Hyperbolic Graph Neural Networks

Requirements

• Python 3.7
• PyTorch >= 1.1
• RDKit
• numpy
• networkx
• scikit-learn

A recipe about installing the requirements is provided in install.sh.

Data Preprocess

For the Ethereum dataset, go to data/ethereum and run download_ethereum.sh

For the node classification dataset, go to data/node and run download_node.sh

For QM8, QM9 and ZINC, go to data/qm8, data/qm9 and data/zinc, respectively and run python get_data.py

For the synthetic dataset, go to data/synthetic and run python generate_graphs.py

For TU Dortmund datasets, go to data/tu and run python data_preprocess.py {REDDIT-MULTI-12K, PROTEINS_full, ENZYMES, DD, COLLAB}

Run Experiments

The code can be run on SLURM and on multiple GPUs. To run on multi GPUs, use

python -m torch.distributed.launch --nproc_per_node=NUM_GPU main.py --task {qm8, qm9, zinc, ethereum, node_classification, synthetic, dd, enzymes, proteins, reddit, collab}

Inputs of Riemannian GNN

Here we introduce the inputs of Riemannian GNN:

• node_repr: representations of each node.
• adj_list: an adjacency list, of which each row i consists of the neighbor IDs of node i. adj_list is padded using 0 to make each row of the same size.
• weight: a weight list for weighted graphs, of which each row i contains the weights of neighbors. weight is padded using 0 to make each row of the same size.
• mask: the i-th row of mask is 0 if the node i is padded. Otherwise, the i-th row is 1.

Directory

• dataset: dataset files.
• gnn: Riemannian graph neural network implementation.
• hyperbolic_module: centroid-based classification and Poincaré distance.
• manifold: Poincaré, Lorentz and Euclidean manifolds.
• optimizer: Riemannian SGD and Riemannian AMSGrad.
• params: parameters for each task.
• task: task code.
• utils: utility modules and functions.

Hyperparameters

Some notable hyperparameters are listed here.

• lr: learning rate for Euclidean variables.
• lr_hyperbolic: learning rate for hyperbolic variables.
• optimizer: optimizer for Euclidean variables.
• hyper_optimizer: optimizer rate for hyperbolic variables.
• num_centroid: the number of centroids for centroid-based prediction.
• gnn_layer: the number of GNN layers.
• embed_size: the embedding size.
• apply_edge_type: a boolean value denotes multi-relational or single-relational.
• edge_type: the number of relations for multi-relational datasets.
• select_manifold: use the Euclidean, Poincaré or Lorentz manifold.
• activation: the activation function.

License

HGNN is licensed under Creative Commons-Non Commercial 4.0. See the LICENSE file for details.