Fraud Dataset Benchmark Eval

---
name: fraud-dataset-benchmark-eval
description: This benchmark evaluates the robustness of fraud detection models to label noise in training data. It measures how effectively various noise-removal techniques preserve predictive performance when tested on clean, unseen data. The protocol specifically probes a model's ability to mitigate artificially injected label corruption across multiple real-world fraud datasets. Use when the user wants to benchmark on Fraud Dataset Benchmark (FDB), or asks about evaluating this task. Reports ROC-AUC.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2208.14417
  bibtex_key: [REDACTED]
  confidence: high
---

# fraud-dataset-benchmark-eval

> Fraud Dataset Benchmark and Applications — Grover et al. (2022) (arXiv:2208.14417, 2022)

## What this evaluates

This benchmark evaluates the robustness of fraud detection models to label noise in training data. It measures how effectively various noise-removal techniques preserve predictive performance when tested on clean, unseen data. The protocol specifically probes a model's ability to mitigate artificially injected label corruption across multiple real-world fraud datasets.

## Datasets

- **Fraud Dataset Benchmark (FDB)** — total ?; splits: train (-1), test (-1); repo https://github.com/amazon-research/fraud-dataset-benchmark

## Metrics

- `ROC-AUC` **(primary)** — range: [0, 1]
  - Area under the Receiver Operating Characteristic curve, measuring the trade-off between true positive rate and false positive rate across classification thresholds.

## Input / output format

**Input**: Tabular feature vectors with potentially noisy training labels; clean ground-truth labels are reserved for the test set.

**Output**: Predicted probabilities for the positive (fraud) class, or binary fraud/non-fraud predictions.

## Scoring recipe

```python
def evaluate(predictions, gold):
    from sklearn.metrics import roc_auc_score
    return roc_auc_score(gold, predictions)
```

## Common pitfalls

- Noise is artificially injected into training labels at fixed rates (10–20%) rather than reflecting natural label noise distributions.
- Evaluation is performed exclusively on clean test data to isolate the effect of training-time noise removal, not to test robustness to noisy test labels.
- Some datasets (e.g., fraudefcom) are highly imbalanced and difficult; without feature engineering, models may perform near random chance regardless of cleaning method.

## Evidence (verbatim from paper)

> First, in Figure 3 we show the ROC-AUC attained on clean test data by a model trained on data cleaned by the various approaches under consideration. Along the x-axis is the amount of noise added to the training data before cleaning, and the y-axis shows the resulting AUC.

## Citation

```bibtex
@misc{grover2022frauddatasetbenchmark,
  title={Fraud Dataset Benchmark and Applications},
  author={Grover et al. (2022)},
  year={2022},
  note={arXiv:2208.14417}
}
```

- arXiv: 2208.14417