• Python 3.6
• Tensorflow 2
• Pygsp
• Scipy
• Scikit-learn

Datasets available:

['ENZYMES', 'PROTEINS', 'PROTEINS_full', 'MUTAG',
 'PTC_FM', 'NCI1', 'PTC_FR', 'DD',
 'DLA', 'IMDB-BINARY', 'MNIST_test']

Note: 'ENZYMES' contains 32 unconnected graphs.

Usage:

python3 hierarchical_skip_gram.py --task=PROTEINS --loss_type=negative_sampling --batch_size=4 --max_depth=5 --num_features=512 --num_epochs=10 --gnn_type=krylov-5 --num_tests=10 --device=0

It will:

• train over PROTEINS dataset
• train the model over 10 epochs with batches of size 4
• the dimension of the embedding will be equal to 512
• it will use GPU 0
• the graph will be coarsened to maximum depth 5
• the vocabulary will be extracted with Truncated Krylov layer, of depth 5
• using negative sampling loss
• repeat this procedure 10 times in order to generate statistics to assess the average performance

It will create a folder 'PROTEINS_weights' with:

• the weights of the node embedder, with h5 format
• the weights of the graph embeddings, with h5 format
• the weights of the graph embedding, in CSV format

To test the quality of the embeddings go to https://projector.tensorflow.org/
You need to upload 'PROTEINS_weights/graph_embeddings.csv' and 'PROTEINS/PROTEINS_graph_labels.txt'

To test the quality of the embeddings just use the SVM:

python3 baselines.py --task=PROTEINS

It will print the accuracy.

This algorithm aims to produce graph embeddings with the use convolutional networks to extract vocabulary.
Then, using Skip Gram with graph embeddings as context embeddings, we generate embeddings for the graph that contain information on the vocabulary inside.
The nodes are pooled with Loukas method.

The loukas_coarsening folder is copied from https://github.com/loukasa/graph-coarsening with only slight modifications.
My work is redistributed under the same license.