Quantum System for Counterfactual Explanation Generation in Financial AI Models
Loading...
Files
Date
Journal Title
Journal ISSN
Volume Title
Publisher
Unuversity Setif 1 Ferhat Abbas . Faculty of Sciences
Abstract
Machine learning models are increasingly used in financial institutions to automate critical decisions such as credit scoring, fraud detection, and risk assessment. While these models achieve strong predictive performance, their black-box nature raises serious concerns regarding transparency and accountability, particularly in light of regulations such as GDPR Article 22 and MiFID II, which require that automated decisions affecting individuals be explainable and contestable.
Existing explainability methods such as LIME and SHAP identify which features influenced a decision but do not indicate what should be changed to obtain a different outcome. Counterfactual explanations address this gap by identifying the minimal modification to an applicant’s profile that would flip a refused decision to an approved one. However, classical counterfactual generation methods rely on random or gradient-based search, which provides no guarantee of finding a minimal, realistic solution to this inherently combinatorial problem.
This thesis proposes a hybrid quantum-classical framework for counterfactual explanation generation in financial decision systems. A Gradient Boosting Machine serves as the black-box credit-scoring oracle. A Variational Quantum Circuit, distilled from this oracle, provides a parameter-efficient validity scorer. The Quantum Approximate Optimization Algorithm then searches the combinatorial space of candidate feature modifications, guided by a cost Hamiltonian that structurally encodes regulatory constraints, including GDPR-mandated attribute immutability.
Evaluated on the German Credit dataset under a stratified, unbiased protocol, the proposed approach achieves an 11.1 % improvement in proximity over DiCE, the strongest classical baseline, and a 48.2 % improvement over a non-quantum ablation, while maintaining 100 % structural regulatory compliance. The framework is further validated under simulated NISQ hardware noise, supporting its near-term practical feasibility.
