Topic Modeling in New Physics Detection

TL;DR

This work addresses model-independent, unsupervised searches for new physics at the LHC by applying topic modeling to events with fully leptonic $t\bar{t}$ backgrounds. It compares three TM methods—$\text{LDA}$, Biterm (BTM), and ProdLDA—using discretized final-state observables and benchmarks against isolation forests and VAEs. In resonant scenarios like $H\to hh$ or $H\to t\bar{t}$, Biterm achieves $\mathrm{AUC}$ values up to around $0.98$, while non-resonant $hh$ and dimension-six operator cases are more challenging with moderate discrimination; some operators yield near-random separation for some models. The results demonstrate that topic modeling can uncover physically meaningful latent patterns using only final-state kinematic variables, offering a robust, model-independent complement to jet-substructure analyses for new-physics searches at the LHC.

Abstract

In this work, we apply topic modeling to detect new physics in proton-proton collisions at the LHC in an unsupervised way. We investigate three new physics scenarios where fully leptonic $t\bar{t}\to b\bar{b}\ell^+\ell^-ν_\ell\barν_\ell$ is the main source of background without relying on jet substructure variables. We demonstrate that the algorithm remains effective even in this low-particle multiplicity framework, complementing jet tagging studies, where it is typically employed. Moreover, we demonstrate that the performance of topic modeling is competitive or even better than well-known outlier detectors, such as isolation forest and variational autoencoders, with moderate and high background pollution in almost all new physics scenarios considered.

Figures (13)

• Figure 1: The normalized distributions of the kinematic variables used for the search for the best representation of the data. The signal events correspond to $H\to hh$. The $\Delta R$ variable is used in the computation of $k_T$ but not included explicitly in the representation of the data.
• Figure 2: Comparison among the ROC and Background Rejection curves from topic modeling models: LDA, biterm, and ProdLDA. The left column shows the case of 50% of background pollution, while the right side shows the 99% scenario. The upper figures show the ROC curves, which represent the relationship between the true positive rate (TPR) and false positive rate (FPR) across many different classification thresholds. The lower figures show the background rejection, which relates the background rejection (inverse of FPR) to the signal efficiency (TPR). The legends show the AUC values, which are used to discriminate the ability of each model in disentangling the signal from the background.
• Figure 3: Comparison among the ROC and Background Rejection curves from biterm, isolation forest, and VAE models in $H\xrightarrow{}hh$ process. The left column shows the case of 50% of background pollution, while the right side shows the 99% scenario. The upper figures show the ROC curves, which represent the relationship between the true positive rate and false positive rate across many different classification thresholds. The lower figures show the background rejection, which relates the background rejection to the signal efficiency. The legends show the AUC values, which are used to discriminate the ability of each model in disentangling the signal from the background.
• Figure 4: Reconstruction of the invariant mass for $H\xrightarrow{}t\bar{t}$ using BTM with 50% background pollution. Upper plot shows the distribution of all combinations together. Each mini-plot is the distribution of each possible combination. The filled histograms refers to the event simulated data, while the line histograms refers to the topic classification.
• Figure 5: Reconstruction of the invariant mass for $H\xrightarrow{}t\bar{t}$ using BTM for 50% background contamination. Continuation of the figure \ref{['fig:mass_reconstructed_Hhh_50_1']}.