Classification and Regression Studies for Flavour Tagging

The identification of jets containing b-hadrons, commonly referred to as b-tagging, plays a crucial role in high-energy physics experiments for both precise measurements of the Standard Model and for exploring scenarios of new physics. b-tagging algorithms are based on a variety of discriminant information related to the long b-hadron lifetime and are constructed by exploiting boosted-decision tree or deep neural-network techniques. he first part of this project aims at comparing the performance of different high-level tagging algorithms. The discrimination power of the b-tagging algorithms, however, is strongly dependent on the jet transverse momentum. The jet transverse momentum is being exploited by the b-tagging algorithms as a proxy for the b-hadron pT , the latter being a quantity not reconstructed by the software that inspects the events and builds the physics objects originating from the collisions. Nevertheless, the b-hadron pT could also be inferred from the outputs of the low level taggers employing a regression model. The second part of the project consists in building such models. The study is carried out in a Python environment, decision trees and neural network are implemented using Scikit-learn [8], XGBoost [4], LightGBM [6] and PyTorch [7].