Extracting reliable information from collider data requires a combination of mathematical, computational and statistical tools to model the observed distributions, often with the help of powerful simulations. But when simulations cannot be fully trusted, data-driven approaches become indispensable. In this talk, I will present probabilistic and Bayesian tools that provide a flexible, data-driven method for unsupervised training on probabilistic models of collider observables, highlighting their advantages for parameter inference and uncertainty estimation. I will illustrate these points in the context of pp > hh > bbbb searches, where recent analyses use a variation of the widely used data-driven "ABCD method". Within our framework, the ABCD method can be generalized and improved, especially when its assumptions are not exactly satisfied, leading to more robust background subtraction. Our approach is able to exploit correlations in multi-dimensional data by modelling kinematic variables and b-tagging scores at the event level. These results also show a way in which simulations can be used as priors, guiding the learning process for unsupervised estimation of the true data distributions. Based on 2402.08001 and 2511.21832.

Speaker Bio: Santiago Tanco recently obtained his PhD at the National University of La Plata (UNLP), Argentina, under the supervision of Aníbal Medina and Alejandro Szynkman. Following his doctoral defense, he is undertaking a series of research visits across Europe. His research focuses on particle phenomenology, with an emphasis on collider signatures of SM and BSM scenarios including dark matter, Z' bosons, and long-lived particles. In particular, he is interested in the intersection of particle physics with the development and application of machine learning and statistical methods for collider data analysis.