Speaker
Description
Machine-learning benchmarks for signal--background separation in high-energy physics are often confounded by inconsistent feature definitions, training procedures, and evaluation metrics. We present a controlled comparison of six architectures for ttH multilepton classification, from XGBoost to Lorentz-equivariant graph networks, under standardised hyperparameter optimisation across a controlled feature hierarchy. We report both ROC AUC and profile likelihood signal-strength uncertainty, investigate model performance as a function of feature set and training-set size, and compare single-channel against unified multi-channel training, demonstrating that the choice of architecture and feature representation has a meaningful impact on analysis sensitivity.