% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@PHDTHESIS{Quadfasel:626057,
author = {Quadfasel, Tobias},
othercontributors = {Kasieczka, Gregor and Schleper, Peter},
title = {{A} {M}odel-{A}gnostic {S}earch for {B}eyond the {S}tandard
{M}odel {P}hysics in the {D}ijet {T}opology using {M}achine
{L}earning},
school = {University of Hamburg},
type = {Dissertation},
reportid = {PUBDB-2025-01279},
pages = {229},
year = {2024},
note = {Dissertation, University of Hamburg, 2024},
abstract = {Despite significant efforts to search for new beyond the
Standard Model phenomena at the LHC physics program and
beyond, no evidence has been found so far. Model-agnostic
searches complement current search efforts, allowing for the
detection of potential new physics anomalies without
targeting a particular signal model.This thesis discusses
the development and application of novel data-driven methods
for model-agnostic anomaly detection. In particular, a new
method based on neural density estimation and weak
classification – named Cathode – is developed, achieving
state-of-the-art performance on a commonly used benchmark
data set. The first application of Cathode and other Machine
Learning-based methods on proton-proton collision data taken
at the CMS experiment from 2016 to 2018 at a centre-of-mass
energy of √𝑠 = 13 TeV is discussed. Specifically,
hadronic resonances in the two-jet final state are targeted.
A generic search for new physics in the invariant mass
spectrum of the two jets revealed no significant excess. In
terms of cross-section limits, Cathode achieves the most
sensitive results for several of the tested mass hypotheses
of an 𝑋 → 𝑌 𝑌 ′ → 4𝑞 signal model, with an
optimal improvement over an inclusive search of factor
1.9.Several improvements to the original Cathode algorithm
are proposed. Latent Cathode allows for signal extraction in
the case of strong correlations between input features and
mass by conducting the weak classification in latent space.
Using Gradient Boosting classifiers instead of Deep
Learning-based methods in weak classification is shown to be
considerably more robust against uninformative features,
allowing for model-agnostic searches without prior feature
selection.},
pnm = {PHGS, VH-GS-500 - PIER Helmholtz Graduate School
$(2015_IFV-VH-GS-500)$},
pid = {$G:(DE-HGF)2015_IFV-VH-GS-500$},
typ = {PUB:(DE-HGF)11},
urn = {urn:nbn:de:gbv:18-ediss-118211},
url = {https://bib-pubdb1.desy.de/record/626057},
}