IQNN-CS: Interpretable Quantum Neural Network for Credit Scoring

Abdul Samad Khan, Nouhaila Innan, Aeysha Khalique, Muhammad ShafiquearXiv 2025 · DOI · arXiv

Introduction

In the financial sector, credit scoring is a high-stakes application where automated decisions must be both accurate and explainable. Regulatory frameworks such as the GDPR and Basel III mandate a "right to explanation," particularly when loan applications are denied. While Quantum Machine Learning (QML) has shown promise in processing high-dimensional financial data, these models are often viewed as "black boxes." IQNN-CS addresses this challenge by developing a hybrid classical-quantum framework that prioritizes transparency alongside performance.

Methodology

The IQNN-CS framework follows a comprehensive five-stage pipeline designed to ensure both predictive power and interpretability.

The IQNN-CS hybrid pipeline architecture, spanning data preparation to evaluation.

The architecture integrates classical preprocessing with a Variational Quantum Circuit (VQC). Data is mapped into the quantum state space using Angle Encoding, where financial features determine the rotation of qubits. The model's interpretability is supported by a suite of tools, including gradient-based attribution, prototype matching, and t-SNE latent space visualization.

A core contribution of this work is the Inter-Class Attribution Alignment (ICAA) metric. ICAA quantifies the distinctness of the model's reasoning across different classes. If the model utilizes similar feature attributions for both "High Risk" and "Low Risk" predictions, the metric flags a potential lack of robust underlying logic, even if the final prediction is correct.

Experimental Setup

The model was rigorously tested on two distinct credit datasets:

  • Dataset 1: A well-structured benchmark used to evaluate the model's ability to define clean decision boundaries.
  • Dataset 2: A complex, imbalanced real-world dataset designed to test robustness and interpretability under noisy conditions.

Results and Discussion

The experimental results highlight the relationship between data structure and quantum model performance.

Training dynamics for Dataset 1 (a, b) and Dataset 2 (c, d), showing accuracy and loss curves.

On Dataset 1, the model achieved near-perfect performance with 100% accuracy and F1-score. The training curves (Figures 3a and 3b) show rapid, stable convergence, indicating that the hybrid architecture effectively captured the dataset's patterns.

Conversely, Dataset 2 presented significant challenges. While the model maintained a respectable 77.3% accuracy, the training process was more erratic (Figures 3c and 3d). The ICAA metric proved vital here, identifying "regions of indecision" where the model's internal logic was less stable. This suggests that while accuracy remains a key metric, it can be deceptive; high accuracy without aligned internal logic (as flagged by ICAA) may indicate a model that is not yet ready for regulated production environments.

Conclusion

IQNN-CS demonstrates that quantum neural networks can be made interpretable for sensitive financial applications. The introduction of the ICAA metric provides a necessary tool for auditing the internal reasoning of QML models. Future research will focus on more advanced quantum embedding techniques to reduce reliance on classical dimensionality reduction and testing the framework on larger, production-grade datasets.

Summary updated

BibTeX
@misc{khan2025iqnn,
  author = {Abdul Samad Khan and Nouhaila Innan and Aeysha Khalique and Muhammad Shafique},
  title = {{IQNN-CS: Interpretable Quantum Neural Network for Credit Scoring}},
  year = {2025},
  eprint = {2510.15044},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.15044},
}
Figure