junhongmit/H2GB

When Heterophily Meets Heterogeneity:
Challenges and a New Large-Scale Graph Benchmark

Junhong Lin¹, Xiaojie Guo², Shuaicheng Zhang³, Yada Zhu², Dawei Zhou³, Julian Shun¹
¹ MIT CSAIL, ² IBM Research, ³ Virginia Tech

This repository hosts a subset of datasets from H2GB, a large-scale benchmark suite designed to evaluate graph learning models on heterophilic and heterogeneous graphs. These graphs naturally arise in real-world applications such as fraud detection, malware detection, academic venue classification, and social network analysis.

The full H2GB benchmark includes 9 datasets across 5 domains (academia, finance, e-commerce, cybersecurity, and social science). This Hugging Face dataset repository provides the following packaged datasets:

• IEEE-CIS-G: Financial transaction fraud detection dataset from IEEE-CIS Kaggle competition
• PDNS: Cybersecurity graph constructed from passive DNS logs
• OAG dataset: Academic paper graphs from the Open Academic Graph (OAG), with domain-specific subsets:
  • oag-cs (computer science)
  • oag-eng (engineering)
  • oag-eng (chemistry)

Other datasets in H2GB (e.g., ogbn-mag, mag-year, H-Pokec, RCDD) are publicly hosted and are automatically downloaded and processed by the H2GB framework during runtime.

---

💾 Dataset Structure

All datasets are formatted for compatibility with PyTorch Geometric using the HeteroData object format.

Included

• ieee_cis_dataset.zip: Raw files representing a heterogeneous bipartite transaction graph with 12 node types and 22 edge types. Nodes represent credit card transactions and associated metadata. Node features are 4823-dimensional numeric vectors.
• pdns_dataset.zip: Cybersecurity graph with domain and IP entities. Includes structural metadata and 10-dimensional node features.
• OAG_CS_20190919.pt, OAG_Engineering.pt and OAG_Chemistry.pt: The .pt files representing academic paper graphs, each as a dictionary object with preprocessed citation relationships and title embeddings.

---

📥 How to Load

To load a specific dataset from this repository using Hugging Face datasets:

from huggingface_hub import hf_hub_download
import zipfile

output_directory = "./" # Your output directory

# Load the IEEE-CIS financial dataset
path = hf_hub_download(repo_id="junhongmit/H2GB", filename="IEEE-CIS.zip", repo_type="dataset")

with zipfile.ZipFile(path, 'r') as zip_ref:
    zip_ref.extractall(output_directory)

📚 Citation

If you use this dataset, please consider citing the following paper (currently on arXiv, will be in KDD procedding soon):

@article{lin2024heterophily,
  title={When Heterophily Meets Heterogeneity: New Graph Benchmarks and Effective Methods},
  author={Lin, Junhong and Guo, Xiaojie and Zhang, Shuaicheng and Zhou, Dawei and Zhu, Yada and Shun, Julian},
  journal={arXiv preprint arXiv:2407.10916},
  year={2024}
}