Qsvm Fraud Detection Eval

---
name: qsvm-fraud-detection-eval
description: Evaluates a quantum support vector machine (QSVM) with quantum feature selection for binary fraud detection on real-world card payment data. It probes the model's ability to identify fraudulent transactions using a balanced dataset and compares performance against classical feature selection baselines. Use when the user wants to benchmark on Real-world card payment data (Balanced Data Set), or asks about evaluating this task. Reports Accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2208.07963
  bibtex_key: grossi2022mixed
  confidence: medium
---

# qsvm-fraud-detection-eval

> Mixed Quantum-Classical Method For Fraud Detection with Quantum Feature Selection — Michele Grossi et al. (arXiv:2208.07963, 2022)

## What this evaluates

Evaluates a quantum support vector machine (QSVM) with quantum feature selection for binary fraud detection on real-world card payment data. It probes the model's ability to identify fraudulent transactions using a balanced dataset and compares performance against classical feature selection baselines.

## Datasets

- **Real-world card payment data (Balanced Data Set)** — total 2500; splits: train (1500), test (1000)

## Metrics

- `Accuracy` **(primary)** — range: [0, 1]
  - Fraction of correctly classified test instances out of the total test instances.
- `AUC` — range: [0, 1]
  - Area under the Receiver Operating Characteristic curve, measuring the model's ability to distinguish between classes across all classification thresholds.

## Input / output format

**Input**: Balanced card payment dataset features (e.g., F_15, F_42, etc.) with binary fraud labels.

**Output**: Binary class prediction (fraud vs. non-fraud) per transaction.

## Scoring recipe

```python
def compute_metrics(predictions, labels):
    accuracy = sum(p == l for p, l in zip(predictions, labels)) / len(labels)
    auc = roc_auc_score(labels, predictions)
    return accuracy, auc
```

## Common pitfalls

- The paper reports results averaged over 5 random trials in the text but 6 trials in Table 6, creating ambiguity in the exact repetition count.
- Performance is highly sensitive to the quantum backend simulation mode; noise models drastically reduce accuracy compared to ideal state-vector simulators.
- Feature selection is integral to the evaluation; swapping quantum-selected features for classical ones changes accuracy by ~0.02, so the protocol must specify the feature set used.

## Evidence (verbatim from paper)

> Due to data under-sampling, we have used 5 random trials to minimize bias. The average KPIs for accuracy and ACU are reported in the Table 6.

## Citation

```bibtex
@misc{grossi2022mixed,
  title={Mixed Quantum-Classical Method For Fraud Detection with Quantum Feature Selection},
  author={Michele Grossi et al.},
  year={2022},
  note={arXiv:2208.07963}
}
```

- arXiv: 2208.07963