Generative Models for High Energy Physics Measurements

In this thesis, we augmented Monte-Carlo simulated data for charged Higgs boson searches. Events are selected if they have two light leptons (electron or muon) of the same sign and exactly one hadronically decaying tau-lepton. For the data generation, a variational autoencoder model was used with evidence lower bound and symmetric equilibrium learning. Both mentioned learning approaches were also tested with hierarchical (Ladder) archi- tecture. For the data quality assessment, both qualitative and quantitative metrics were taken into account. The standard evidence lower bound (ELBO) learning model was selected as the best-performing option. The model was then used to generate data for the signal and background separation analysis experiments. The dependence of classifier performance on the training dataset size was demonstrated using two widely used machine learning paradigms for tabular data classification: gradient-boosted decision trees and deep neural networks.