Weakly Supervised Learning for Quark vs Gluon Jet Discrimination
Pakarinen, Kalle (2025)
2025
Teknis-luonnontieteellinen DI-ohjelma - Master's Programme in Science and Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-08-22
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202508228412
https://urn.fi/URN:NBN:fi:tuni-202508228412
Tiivistelmä
Jets are collimated sprays of particles resulting from the hadronization of high-energy quarks and gluons produced in proton–proton collisions. They are among the most frequently observed objects at the Large Hadron Collider, and distinguishing between quark- and gluon-initiated jets is essential for many physics analyses. Conventional machine learning approaches for jet classification rely on Monte Carlo (MC) simulations for training, which introduces modeling uncertainties due to potential mismatches with real data. Weakly supervised methods offer an alternative by enabling training directly on data without requiring per-jet truth labels, thereby reducing dependence on simulation.
This thesis investigates the use of weakly supervised machine learning techniques for quark versus gluon jet discrimination using data from the CMS experiment. Two complementary methods are studied: Classification Without Labels (CWOLA) and Jet Topics. Both allow model training without access to simulation-based truth labels, reducing reliance on MC simulation data.
The CWOLA method trains a classifier to distinguish between quark and gluon jets using mixed samples with differing class proportions, without any individual jet labels. It was implemented using a graph neural network and evaluated on both simulated and detector-level data. The results show that CWOLA can approach the performance of fully supervised training, particularly when the dataset is large and the quark–gluon fractions in the training samples are sufficiently distinct. This demonstrates its viability for data-driven jet tagging in scenarios where reliable truth information is unavailable.
Jet Topics, based on topic modeling, is applied to extract the underlying quark and gluon feature distributions and to estimate the corresponding jet fractions directly from data. Applied to jet multiplicity, the method successfully identified interpretable quark-like and gluon-like topics. Comparisons with distributions constructed using jet flavour information from MC simulated data revealed significant mismatches, highlighting limitations of flavour-based definitions in simulation.
This thesis investigates the use of weakly supervised machine learning techniques for quark versus gluon jet discrimination using data from the CMS experiment. Two complementary methods are studied: Classification Without Labels (CWOLA) and Jet Topics. Both allow model training without access to simulation-based truth labels, reducing reliance on MC simulation data.
The CWOLA method trains a classifier to distinguish between quark and gluon jets using mixed samples with differing class proportions, without any individual jet labels. It was implemented using a graph neural network and evaluated on both simulated and detector-level data. The results show that CWOLA can approach the performance of fully supervised training, particularly when the dataset is large and the quark–gluon fractions in the training samples are sufficiently distinct. This demonstrates its viability for data-driven jet tagging in scenarios where reliable truth information is unavailable.
Jet Topics, based on topic modeling, is applied to extract the underlying quark and gluon feature distributions and to estimate the corresponding jet fractions directly from data. Applied to jet multiplicity, the method successfully identified interpretable quark-like and gluon-like topics. Comparisons with distributions constructed using jet flavour information from MC simulated data revealed significant mismatches, highlighting limitations of flavour-based definitions in simulation.