In this example, we implement HGCN (Ines Chami, et al., Hyperbolic Graph Convolutional Neural Networks, NeurIPS'19) by curvlearn over the OpenFlights airport dataset.
3,188 nodes are kept in the dataset. The adjacency matrix is recorded in adj.pkl, and the numeric features are collected in features.pkl.
The configurations of training are listed in config.py, leading to the following performance.
| Manifold | AUC |
|---|---|
| Euclidean | 93.68 |
| PoincareBall | 94.51 |
| Stereographic | 95.13 |
The entry of the training is train.py. Launch the training by
python examples/hgcn/train.py
and have fun!
The code has been tested under the following environment settings:
Hardware:
Tesla P100 - 16GB (Actual consumption: 1.4GB)
Intel Xeon E5-2682 v4 @ 2.50GHz
Python dependencies:
tensorflow-gpu==1.15.0
numpy==1.16.5