IQNN-CS: Interpretable Quantum Neural Network for Credit Scoring

AI summary

Overview

Research area: Quantum machine learning (QML) applied to financial credit scoring, combined with post-hoc explainable AI for structured/tabular data.

Technical level: Intermediate. Readers need some familiarity with variational quantum circuits, parameter-shift gradients, and classical attribution methods, but the paper is written as an application-oriented study rather than a theoretical one.

Scope: The paper proposes and evaluates IQNN-CS, a hybrid classical–quantum pipeline for multiclass credit risk classification that is evaluated on two real-world credit datasets for predictive performance and, centrally, for interpretability.

What This Paper Is About

Credit scoring decides who gets loans, at what limits, and at what rates, so models used in this domain must satisfy regulators such as GDPR and Basel III that demand transparency. Quantum neural networks can represent complex decision boundaries over structured financial data, but they are essentially black boxes, and no standardized interpretability tooling exists for them, particularly for multiclass problems. This paper asks how a QNN can be designed for real-world credit scoring where accuracy alone is not sufficient, and answers with an interpretability-first architecture plus a new metric for auditing how distinctly a model reasons about different risk classes.

Key Contributions

1. A hybrid classical–quantum pipeline for structured credit data. A variational QNN performs multiclass classification on PCA-compressed financial features, wrapped in classical pre- and post-processing blocks, with interpretability provided by adapted classical and quantum explanation techniques.

2. Inter-Class Attribution Alignment (ICAA), a new metric. ICAA quantifies attribution divergence across predicted classes by computing cosine similarity between the attribution vectors of different classes, exposing whether the model uses distinct reasoning per credit-risk category or overlapping reasoning.

3. Evaluation on two real-world credit datasets. The framework is assessed for robustness, attribution behavior, and interpretability rather than for quantum advantage — the authors explicitly state the goal is not to demonstrate quantum supremacy or to beat classical machine learning models.

4. A comparative utility assessment of interpretability methods in a quantum setting, summarizing which techniques (saliency, gradient × input, integrated gradients, SmoothGrad, occlusion, ICAA, example-based attribution, indecision detection) remain reliable when the underlying model trains well versus poorly.

Main Findings

• Dataset 1 saturates at perfect performance: IQNN-CS reached 100% accuracy and 1.00 precision, recall, and F1 for both the Low and High classes. Training, validation, and test accuracy curves stabilized above 98% across all splits, with a sharp loss decrease and a smooth plateau and minimal train–validation discrepancy.

• Dataset 2 is substantially harder: overall accuracy was 77.3%. The Low class had precision 0.64, recall 0.97, and F1 0.77; the High class had precision 0.95, recall 0.67, and F1 0.79; averages were 0.73 precision, 0.84 recall, 0.78 F1. Validation and test accuracy plateaued below 80%, with slower and noisier loss decline and a wider train–validation gap.

• Latent quantum geometry mirrors predictive quality: t-SNE projections of QNN activations showed well-separated manifolds for Dataset 1, but entangled clusters for Dataset 2, particularly between the Average and High classes.

• Attribution patterns are concentrated when the model works, diffuse when it does not: saliency maps for a representative test instance were focused and interpretable on Dataset 1, while Dataset 2 produced diffuse, noisy attributions that varied across runs.

• Occlusion reveals reliance on informative features: occluding top-ranked features caused sharp confidence drops on Dataset 1, but gradual, unstructured degradation on Dataset 2.

• ICAA separates explanations on Dataset 1 and exposes overlap on Dataset 2: Dataset 1 yielded low inter-class attribution similarity, while Dataset 2 showed significant overlap, indicating a less distinct decision rationale across classes.

• Example-based attribution can break down under high confidence: for test sample 7, Dataset 1's five most influential training examples were all from the same class as the test sample (cosine similarities 0.9989, 0.9976, 0.9976, 0.9930, 0.9416). In Dataset 2, the top matches (indices 503, 19, 433, 588) were all from a different class at cosine similarity 1.0000, with only one same-class match (index 173, label 2, also 1.0000).

• Attribution stability flagged one indecisive case: under 20 random Gaussian perturbations, Sample 3 and Sample 7 were judged not indecisive, but Sample 12 in Dataset 2 had a saliency standard deviation of 0.2797 (versus 0.1426 on Dataset 1) and was flagged as indecisive.

• Prediction confidence tracked dataset structure: Dataset 1 produced sharper softmax distributions (lower uncertainty), Dataset 2 broader ones, consistent with noisier convergence and less reliable explanations.

• Method utility degrades with data difficulty: occlusion and ICAA were rated "Very High" utility on Dataset 1 and "Medium" on Dataset 2; saliency and integrated gradients were High on Dataset 1 and Medium on Dataset 2; gradient × input was Moderate then Low; SmoothGrad was Low on both; example-based attribution was High then Low.

Methodology in Plain English

The authors built a five-stage pipeline. First, two benchmark credit datasets with numerical and categorical financial attributes are preprocessed: class imbalance is handled by undersampling for Dataset 1 and SMOTE for Dataset 2, numerical features are standardized by subtracting the mean and dividing by the standard deviation, and PCA reduces the data to quantum-compatible vectors (6 components) matching the available qubit count. Stratified sampling splits each dataset 70/15/15 into train, validation, and test sets while preserving class distribution.

Second, the model is a three-part stack: a classical block that maps the PCA-compressed input into a latent vector through fully connected layers with ReLU activations; a quantum layer that angle-encodes those features onto 6 qubits initialized to |0⟩, applies a multi-layer entangling variational circuit (4 StronglyEntanglingLayers) with trainable angles, and measures Pauli-Z expectation values per qubit; and a classical post-processing head with dropout and linear projections that produces logits over the credit-risk classes. The whole system is end-to-end differentiable via PennyLane's TorchLayer and PyTorch autodifferentiation.

Third, training uses the negative log-likelihood loss (Table 2 lists a class-weighted cross-entropy loss), AdamW with decoupled weight decay, a learning rate of 0.01, a StepLR scheduler, batch size 16, 50 epochs, and early stopping on validation loss. Quantum gradients use the parameter-shift rule with shift s = π/2 for single-qubit gates.

Fourth, interpretability is assessed four ways: gradient-based attribution (saliency maps, gradient × input, integrated gradients, SmoothGrad); prototype matching in quantum feature space, where cosine similarity between a test instance's QNN activations and training activations retrieves similar historical cases; occlusion, where features are masked and confidence degradation is tracked; and the new ICAA metric, which builds a matrix of pairwise cosine similarities between per-class attribution vectors to show how distinguishable the model's reasoning is across classes. Regions of indecision are probed by perturbing inputs and measuring attribution variance.

Fifth, evaluation uses accuracy, macro F1-score, confusion matrices, and interpretability visualizations including t-SNE projections of quantum activations colored by true label. Everything ran on an Apple M3 CPU with 16 GB RAM using PennyLane's default.qubit simulator, without GPU acceleration. Dataset sizes, in terms of number of instances, are not reported in the paper content.

Why This Matters

Impact on research: Interpretability for QML is still in its infancy, and prior quantum credit-scoring work treated the task as binary classification or as an optimization problem, overlooking transparency. This paper reframes credit scoring as a multiclass QML problem and supplies a quantitative interpretability metric (ICAA) that can be applied to any multiclass model, quantum or classical. Its most provocative finding is methodological: a model can output high-confidence probabilities while its internal evidence points to the wrong class, so interpretability doubles as a diagnostic rather than just an explanation.

Real-world applications:

• Credit scoring and loan underwriting in regulated banking, where GDPR and Basel III require justification of automated decisions.
• Audit and compliance workflows that need to demonstrate why a particular applicant received a given risk tier.
• Model-risk management for institutions piloting quantum or hybrid models, using ICAA overlap as an early warning that class-level reasoning has collapsed.
• Interpretability-aware design in other high-stakes domains the authors name, such as healthcare, where the same tension between performance and transparency applies.

Industry relevance: The paper is explicit that it does not claim quantum advantage or superior accuracy over classical models. Its practical message is that QNNs can be made auditable at the level of individual decisions and that interpretability reliability is tightly coupled to training convergence quality and dataset structure — a dependency that any team deploying these models would need to monitor.

Future Directions

• Scaling beyond simulated small circuits. The study uses 6 qubits and the default.qubit simulator on a CPU; how ICAA, occlusion, and attribution stability behave on larger or real quantum hardware is an open question.

• Establishing classical baselines. No comparison against classical credit-scoring models is reported, so it remains unclear whether the interpretability findings are specific to QNNs or shared by any model trained on the same noisy data.

• Making noisy explanations more reliable. Gradient-based methods were inconsistent on the harder dataset, and SmoothGrad was rated low utility on both, suggesting a need for explanation techniques robust to unstable convergence.

• Generalizing ICAA. The metric is demonstrated on two credit datasets with a specific multiclass label scheme; testing it across other multiclass structured domains and other model families would clarify whether attribution overlap is a general signal of model confusion.

• Addressing data complexity and imbalance directly. Dataset 2's instability was attributed to feature diversity and class imbalance, which points to data-centric remedies (better balancing, feature engineering, or encoding choices) as complementary to architectural fixes.

Target Audience

Researchers working at the intersection of quantum machine learning and explainable AI; quantum finance practitioners evaluating whether QNNs are viable in regulated settings; model-risk, compliance, and audit professionals interested in quantitative tools for auditing opaque models; and graduate students or applied scientists with intermediate background in variational quantum circuits and PyTorch who want a concrete, fully specified example of an interpretability-first hybrid quantum–classical pipeline.

Authors’ abstract

Credit scoring is a high-stakes task in financial services, where model decisions directly impact individuals' access to credit and are subject to strict regulatory scrutiny. While Quantum Machine Learning (QML) offers new computational capabilities, its black-box nature poses challenges for adoption in domains that demand transparency and trust. In this work, we present IQNN-CS, an interpretable quantum neural network framework designed for multiclass credit risk classification. The architecture combines a variational QNN with a suite of post-hoc explanation techniques tailored for structured data. To address the lack of structured interpretability in QML, we introduce Inter-Class Attribution Alignment (ICAA), a novel metric that quantifies attribution divergence across predicted classes, revealing how the model distinguishes between credit risk categories. Evaluated on two real-world credit datasets, IQNN-CS demonstrates stable training dynamics, competitive predictive performance, and enhanced interpretability. Our results highlight a practical path toward transparent and accountable QML models for financial decision-making.