KIGNet: Physics-Motivated Multi-Graph Representation Learning for Explainable Jet Tagging

Jet tagging, the task of identifying the initiating particle of a jet produced in high-energy collisions, is essential for precision measurements and new-physics searches at the Large Hadron Collider. While deep learning has improved jet-tagging accuracy, existing models often lack interpretability. This project proposes KIGNet, a physics-motivated multi-graph neural network that represents jets using four kinematic quantities: angular separation, relative transverse momentum, momentum fraction, and invariant mass. By assigning each quantity its own weighted graph and neural network, we aim to improve interpretability while maintaining predictive performance. Grad-CAM will be used to evaluate feature contributions against QCD expectations, and the framework will be validated using real CMS collision data. An early version of this work is available on arXiv.

Team: MD Raqibul Islam, Adrita Khan, Mir Sazzat Hossain, Choudhury Ben Yamin Siddiqui, Md. Zakir Hossan, Tanjib Khan, M Arshad Momen, PhD, Prof. Amin Ahsan Ali, PhD, and AKM Mahbubur Rahman, PhD.