Machine learning for ttH mechanism Higgs boson detection from CERN ATLAS data

One aspect of studying subatomic particles by observing proton-proton collision is being able to identify those collisions where the particles of interest occur, since thousands of collisions are happening in an accelerator such as the Large Hadron Collider (LHC) at any given time. Machine learning methods have shown the potential to improve the performance of the detection while using either hand-engineered features or low-level measurements from the detector as input features. One such particle, which has been studied by multiple research groups, is the Higgs boson. The aim of this thesis is to test and compare several machine learning algorithms and compare the usage of hand-engineered features with the usage of direct measurements of the detector on the task of detecting Higgs boson events, namely the $t\bar{t}H$ process. Gradient boosting, multi-layered perceptron (MLP) and TabNet algorithms were tested and the results show superior performance of gradient boosting algorithms. Hand-engineered features show superior performance as opposed to direct measurements from the detector. Combination of all types of features show the best performance. We also show that classifiers training with only the most important features can achieve results with only a small performance decrease, while on the other hand providing benefits in terms of training time and model simplicity. In addition, it is shown that for an increased amount of training data, the performance of the classifiers is expected to improve.