Transforming tabular data into graphs for GNN-driven classification of BSM and standard model events in high-energy collisions

This study investigates the application of Graph Neural Networks (GNNs) in classifying a Beyond Standard Model (BSM) signal and Standard Model (SM) backgrounds in high-energy particle physics. Motivated by the challenges of identifying rare BSM events amidst substantial SM backgrounds, we propose three distinct methods to transform particle data into graph representations. These methods differ in graph topology and feature encoding, enabling a comprehensive analysis of GNN performance. Dataset utilized in this work comprises simulated collision events featuring gluino pair production as the BSM signal and various SM processes as the background. Three GNN architectures were developed, with graph sizes ranging from 9 to 36 nodes, and evaluated against each other. Performance metrics such as accuracy, precision, recall, and area under the ROC curve were used to compare performance of the models. The findings demonstrate that GNNs are particularly effective at exploiting the intrinsic relationship among the particles produced in proton-proton collisions, with the third method achieving the highest accuracy of 92.35% and AUC of 0.9668. These relationships are captured through the graph structure, where particles are represented as nodes and their features are connected to reflect patterns or correlations within an event. By modeling these connections, GNNs effectively distinguish signal events from background processes, highlighting their potential to enhance sensitivity in BSM searches and providing a framework for future applications in particle physics.