-
Challenging Gradient Boosted Decision Trees with Tabular Transformers for Fraud Detection at Booking.com
Authors:
Sergei Krutikov,
Bulat Khaertdinov,
Rodion Kiriukhin,
Shubham Agrawal,
Kees Jan De Vries
Abstract:
Transformer-based neural networks, empowered by Self-Supervised Learning (SSL), have demonstrated unprecedented performance across various domains. However, related literature suggests that tabular Transformers may struggle to outperform classical Machine Learning algorithms, such as Gradient Boosted Decision Trees (GBDT). In this paper, we aim to challenge GBDTs with tabular Transformers on a typ…
▽ More
Transformer-based neural networks, empowered by Self-Supervised Learning (SSL), have demonstrated unprecedented performance across various domains. However, related literature suggests that tabular Transformers may struggle to outperform classical Machine Learning algorithms, such as Gradient Boosted Decision Trees (GBDT). In this paper, we aim to challenge GBDTs with tabular Transformers on a typical task faced in e-commerce, namely fraud detection. Our study is additionally motivated by the problem of selection bias, often occurring in real-life fraud detection systems. It is caused by the production system affecting which subset of traffic becomes labeled. This issue is typically addressed by sampling randomly a small part of the whole production data, referred to as a Control Group. This subset follows a target distribution of production data and therefore is usually preferred for training classification models with standard ML algorithms. Our methodology leverages the capabilities of Transformers to learn transferable representations using all available data by means of SSL, giving it an advantage over classical methods. Furthermore, we conduct large-scale experiments, pre-training tabular Transformers on vast amounts of data instances and fine-tuning them on smaller target datasets. The proposed approach outperforms heavily tuned GBDTs by a considerable margin of the Average Precision (AP) score. Pre-trained models show more consistent performance than the ones trained from scratch when fine-tuning data is limited. Moreover, they require noticeably less labeled data for reaching performance comparable to their GBDT competitor that utilizes the whole dataset.
△ Less
Submitted 22 May, 2024;
originally announced May 2024.
-
Machine Learning for Fraud Detection in E-Commerce: A Research Agenda
Authors:
Niek Tax,
Kees Jan de Vries,
Mathijs de Jong,
Nikoleta Dosoula,
Bram van den Akker,
Jon Smith,
Olivier Thuong,
Lucas Bernardi
Abstract:
Fraud detection and prevention play an important part in ensuring the sustained operation of any e-commerce business. Machine learning (ML) often plays an important role in these anti-fraud operations, but the organizational context in which these ML models operate cannot be ignored. In this paper, we take an organization-centric view on the topic of fraud detection by formulating an operational m…
▽ More
Fraud detection and prevention play an important part in ensuring the sustained operation of any e-commerce business. Machine learning (ML) often plays an important role in these anti-fraud operations, but the organizational context in which these ML models operate cannot be ignored. In this paper, we take an organization-centric view on the topic of fraud detection by formulating an operational model of the anti-fraud departments in e-commerce organizations. We derive 6 research topics and 12 practical challenges for fraud detection from this operational model. We summarize the state of the literature for each research topic, discuss potential solutions to the practical challenges, and identify 22 open research challenges.
△ Less
Submitted 5 July, 2021;
originally announced July 2021.
-
Likelihood Analysis of Supersymmetric SU(5) GUTs
Authors:
E. Bagnaschi,
J. C. Costa,
K. Sakurai,
M. Borsato,
O. Buchmueller,
R. Cavanaugh,
V. Chobanova,
M. Citron,
A. De Roeck,
M. J. Dolan,
J. R. Ellis,
H. Flächer,
S. Heinemeyer,
G. Isidori,
M. Lucio,
D. Martínez Santos,
K. A. Olive,
A. Richards,
K. J. de Vries,
G. Weiglein
Abstract:
We perform a likelihood analysis of the constraints from accelerator experiments and astrophysical observations on supersymmetric (SUSY) models with SU(5) boundary conditions on soft SUSY-breaking parameters at the GUT scale. The parameter space of the models studied has 7 parameters: a universal gaugino mass $m_{1/2}$, distinct masses for the scalar partners of matter fermions in five- and ten-di…
▽ More
We perform a likelihood analysis of the constraints from accelerator experiments and astrophysical observations on supersymmetric (SUSY) models with SU(5) boundary conditions on soft SUSY-breaking parameters at the GUT scale. The parameter space of the models studied has 7 parameters: a universal gaugino mass $m_{1/2}$, distinct masses for the scalar partners of matter fermions in five- and ten-dimensional representations of SU(5), $m_5$ and $m_{10}$, and for the $\mathbf{5}$ and $\mathbf{\bar 5}$ Higgs representations $m_{H_u}$ and $m_{H_d}$, a universal trilinear soft SUSY-breaking parameter $A_0$, and the ratio of Higgs vevs $\tan β$. In addition to previous constraints from direct sparticle searches, low-energy and flavour observables, we incorporate constraints based on preliminary results from 13 TeV LHC searches for jets + MET events and long-lived particles, as well as the latest PandaX-II and LUX searches for direct Dark Matter detection. In addition to previously-identified mechanisms for bringing the supersymmetric relic density into the range allowed by cosmology, we identify a novel ${\tilde u_R}/{\tilde c_R} - \tildeχ^0_1$ coannihilation mechanism that appears in the supersymmetric SU(5) GUT model and discuss the role of ${\tilde ν_τ}$ coannihilation. We find complementarity between the prospects for direct Dark Matter detection and SUSY searches at the LHC.
△ Less
Submitted 26 April, 2017; v1 submitted 31 October, 2016;
originally announced October 2016.
-
Supersymmetric Dark Matter after LHC Run 1
Authors:
E. A. Bagnaschi,
O. Buchmueller,
R. Cavanaugh,
M. Citron,
A. De Roeck,
M. J. Dolan,
J. R. Ellis,
H. Flaecher,
S. Heinemeyer,
G. Isidori,
S. Malik,
D. Martinez Santos,
K. A. Olive,
K. Sakurai,
K. J. de Vries,
G. Weiglein
Abstract:
Different mechanisms operate in various regions of the MSSM parameter space to bring the relic density of the lightest neutralino, neutralino_1, assumed here to be the LSP and thus the Dark Matter (DM) particle, into the range allowed by astrophysics and cosmology. These mechanisms include coannihilation with some nearly-degenerate next-to-lightest supersymmetric particle (NLSP) such as the lighte…
▽ More
Different mechanisms operate in various regions of the MSSM parameter space to bring the relic density of the lightest neutralino, neutralino_1, assumed here to be the LSP and thus the Dark Matter (DM) particle, into the range allowed by astrophysics and cosmology. These mechanisms include coannihilation with some nearly-degenerate next-to-lightest supersymmetric particle (NLSP) such as the lighter stau (stau_1), stop (stop_1) or chargino (chargino_1), resonant annihilation via direct-channel heavy Higgs bosons H/A, the light Higgs boson h or the Z boson, and enhanced annihilation via a larger Higgsino component of the LSP in the focus-point region. These mechanisms typically select lower-dimensional subspaces in MSSM scenarios such as the CMSSM, NUHM1, NUHM2 and pMSSM10. We analyze how future LHC and direct DM searches can complement each other in the exploration of the different DM mechanisms within these scenarios. We find that the stau_1 coannihilation regions of the CMSSM, NUHM1, NUHM2 can largely be explored at the LHC via searches for missing E_T events and long-lived charged particles, whereas their H/A funnel, focus-point and chargino_1 coannihilation regions can largely be explored by the LZ and Darwin DM direct detection experiments. We find that the dominant DM mechanism in our pMSSM10 analysis is chargino_1 coannihilation: {parts of its parameter space can be explored by the LHC, and a larger portion by future direct DM searches.
△ Less
Submitted 5 August, 2015;
originally announced August 2015.
-
The pMSSM10 after LHC Run 1
Authors:
K. J. de Vries,
E. A. Bagnaschi,
O. Buchmueller,
R. Cavanaugh,
M. Citron,
A. De Roeck,
M. J. Dolan,
J. R. Ellis,
H. Flaecher,
S. Heinemeyer,
G. Isidori,
S. Malik,
J. Marrouche,
D. Martinez Santos,
K. A. Olive,
K. Sakurai,
G. Weiglein
Abstract:
We present a frequentist analysis of the parameter space of the pMSSM10, in which the following 10 soft SUSY-breaking parameters are specified independently at the mean scalar top mass scale Msusy = Sqrt[M_stop1 M_stop2]: the gaugino masses M_{1,2,3}, the 1st-and 2nd-generation squark masses M_squ1 = M_squ2, the third-generation squark mass M_squ3, a common slepton mass M_slep and a common triline…
▽ More
We present a frequentist analysis of the parameter space of the pMSSM10, in which the following 10 soft SUSY-breaking parameters are specified independently at the mean scalar top mass scale Msusy = Sqrt[M_stop1 M_stop2]: the gaugino masses M_{1,2,3}, the 1st-and 2nd-generation squark masses M_squ1 = M_squ2, the third-generation squark mass M_squ3, a common slepton mass M_slep and a common trilinear mixing parameter A, the Higgs mixing parameter mu, the pseudoscalar Higgs mass M_A and tan beta. We use the MultiNest sampling algorithm with 1.2 x 10^9 points to sample the pMSSM10 parameter space. A dedicated study shows that the sensitivities to strongly-interacting SUSY masses of ATLAS and CMS searches for jets, leptons + MET signals depend only weakly on many of the other pMSSM10 parameters. With the aid of the Atom and Scorpion codes, we also implement the LHC searches for EW-interacting sparticles and light stops, so as to confront the pMSSM10 parameter space with all relevant SUSY searches. In addition, our analysis includes Higgs mass and rate measurements using the HiggsSignals code, SUSY Higgs exclusion bounds, the measurements B-physics observables, EW precision observables, the CDM density and searches for spin-independent DM scattering. We show that the pMSSM10 is able to provide a SUSY interpretation of (g-2)_mu, unlike the CMSSM, NUHM1 and NUHM2. As a result, we find (omitting Higgs rates) that the minimum chi^2/dof = 20.5/18 in the pMSSM10, corresponding to a chi^2 probability of 30.8 %, to be compared with chi^2/dof = 32.8/24 (31.1/23) (30.3/22) in the CMSSM (NUHM1) (NUHM2). We display 1-dimensional likelihood functions for SUSY masses, and show that they may be significantly lighter in the pMSSM10 than in the CMSSM, NUHM1 and NUHM2. We discuss the discovery potential of future LHC runs, e+e- colliders and direct detection experiments.
△ Less
Submitted 13 April, 2015;
originally announced April 2015.
-
SUSY fits with full LHC Run I data
Authors:
Kees Jan de Vries
Abstract:
We present the latest results from the MasterCode Collaboration on supersymmetric models, in particular on the CMSSM, the NUHM1, the NUHM2 and the pMSSM. We combine the data from LHC Run I with astrophysical observables, flavor and electroweak precision observables. We determine the best fit regions of these models and analyze the discovery potential of squarks and gluinos at LHC Run II and direct…
▽ More
We present the latest results from the MasterCode Collaboration on supersymmetric models, in particular on the CMSSM, the NUHM1, the NUHM2 and the pMSSM. We combine the data from LHC Run I with astrophysical observables, flavor and electroweak precision observables. We determine the best fit regions of these models and analyze the discovery potential of squarks and gluinos at LHC Run II and direct detection experiments.
△ Less
Submitted 24 October, 2014;
originally announced October 2014.
-
The NUHM2 after LHC Run 1
Authors:
O. Buchmueller,
R. Cavanaugh,
M. Citron,
A. De Roeck,
M. J. Dolan,
J. R. Ellis,
H. Flaecher,
S. Heinemeyer,
S. Malik,
J. Marrouche,
D. Martinez Santos,
K. A. Olive,
K. J. De Vries,
G. Weiglein
Abstract:
We make a frequentist analysis of the parameter space of the NUHM2, in which the soft supersymmetry (SUSY)-breaking contributions to the masses of the two Higgs multiplets, $m^2_{H_{u,d}}$, vary independently from the universal soft SUSY-breaking contributions $m^2_0$ to the masses of squarks and sleptons. Our analysis uses the MultiNest sampling algorithm with over $4 \times 10^8$ points to sampl…
▽ More
We make a frequentist analysis of the parameter space of the NUHM2, in which the soft supersymmetry (SUSY)-breaking contributions to the masses of the two Higgs multiplets, $m^2_{H_{u,d}}$, vary independently from the universal soft SUSY-breaking contributions $m^2_0$ to the masses of squarks and sleptons. Our analysis uses the MultiNest sampling algorithm with over $4 \times 10^8$ points to sample the NUHM2 parameter space. It includes the ATLAS and CMS Higgs mass measurements as well as their searches for supersymmetric jets + MET signals using the full LHC Run~1 data, the measurements of $B_s \to μ^+ μ^-$ by LHCb and CMS together with other B-physics observables, electroweak precision observables and the XENON100 and LUX searches for spin-independent dark matter scattering. We find that the preferred regions of the NUHM2 parameter space have negative SUSY-breaking scalar masses squared for squarks and sleptons, $m_0^2 < 0$, as well as $m^2_{H_u} < m^2_{H_d} < 0$. The tension present in the CMSSM and NUHM1 between the supersymmetric interpretation of $g_μ- 2$ and the absence to date of SUSY at the LHC is not significantly alleviated in the NUHM2. We find that the minimum $χ^2 = 32.5$ with 21 degrees of freedom (dof) in the NUHM2, to be compared with $χ^2/{\rm dof} = 35.0/23$ in the CMSSM, and $χ^2/{\rm dof} = 32.7/22$ in the NUHM1. We find that the one-dimensional likelihood functions for sparticle masses and other observables are similar to those found previously in the CMSSM and NUHM1.
△ Less
Submitted 18 August, 2014;
originally announced August 2014.
-
The CMSSM and NUHM1 after LHC Run 1
Authors:
O. Buchmueller,
R. Cavanaugh,
A. De Roeck,
M. J. Dolan,
J. R. Ellis,
H. Flacher,
S. Heinemeyer,
G. Isidori,
J. Marrouche,
D. Martinez Santos,
K. A. Olive,
S. Rogerson,
F. J. Ronga,
K. J. de Vries,
G. Weiglein
Abstract:
We analyze the impact of data from the full Run 1 of the LHC at 7 and 8 TeV on the CMSSM with mu > 0 and < 0 and the NUHM1 with mu > 0, incorporating the constraints imposed by other experiments such as precision electroweak measurements, flavour measurements, the cosmological density of cold dark matter and the direct search for the scattering of dark matter particles in the LUX experiment. We us…
▽ More
We analyze the impact of data from the full Run 1 of the LHC at 7 and 8 TeV on the CMSSM with mu > 0 and < 0 and the NUHM1 with mu > 0, incorporating the constraints imposed by other experiments such as precision electroweak measurements, flavour measurements, the cosmological density of cold dark matter and the direct search for the scattering of dark matter particles in the LUX experiment. We use the following results from the LHC experiments: ATLAS searches for events with MET accompanied by jets with the full 7 and 8 TeV data, the ATLAS and CMS measurements of the mass of the Higgs boson, the CMS searches for heavy neutral Higgs bosons and a combination of the LHCb and CMS measurements of B_s to mu+mu- and B_d to mu+mu-. Our results are based on samplings of the parameter spaces of the CMSSM for both mu>0 and mu<0 and of the NUHM1 for mu > 0 with 6.8 x 10^6, 6.2 x 10^6 and 1.6 x 10^7 points, respectively, obtained using the MultiNest tool. The impact of the Higgs mass constraint is assessed using FeynHiggs 2.10.0, which provides an improved prediction for the masses of the MSSM Higgs bosons in the region of heavy squark masses. It yields in general larger values of M_h than previous versions of FeynHiggs, reducing the pressure on the CMSSM and NUHM1. We find that the global chi^2 functions for the supersymmetric models vary slowly over most of the parameter spaces allowed by the Higgs mass and the MET searches, with best-fit values that are comparable to the chi^2/dof for the best Standard Model fit. We provide 95% CL lower limits on the masses of various sparticles and assess the prospects for observing them during Run 2 of the LHC.
△ Less
Submitted 18 December, 2013;
originally announced December 2013.
-
The End of the CMSSM Coannihilation Strip is Nigh
Authors:
M. Citron,
J. Ellis,
F. Luo,
J. Marrouche,
K. A. Olive,
K. J. de Vries
Abstract:
A recent global fit to the CMSSM incorporating current constraints on supersymmetry, including missing transverse energy searches at the LHC, BR(B_s to mu+ mu-) and the direct XENON100 search for dark matter, favours points towards the end of the stau-neutralino (stau_1- chi) coannihilation strip with relatively large m_1/2 and 10 < tan beta < 40 and points in the H/A rapid-annihilation funnel wit…
▽ More
A recent global fit to the CMSSM incorporating current constraints on supersymmetry, including missing transverse energy searches at the LHC, BR(B_s to mu+ mu-) and the direct XENON100 search for dark matter, favours points towards the end of the stau-neutralino (stau_1- chi) coannihilation strip with relatively large m_1/2 and 10 < tan beta < 40 and points in the H/A rapid-annihilation funnel with tan beta ~ 50. The coannihilation points typically have m_stau_1-m_chi < 5 GeV, and a significant fraction, including the most-favoured point, has m_stau_1-m_chi < m_tau. In such a case, the stau_1 lifetime would be so long that the stau_1 would be detectable as a long-lived massive charged particle that may decay inside or outside the apparatus. We show that CMSSM scenarios close to the tip of the coannihilation strip for tan beta < 40 are already excluded by LHC searches for massive charged particles, and discuss the prospects for their detection in the CMS and ATLAS detectors via time-of-flight measurements, anomalous heavy ionization or decays into one or more soft charged particles.
△ Less
Submitted 12 December, 2012;
originally announced December 2012.
-
The CMSSM and NUHM1 in Light of 7 TeV LHC, B_s to mu+mu- and XENON100 Data
Authors:
O. Buchmueller,
R. Cavanaugh,
M. Citron,
A. De Roeck,
M. J. Dolan,
J. R. Ellis,
H. Flacher,
S. Heinemeyer,
G. Isidori,
J. Marrouche,
D. Martinez Santos,
S. Nakach,
K. A. Olive,
S. Rogerson,
F. J. Ronga,
K. J. de Vries,
G. Weiglein
Abstract:
We make a frequentist analysis of the parameter space of the CMSSM and NUHM1, using a Markov Chain Monte Carlo (MCMC) with 95 (221) million points to sample the CMSSM (NUHM1) parameter spaces. Our analysis includes the ATLAS search for supersymmetric jets + MET signals using ~ 5/fb of LHC data at 7 TeV, which we apply using PYTHIA and a Delphes implementation that we validate in the relevant param…
▽ More
We make a frequentist analysis of the parameter space of the CMSSM and NUHM1, using a Markov Chain Monte Carlo (MCMC) with 95 (221) million points to sample the CMSSM (NUHM1) parameter spaces. Our analysis includes the ATLAS search for supersymmetric jets + MET signals using ~ 5/fb of LHC data at 7 TeV, which we apply using PYTHIA and a Delphes implementation that we validate in the relevant parameter regions of the CMSSM and NUHM1. Our analysis also includes the constraint imposed by searches for B_s to mu+mu- by LHCb, CMS, ATLAS and CDF, and the limit on spin-independent dark matter scattering from 225 live days of XENON100 data. We assume M_h ~ 125 GeV, and use a full set of electroweak precision and other flavour-physics observables, as well as the cold dark matter density constraint. The ATLAS 5/fb constraint has relatively limited effects on the 68 and 95% CL regions in the (m_0, m_1/2) planes of the CMSSM and NUHM1. The new B_s to mu+mu- constraint has greater impacts on these CL regions, and also impacts significantly the 68 and 95% CL regions in the (M_A, tan beta) planes of both models, reducing the best-fit values of tan beta. The recent XENON100 data eliminate the focus-point region in the CMSSM and affect the 68 and 95% CL regions in the NUHM1. In combination, these new constraints reduce the best-fit values of m_0, m_1/2 in the CMSSM, and increase the global chi^2 from 31.0 to 32.8, reducing the p-value from 12% to 8.5%. In the case of the NUHM1, they have little effect on the best-fit values of m_0, m_1/2, but increase the global chi^2 from 28.9 to 31.3, thereby reducing the p-value from 15% to 9.1%.
△ Less
Submitted 31 July, 2012;
originally announced July 2012.
-
Higgs and Supersymmetry
Authors:
O. Buchmueller,
R. Cavanaugh,
A. De Roeck,
M. J. Dolan,
J. R. Ellis,
H. Flacher,
S. Heinemeyer,
G. Isidori,
J. Marrouche,
D. Martinez Santos,
K. A. Olive,
S. Rogerson,
F. J. Ronga,
K. J. de Vries,
G. Weiglein
Abstract:
Global frequentist fits to the CMSSM and NUHM1 using the
MasterCode framework predicted m_h \simeq 119 GeV in fits incorporating the g_mu-2 constraint and \simeq 126 GeV without it. Recent results by ATLAS and CMS could be compatible with a Standard Model-like Higgs boson around m_h \simeq 125 GeV. We use the previous MasterCode analysis to calculate the likelihood for a measurement of any nomin…
▽ More
Global frequentist fits to the CMSSM and NUHM1 using the
MasterCode framework predicted m_h \simeq 119 GeV in fits incorporating the g_mu-2 constraint and \simeq 126 GeV without it. Recent results by ATLAS and CMS could be compatible with a Standard Model-like Higgs boson around m_h \simeq 125 GeV. We use the previous MasterCode analysis to calculate the likelihood for a measurement of any nominal Higgs mass within the range of 115 to 130 GeV. Assuming a Higgs mass measurement at m_h \simeq 125 GeV, we display updated global likelihood contours in the (m_0, m_{1/2}) and other parameter planes of the CMSSM and NUHM1, and present updated likelihood functions for m_gluino, m_squark, B to mu mu, and the spin-independent dark matter cross section σ^si. The implications of drop** g_mu-2 from the fits are also discussed. We furthermore comment on a hypothetical measurement of m_h \simeq 119 GeV.
△ Less
Submitted 24 May, 2012; v1 submitted 15 December, 2011;
originally announced December 2011.
