Q-Trace: A Noise-Aware Quantum Feature Attribution and Explanation Stability Framework for Hybrid Learning Models
Keywords:
Quantum Machine Learning, Explainable AI, Variational Quantum Circuits, Feature Attribution, NISQ Noise, Shapley Values, Parameter-shift Rule, Explanation StabilityAbstract
Hybrid Hybrid quantum-classical learning models increasingly rely on parameterized quantum circuits to encode classical data into high-dimensional Hilbert spaces. However, the explainability of these models under realistic hardware noise remains poorly understood. Existing feature-attribution techniques adapted from classical machine learning generally assume noiseless, deterministic function evaluation. This assumption is violated on near-term quantum devices, where gate errors, decoherence, and finite-shot sampling affect the expectation values used to construct explanations. This paper introduces Q-Trace, a framework that formulates feature attribution as an operator acting on a noisy quantum channel rather than on an idealized circuit, while explicitly quantifying the extent to which explanations may drift as noise increases. Q-Trace comprises three components: Noise-Aware Quantum Gradient Attribution (NQGA), which extends the parameter-shift rule to noisy quantum channels; Noise-Conditioned Shapley Estimation (NCSE), which computes exact Shapley-value attributions directly from the noisy expectation-value function; and the Explanation Stability Index (ESI), a similarity-based metric for measuring attribution drift across noise levels. To the best of our knowledge, Q-Trace is the first framework to explicitly incorporate a parameterized noise channel into input-feature attribution, evaluate gradient-based and coalition-based attribution operators under the same noise channel, and quantify explanation stability through a dedicated metric. We derive a closed-form bound relating ESI degradation to the contraction rate of the underlying depolarizing channel and validate the framework using a four-qubit hybrid variational classifier simulated exactly at the density-matrix level. The results show that Shapley-based attributions are consistently more stable than gradient-based attributions. At a noise level of (p = 0.20), the cosine and rank ESI values for the Shapley approach reach 0.991 and 0.927, respectively, averaged across five trained models, compared with 0.963 and 0.854 for the gradient approach. Furthermore, finite-difference gradient estimators exhibit substantially greater sensitivity to shot noise, with the standard deviation of gradient estimates increasing by a factor of 52.15 on average across six shot budgets.
Downloads
Downloads
Published
How to Cite
Issue
Section
License
Copyright (c) 2026 Journal of Transactions in Systems Engineering

This work is licensed under a Creative Commons Attribution 4.0 International License.
