Machine learning model performances for the Z boson mass

Abstract

Z bosons, one of the electroweak gauge bosons, are produced in proton-proton (pp) collisions mainly by Drell-Yan (DY) process. Having short lifetime results in study of the vector boson production via reconstruction of its decay products. Because of its remarkable success in various disciplines, in this study machine learning (ML) models were applied to identify the invariant mass spectrum of Z boson produced in pp collisions at root s = 7 TeV at the Large Hadron Collider (LHC) by reconstruction of two oppositely charged same-flavor leptons, muons or electrons, in its peak region (60-120 GeV/c(2)). To introduce an alternative method for classic vector boson analysis, ensemble models such as Random Forest (RF), Weighted Random Forest (WRF), Balanced Random Forest (BRF), a gradient boosting framework as Light Gradient Boosted Machine (LightGBM), and Deep Neural Networks (DNNs) were preferred for identification of Z boson spectrum from its dielectron and dimuon decay channels, separately. Each model's performances were assessed by ML metrics such as area under receiver operating characteristic curve (AUC ROC), sensitivity, precision, and F-1 score. It is revealed that LightGBM algorithm has 99.311% success for predicting the Z boson invariant mass spectrum from its dielectron decay channel with 96.326% sensitivity and 93.500% precision. In dimuon analysis, RF model demonstrated 99.980% success for predicting the spectrum of the vector boson with 99.064% sensitivity and 99.528% precision. The results confirm that the data structure effects the model performances. Predicted and correctly predicted Z boson invariant mass spectrum by outperformed ML models in each analysis was fitted with Breit-Wigner (BW) convoluted Crystal Ball (CB) function to investigate the misclassification effect of the techniques on the mass peak position by comparison of the mean values of the CB. It is represented that due to more than 99% success of the algorithms, the mean values are consistent within errors for the relevant analysis.