Does SUSY have friends? A new approach for LHC event analysis

TL;DR

This paper introduces a graph network-based method for analyzing LHC collision events, improving signal-background separation and background estimation in searches for supersymmetric particles.

Contribution

It presents a novel network metric approach for event analysis that outperforms traditional methods and can handle differing Monte Carlo and real data event counts.

Findings

Network variables outperform cut-and-count and BDT methods in electroweakino analysis.

The approach enables accurate background estimation despite differing MC and real data.

Network metrics show potential for challenging stop searches with further validation.

Abstract

We present a novel technique for the analysis of proton-proton collision events from the ATLAS and CMS experiments at the Large Hadron Collider. For a given final state and choice of kinematic variables, we build a graph network in which the individual events appear as weighted nodes, with edges between events defined by their distance in kinematic space. We then show that it is possible to calculate local metrics of the network that serve as event-by-event variables for separating signal and background processes, and we evaluate these for a number of different networks that are derived from different distance metrics. Using a supersymmetric electroweakino and stop production as examples, we construct prototype analyses that take account of the fact that the number of simulated Monte Carlo events used in an LHC analysis may differ from the number of events expected in the LHC dataset, allowing an accurate background estimate for a particle search at the LHC to be derived. For the electroweakino example, we show that the use of network variables outperforms both cut-and-count analyses that use the original variables and a boosted decision tree trained on the original variables. The stop example, deliberately chosen to be difficult to exclude due its kinematic similarity with the top background, demonstrates that network variables are not automatically sensitive to BSM physics. Nevertheless, we identify local network metrics that show promise if their robustness under certain assumptions of node-weighted networks can be confirmed.