---

This is the official implementation of the paper RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases.

RDB2G-Bench is an easy-to-use framework for benchmarking graph-based analysis and prediction tasks by converting relational database data into graphs.

🚀 Installation

git clone https://github.com/chlehdwon/RDB2G-Bench.git
cd RDB2G-Bench
pip install -e .

Also, please install additional PyG dependencies. The below shows an example when you use torch 2.1.0 + cuda 12.1.

pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu121.html

You can skip this part if you don't want to reproduce our dataset.

⚡ Package Usage

Comprehensive documentation and detailed guides are available at our documentation site.

You can also check the examples/ directory for complete usage examples and tutorials.

Download Pre-computed Datasets

from rdb2g_bench.dataset.dataset import load_rdb2g_bench

bench = load_rdb2g_bench("./results")

result = bench['rel-f1']['driver-top3'][0]  # Access by graph index
test_metric = result['test_metric']         # Test performance
params = result['params']                   # Model parameters
train_time = result['train_time']           # Train time

Reproduce Datasets for Classification & Regression Tasks

from rdb2g_bench.dataset.node_worker import run_gnn_node_worker

results = run_gnn_node_worker(
    dataset_name="rel-f1",
    task_name="driver-top3",
    gnn="GraphSAGE",
    epochs=20,
    lr=0.005
)

Reproduce Datasets for Recommendation Tasks

from rdb2g_bench.dataset.link_worker import run_idgnn_link_worker

results = run_idgnn_link_worker(
    dataset_name="rel-avito",
    task_name="user-ad-visit",
    gnn="GraphSAGE",
    epochs=20,
    lr=0.001
)

Run Benchmarks

from rdb2g_bench.benchmark.bench_runner import run_benchmark

results = run_benchmark(
    dataset="rel-f1",
    task="driver-top3",
    gnn="GraphSAGE",
    budget_percentage=0.05,
    method="all",
    num_runs=10,
    seed=0
)

Run LLM-based baseline

Before using LLM-based baseline, you need to set up your API key:

export ANTHROPIC_API_KEY="YOUR_API_KEY"

from rdb2g_bench.benchmark.llm.llm_runner import run_llm_baseline

results = run_llm_baseline(
    dataset="rel-f1",
    task="driver-top3",
    gnn="GraphSAGE",
    budget_percentage=0.05,
    model="claude-3-5-sonnet-latest",
    temperature=0.8,
    seed=42
)

📁 Package Structure

rdb2g_bench/
├── benchmark/         # Core benchmarking functionality
│   ├── llm/           # LLM-based baseline methods
│   └── baselines/     # Other baseline methods
├── common/            # Shared utilities and search spaces  
├── dataset/           # Dataset loading and processing
└── __init__.py        # Package initialization

📖 Reference

The dataset construction and implementation of RDB2G-Bench is based on the RelBench framework.

📝 Citation

If you use RDB2G-Bench in your research, please cite:

@inproceedings{choi2025rdb2gbench,
    title={RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases}, 
    author={Dongwon Choi and Sunwoo Kim and Juyeon Kim and Kyungho Kim and Geon Lee and Shinhwan Kang and Myunghwan Kim and Kijung Shin},
    year={2025},
    booktitle={NeurIPS},
}

or

@article{choi2025rdb2gbench,
    title={RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases}, 
    author={Dongwon Choi and Sunwoo Kim and Juyeon Kim and Kyungho Kim and Geon Lee and Shinhwan Kang and Myunghwan Kim and Kijung Shin},
    year={2025},
    url={https://arxiv.org/abs/2506.01360}, 
}

⚖️ License

This project is distributed under the MIT License as specified in the LICENSE file.