Unsupervised quark/gluon jet tagging with Poissonian Mixture Models

TL;DR

This paper introduces an unsupervised learning method using Poissonian Mixture Models to classify quark and gluon jets, reducing reliance on simulations and providing interpretable results with competitive accuracy.

Contribution

It presents a novel unsupervised algorithm for quark-gluon jet tagging that estimates mixture parameters directly from data, minimizing theoretical biases.

Findings

Achieves quark-gluon classification accuracy of 0.65-0.7.

Performs well under simulated detector effects.

Enables hyperparameter tuning via unsupervised metrics.

Abstract

The classification of jets induced by quarks or gluons is important for New Physics searches at high-energy colliders. However, available taggers usually rely on modelling the data through Monte Carlo simulations, which could veil intractable theoretical and systematical uncertainties. To significantly reduce biases, we propose an unsupervised learning algorithm that, given a sample of jets, can learn the SoftDrop Poissonian rates for quark- and gluon-initiated jets and their fractions. We extract the Maximum Likelihood Estimates for the mixture parameters and the posterior probability over them. We then construct a quark-gluon tagger and estimate its accuracy in actual data to be in the $0.65-0.7$ range, below supervised algorithms but nevertheless competitive. We also show how relevant unsupervised metrics perform well, allowing for an unsupervised hyperparameter selection. Further, we find that this result is not affected by an angular smearing introduced to simulate detector effects for central jets. The presented unsupervised learning algorithm is simple; its result is interpretable and depends on very few assumptions.