IQNN-CS: Interpretable Quantum Neural Network for Credit Scoring

Abdul Samad Khan, Nouhaila Innan, Aeysha Khalique, Muhammad ShafiqueCommunications 2026 · DOI · arXiv · OpenAlex

In the financial sector, credit scoring models do more than just process numbers; they make life-altering decisions that are subject to strict regulatory oversight. Frameworks like the EU’s General Data Protection Regulation (GDPR) and Basel III mandate a "right to explanation," requiring that automated decisions be transparent and interpretable. While Quantum Neural Networks (QNNs) are praised for their ability to represent complex data with fewer parameters, they are often criticized for being opaque "black boxes."

This paper introduces IQNN-CS, a hybrid quantum-classical framework designed to bridge the gap between the computational potential of Quantum Machine Learning (QML) and the practical, regulatory need for transparency in multiclass credit risk assessment.

The IQNN-CS Framework

The authors propose a comprehensive five-stage pipeline that integrates a hybrid model with a suite of interpretability tools. This pipeline ensures that the model's decisions are not only accurate but also explainable.

The IQNN-CS five-stage pipeline: Data Preparation, Hybrid Model, Training, Interpretability, and Evaluation.
  1. Hybrid Architecture: The model employs a "sandwich" approach. Classical preprocessing layers compress input data before it is fed into a Variational Quantum Circuit (VQC). The VQC utilizes angle encoding and strongly entangling layers to process features in a high-dimensional Hilbert space. Finally, a classical post-processing head generates the credit risk classification.
  2. Post-hoc Interpretability: The framework adapts several classical techniques to the quantum domain, including Saliency Maps, Integrated Gradients, and Occlusion Sensitivity.
  3. The ICAA Metric: A core innovation of this work is the Inter-Class Attribution Alignment (ICAA) metric. It quantifies the overlap in feature attribution between different predicted classes. A low ICAA score suggests the model has learned distinct, well-separated rationales for different risk categories, which is essential for regulatory compliance.

Experimental Results and Performance

The framework was evaluated on two real-world benchmark credit scoring datasets. Dataset 1 served as a baseline for ideal quantum learning behavior, while Dataset 2 provided a more complex, imbalanced environment to test the limits of the model and its diagnostic tools.

Performance comparison across datasets: (a, b) show rapid convergence and near-perfect accuracy on Dataset 1, while (c, d) illustrate the challenges and limited generalization encountered with Dataset 2.

On Dataset 1, IQNN-CS achieved 100% accuracy and F1-score, with visualizations showing perfectly separated manifolds in the quantum latent space. On the more challenging Dataset 2, the model maintained a respectable 77.3% accuracy. However, the interpretability tools revealed that the model's internal logic for Dataset 2 was more "diffuse" and "noisy," signaling that the raw accuracy might overstate the model's reliability.

Key Insights

  • Interpretability as a Diagnostic Tool: Beyond providing explanations for end-users, interpretability serves as a vital debugging tool. It can identify "regions of indecision" where model uncertainty is high, allowing these cases to be flagged for human review.
  • Quantum Inductive Bias: The study highlights that QNN performance is deeply tied to how well the data’s structure aligns with quantum encoding. When the data "fits" the quantum circuit, the model produces exceptionally clean decision patterns.
  • Validation of ICAA: The ICAA metric successfully distinguished between the two datasets, effectively flagging when a model is "guessing" based on shared noise rather than distinct features.

Limitations and Future Work

While IQNN-CS demonstrates the feasibility of interpretable QML in finance, challenges remain regarding scalability to datasets with hundreds of features and the impact of hardware noise on Noisy Intermediate-Scale Quantum (NISQ) devices. Future work will explore trainable quantum feature maps to automatically optimize data encoding for both accuracy and clarity.

Summary updated

BibTeX
@inproceedings{khan2026iqnn,
  author = {Abdul Samad Khan and Nouhaila Innan and Aeysha Khalique and Muhammad Shafique},
  title = {{IQNN-CS: Interpretable Quantum Neural Network for Credit Scoring}},
  booktitle = {Communications},
  year = {2026},
  doi = {10.1007/978-3-032-13855-2\_8},
}
Figure