Feasibility and Reliability Studies of Graph Neural Networks for Multivariate tt+X Event Classification at the CMS Experiment at CERN

Graph neural networks (GNNs) are a novel machine learning technique dedicated to processing graph data. In fact, a lot of high energy physics data can be more naturally represented by graphs rather than, for instance, by vectors, which deep neural networks (DNNs) use. This applies to \(\text{t}\bar{\text{t}}+\text{b}\bar{\text{b}}\), \(\text{t}\bar{\text{t}}\text{H}(\text{b}\bar{\text{b}})\) and \(\text{t}\bar{\text{t}}\text{Z}(\text{b}\bar{\text{b}})\) processes as well, which are irreducible backgrounds to each other and need to be separated with high efficiency due to the importance of each process in itself. Therefore, GNNs are presumably a promising approach for this classification task. To validate this assumption, the general feasibility of GNNs for multivariate \(\text{t}\bar{\text{t}}+\text{X}\) event classification as well as their reliability are examined. The studies are concluded with comparisons between GNNs and DNNs under fair conditions.