SHGLNN

Self-supervised Heterogeneous Hypergraph Learning with Context-aware Pooling for Graph-level Classification

SHGLNN is an algorithm for graph embeddings, harnessing the power of hypergraphs for intra and inter-graph contexts. This repository provides the code implementation of the algorithm as described in the paper.

Algorithm Overview

The algorithm receives a set of graphs and processes them in batches through a specified number of epochs. It leverages intra-graph hyperedges, inter-graph hyperedges, type-specific node and hyperedge attentions, and contrastive loss to generate embeddings for the input graphs.

See Algorithm 1 in our paper for a detailed step-by-step explanation.

Code Structure

Core Files:

• kMatrix.py: Responsible for generating k-specific matrices corresponding to intra-graph hyperedges.
• sMatrix.py: Generates matrix pertaining to inter-graph hyperedges.
• hypergraphGen.py: Constructs hypergraph using the intra and inter-graph hyperedges.
• dataset_loader.py: Helper functions to load and preprocess the datasets.
• layers.py: Contains various neural network layers utilized in the model - node convolution, hyperedge convolution, context-aware graph-level pooling
• model.py: Defines the main SHGLNN neural network model.
• train.py: Orchestrates the training process of the algorithm, including the calculation of contrastive loss and back propagation.

Prerequisites

• Python 3.8 or above
• PyTorch Geometric 2.4.0
• NumPy 1.26.0
• NetworkX 3.1
• HypernetX 2.0.5

Datasets

We use the TUDatasets mentioned in the paper from PyTorch Geometric TUDatasets (https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.datasets.TUDataset.html#torch_geometric.datasets.TUDataset)

Usage

1. Clone the repository git clone https://github.com/YOUR_USERNAME/SHGLNN.git cd SHGLNN
2. Install the required packages/libraries: pip install numpy networkx hypernetx
3. Generate the matrices K and S for intra- and inter-graph hyperedges
   • Note that for all datasets, multiple K matrices would be generated as we go through different values of K. However, we will generate a single S matrix for each dataset.
4. Generate hypergraph using hypergraphGen.py
5. Develop layers using layers.py
6. Run the model.py to develop the model
7. Then, train the model using train.py
   • Note that for datasets, we used the built-in TU dataset from PyTorch Geometric

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 [email protected] or [email protected].

License

MIT

Citation

If you use this work, please cite the following:

@inproceedings{hayatSHGLNN2023, 
    author = {Hayat, Malik Khizar, Xue, Shan and Yang, Jian},
    title = {Self-supervised Heterogeneous Hypergraph Learning with Context-aware Pooling for Graph-level Classification},
    booktitle = {ICDM},
    year = {2023},
    pages=---,
    doi={---},
}