Yannik Fausch

2024
Max Planck Institut for Physics Garching

Abstract: This thesis aims to compare TabNet, a Deep Learning (DL) algorithm for classi- fication of tabular data, to the established Boosted Decision Tree (BDT) used for event selection of the τ − → π−π−π+ντ decay. The goal is to see whether Tab- Net can outperform the established BDT. First the impact of hyperparameter values of TabNet were studied manually before conducting a high-dimensional automated hyperparameter optimization. The established BDT still yields a better performance over all purities than TabNet. Additionally, the feature im- portance of TabNet was studied and compared to the established BDT, to see the influence of features on to the prediction of these two models. For both models the most important category was the event-shape and the third most important were the vertex reconstruction features. The second and fourth fea- tures differ for TabNet and the established BDT. The second most important feature for TabNet were the γ and π0 rejection features and the forth most im- portant features were the kinematic features. For the established BDT it was the other way around.

Note: Presented on 29 10 2024
Note: BSc