tt̄H Events Classification with Graph Neural Networks in 2L(SS)+1τHad Channel

Particle physics experiments at the Large Hadron Collider (LHC) require robust methods to distinguish between signal and background events. This study focuses on the classification of tt̄H events in the 2L(SS)+1τHad channel using machine learning techniques. Specifically, we compared the performance of Boosted Decision Trees (BDTs), Fully Connected Artificial Neural Networks (ANNs), and Graph Neural Networks (GNNs) in identifying signal events against the background processes tt̄W and tt̄Z. The models were trained using a set of 159 variables, encompassing object-specific and relational attributes, derived from Monte Carlo simulations. The graph structure in GNNs leverages the natural relational data between particles in an event. Our results demonstrate comparable performance across all three algorithms, with average AUC scores of approximately 0.695 for GNNs, 0.694 for BDTs, and 0.689 for ANNs. The study suggests that while GNNs did not significantly outperform traditional methods in this specific task, there is potential for improvement, possibly by increasing the sample size.