Dissecting flavor tagging with graph neural networks
Abstract
Graph neural networks such as ParticleNet are widely used for quark/gluon jet tagging. We study how the discriminating power of such a network is distributed between constituent multiplicity and the remaining structure of the jet, differentially in jet transverse momentum and multiplicity. To this end, we train ParticleNet on quark and gluon jets from Z+jet production in Pythia, with the two classes matched in transverse momentum and rapidity over the wide range 80≤ pTjet≤ 400 GeV. We find that the network learns how the fraction of gluon jets changes with the multiplicity in every pTjet bin, including the shift of the multiplicity at which quark and gluon jets are equally likely, which implies that it uses the jet scale (reconstructed from its constituents) jointly with its substructure. At fixed multiplicity, the gluon probabilities it assigns agree on average with the observed gluon fractions, but are widely spread about the mean. This allows us to interpret its output as the sum of the multiplicity log-likelihood ratio and a residual log-likelihood ratio at fixed multiplicity, which allows the network to correctly identify 20-34% of the jets whose multiplicity favors the opposite class. By rescaling the radial spread of the constituents at fixed multiplicity, we find that the network considers broader jets more gluon-like, and that this response is largest at low and intermediate multiplicity, where the multiplicity alone does not decide the class of a jet.
Related
- Quantum Graph Neural Networks for Jet Tagging on Quantum Hardware
- Using Graph Neural Networks for the segmentation of overlapping objects in high granularity calorimeters
- Emergent Loop Geometry in Gauge-Equivariant Graph Neural Networks
- Graph neural networks and the energetic cavity method for combinatorial optimization
- Non-linear mechanical field reconstruction coupling recurrent neural networks with physics-informed graph neural networks
BibTeX
@misc{budhraja2026dissecting,
title = {{Dissecting flavor tagging with graph neural networks}},
author = {Ankita Budhraja and Sunny Kumar Singh and Balbeer Singh},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.33757},
eprint = {2610.33757},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.33757},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Dissecting flavor tagging with graph neural networks
AU - Ankita Budhraja
AU - Sunny Kumar Singh
AU - Balbeer Singh
PY - 2026
DA - 2026/10/24
PB - arXiv
DO - 10.48550/arxiv.2610.33757
UR - https://arxiv.org/abs/2610.33757
AB - Graph neural networks such as ParticleNet are widely used for quark/gluon jet tagging. We study how the discriminating power of such a network is distributed between constituent multiplicity and the remaining structure of the jet, differentially in jet transverse momentum and multiplicity. To this end, we train ParticleNet on quark and gluon jets from Z+jet production in Pythia, with the two classes matched in transverse momentum and rapidity over the wide range 80≤ p_T^jet≤ 400 GeV. We find that the network learns how the fraction of gluon jets changes with the multiplicity in every p_T^jet bin, including the shift of the multiplicity at which quark and gluon jets are equally likely, which implies that it uses the jet scale (reconstructed from its constituents) jointly with its substructure. At fixed multiplicity, the gluon probabilities it assigns agree on average with the observed gluon fractions, but are widely spread about the mean. This allows us to interpret its output as the sum of the multiplicity log-likelihood ratio and a residual log-likelihood ratio at fixed multiplicity, which allows the network to correctly identify 20-34% of the jets whose multiplicity favors the opposite class. By rescaling the radial spread of the constituents at fixed multiplicity, we find that the network considers broader jets more gluon-like, and that this response is largest at low and intermediate multiplicity, where the multiplicity alone does not decide the class of a jet.
ER -