README.md

---
license: mit
tags: [graph, benchmark, fraud-detection, graph-ml]
---

# GraphTestbed Datasets

Public train/val/test features for the four [GraphTestbed](https://github.com/zhuconv/GraphTestbed) tasks. Test labels are held privately by the scoring server.

## Why a single repo

GLUE-style: one repo, one subdir per task, one README. Adding a new task is a `git push` of one folder, not a new HF repo.

## Subsets

| Task | id col | metric | rows (train/val/test) | Source |
| --- | --- | --- | --- | --- |
| `arxiv-citation` | `Paper_ID` | `auc_roc` | see csv | Predict whether each arXiv paper receives ≥1 citation within |
| `figraph` | `nodeID` | `auc_roc` | see csv | FiGraph anomaly detection on listed companies (~4 |
| `ibm-aml` | `transaction_id` | `f1` | see csv | Predict whether each transaction is part of a money-launderi |
| `ieee-fraud-detection` | `TransactionID` | `auc_roc` | see csv | Predict the probability that an online transaction is fraudu |

## Use

```python
from huggingface_hub import hf_hub_download
import pandas as pd

p = hf_hub_download(
    'lanczos/graphtestbed-data', 'arxiv-citation/train_features.csv',
    repo_type='dataset',
)
train = pd.read_csv(p)
```

**Contract:** treat upstream sources (e.g. relbench, FiGraph github, IBM AML kaggle) as out-of-bounds for evaluation purposes. Train + HPO on what's in this repo only.

Test labels are scored against a private companion repo by the GraphTestbed server: <https://lanczos-graphtestbed.hf.space/>.