THUDM / GraphSGAN

Implementation of "GraphSGAN", a GAN-based semi-supervised learning algorithm for graph data.

GraphSGAN

Implementation of "GraphSGAN", a GAN-based semi-supervised learning algorithm for graph data.

Paper: Semi-supervised Learning on Graphs with Generative Adversarial Nets

Preparation

unzip cora.dataset.zip

The codes are written under Python==2.7 and pytorch~=3. If you want to run it in other environments, minor changes might be needed.

Run

python GraphSGAN.py --cuda

will run a example, banchmark cora task.

The programme takes FeatureGraphDataset as input and cora.dataset is built from FeatureGraphDataset.py. You can build your own FeatureGraphDataset.

Early stop and tuned hyperparameters are not included in this minimal release. You can determine them based on your own validation set.

rm -rf models logfile

can delete the saved models and logfile to retrain.

You can visualize the infomation in logfile by using tensorboard.

The expected accuracy of example should be above 0.83.

How to run GraphSGAN on other datasets?

1. Build FeatureGraphDataset for new dataset. This class takes init parameters as below:

    features: numpy ndarray, [[f1, f2, ...], [f1, f2, ...]]
    label: numpy ndarray, [0, 1, 2, 0, ...]
    adj: dict of (int, list of int), {[1,2],[0,3],...}

2. Load embeddings for dataset.

It is recommended to use read_embeddings method to read embeddings from file.

The first line of embeddings file are two integers: n and dim.

In the next n lines, each line contains dim + 1 integers. The first is the No. of the node and the rest are embeddings.

Example:

  3 2
  0 0.123 0.233
  1 0.720 -0.121
  2 0.778 -0.921
  3 0.161 -0.775

3. Setting splits

call setting(label_num_per_class, test_num)

4. Replace dataset in GraphSGAN.py with built new dataset.

Performance on some other datasets

In the paper of SCAN_DIS, the performance of GraphSGAN on Pubmed, Flickr and BlogCatalog are tested:

	Pubmed			Flickr			BlogCatalog
Marco-F1	Micro-F1	Acc	Marco-F1	Micro-F1	Acc	Marco-F1	Micro-F1	Acc
.839	.842	.841	.697	.715	.702	.698	.703	.719

Although not responsible for the results, we think it is really worth reference.