DyH₂GNet

Self-supervised Time-aware Heterogeneous Hypergraph Learning for Dynamic Graph-level Classification

DyH₂GNet is a novel algorithm for learning high-order temporal patterns in dynamic graphs. It constructs heterogeneous hypergraphs by combining structural and semantic interactions and employs a self-supervised learning (SSL) mechanism to enhance graph-level classification. This repository provides the code implementation as described in the paper.

---

Framework Overview

The framework processes dynamic graphs through consecutive snapshots. It constructs heterogeneous hypergraphs using $k$-hop neighborhoods for structural interactions and node correlation for behavioral interactions. DyH₂GNet employs time-aware graph augmentations, intra/inter-snapshot attentions, and contrastive loss to generate embeddings that capture temporal dynamics effectively.

See Algorithm 1 in our paper for a step-by-step explanation of the process.

Code Structure

Core Files:

• dataset_loader.py: Loads and preprocesses datasets.
• kMatrix.py: Generates $k$-specific matrices for structural interactions.
• sMatrix.py: Generates correlation-based matrices for behavioral interactions.
• hypergraphGen.py: Constructs the heterogeneous hypergraph from structural and behavioral matrices.
• models/: Contains modules for:
  • Intra- and inter-snapshot node and hyperedge attention mechanisms.
  • Temporal attention, contrastive loss, and pooling layers.
• train_dyh2gnet.py: Implements the training process for DyH₂GNet.
• analysis/: Evaluation scripts for:
  • Graph Similarity Ranking
  • Anomaly Detection
  • Trend Analysis
  • Ablation Analysis

---

Prerequisites

• Python 3.8 or above
• PyTorch Geometric 2.4.0
• NumPy 1.26.0
• SciPy 1.11.0
• NetworkX 3.1
• Matplotlib 3.7.2
• tqdm 4.65.0

---

Datasets

This implementation supports dynamic graph datasets. Example datasets include:

• Enron (communication network)
• Facebook (social interactions)
• Reddit GOT (user discussions)
• Formula1 (sports events)

Place dataset files (e.g., .pkl format) in the data/ folder for training and evaluation.

---

Usage

1. Clone the repository: git clone https://github.com/YOUR_USERNAME/DyH2GNet.git, then cd DyH2GNet
2. pip install -r requirements.txt
3. Place the dataset files in the data/ folder
4. Train the model python train_dyh2gnet.py
5. Evaluate the model:

• python analysis/graph_similarity_ranking.py
• python analysis/anomaly_detection.py
• python analysis/trend_analysis.py
• python analysis/ablation_analysis.py

Contribution

Pull requests are welcome. For major changes or any issues, please open an issue first for discussion. For further inquiries, contact the first author at malik.khizar@hdr.mq.edu.au or khizerhayat92@gmail.com.

License

MIT

Citation

If you use this work, please cite the following:

@inproceedings{Hayat2025DyH2GNet,
  author    = {Malik Khizar Hayat and Shan Xue and Jia Wu and Bilal Khan and Jian Yang},
  title     = {Self-supervised Time-aware Heterogeneous Hypergraph Learning for Dynamic Graph-level Classification},
  booktitle = {Proceedings of the ACM International Conference on Web Search and Data Mining (WSDM)},
  year      = {2025}
}