Corporate Fraud Detection Eval

---
name: corporate-fraud-detection-eval
description: Evaluates a model's ability to detect corporate fraud using financial graphs. It specifically probes robustness to information overload from noisy support nodes (e.g., directors) and label noise caused by delayed fraud detection. Use when the user wants to benchmark on MBM, SME, GEM, or asks about evaluating this task. Reports AUC.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2502.19305
  bibtex_key: wang2025corporatefraud
  confidence: high
---

# corporate-fraud-detection-eval

> Corporate Fraud Detection in Rich-yet-Noisy Financial Graph — Wang et al. (2025) (arXiv:2502.19305, 2025)

## What this evaluates

Evaluates a model's ability to detect corporate fraud using financial graphs. It specifically probes robustness to information overload from noisy support nodes (e.g., directors) and label noise caused by delayed fraud detection.

## Datasets

- **MBM** — total ?; splits: train (-1), valid (-1), test (-1)
- **SME** — total ?; splits: train (-1), valid (-1), test (-1)
- **GEM** — total ?; splits: train (-1), valid (-1), test (-1)

## Metrics

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

## Input / output format

**Input**: Graph-structured financial data containing company nodes and support nodes (e.g., directors, related parties). Node features are financial attributes (min-max normalized, missing values imputed with mean).

**Output**: Fraud probability or decision score for each company node.

## Scoring recipe

```python
def compute_auc(y_true, y_pred):
    fpr, tpr, _ = roc_curve(y_true, y_pred)
    return auc(fpr, tpr)
```

## Common pitfalls

- Test set is strictly restricted to non-fraud companies labeled at least 8 years ago to prevent hidden fraud bias.
- Standard GNNs suffer from information overload when processing the full graph with many support nodes; models must distill support node info via knowledge embeddings.
- Label noise is asymmetric and neighborhood-dependent due to delayed fraud detection, requiring specialized robust training.

## Evidence (verbatim from paper)

> We exploit AUC as the evaluation metric following the common practice in fraud detection studies [[18], [17]]. We randomly split each of our datasets into train/valid/test following the proportion of 6:2:2 with the additional constraint that the test set contains no hidden fraud.

## Citation

```bibtex
@misc{wang2025corporatefraud,
  title={Corporate Fraud Detection in Rich-yet-Noisy Financial Graph},
  author={Wang et al. (2025)},
  year={2025},
  note={arXiv:2502.19305}
}
```

- arXiv: 2502.19305