation first and uses additive and multiplicative decision making process within it afterward. The intuition is based on the fact that collider observables are heterogeneous physical quantities; hence, discriminative evidence can be often found in correlations between transverse momenta, angles, missing transverse energy, jets, and various derived mass combinations. Using stratified splits on a dataset of 818,238 simulated events, experiments have shown that representation-based model significantly outperforms the direct approach in terms of testing accuracy (0.7524 vs. 0.8426), ROC-AUC (0.8152 vs. 0.9113), signal class recall (0.53 vs. 0.73), and signal class F1-score (0.60 vs. 0.76). These results suggest that additive and multiplicative evidence becomes more informative when computed on top of the learned features as opposed to directly on scaled raw kinematics. The research should be treated as an architecture level comparison task, not the complete analysis of Higgs boson search since event level weights, Approximate Median Significance optimization, systematics, calibration, and inference procedures are out of the scope of this research.
Source: Özgür Yayınları eBooks
Authors: Şafak Yasun, Ferdi Güler
Institution: Istanbul Arel University
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7169649193