Generating Explanations for Graph Neural Networks for tt+X events at the CMS Experiment

In the last years, efforts have been made to use various machine learning models in the multiclass classification of the tt+X events. Especially Graph Neural Networks (GNN) architectures have received attention, due to the data structure lending itself to the representation as a multi-relational graph. Naturally, when using a GNN to classify an event, it is of interest why it makes a particular prediction. GNNs however, due to the complex structure of graph data and black-box nature, are hard to interpret.

This thesis focuses on examining a GNN trained for this classification task using the GNNExplainer method to better understand the underlying decisions taken by the GNN when using different tt+X events as input.