RWR-GAE

Code for the paper "Random Walk Regularized Graph Auto Encoder"

The base code is a PyTorch implementation of the Variational Graph Auto-Encoder model described in the paper: T. N. Kipf, M. Welling, Variational Graph Auto-Encoders, NIPS Workshop on Bayesian Deep Learning (2016)

The code in this repo is based on or refers to https://github.com/tkipf/gae, https://github.com/tkipf/pygcn and https://github.com/vmasrani/gae_in_pytorch.

Requirements

• Python 3
• PyTorch 0.4

To train a model run the following command

cd gae
python train.py --model="gcn_ae" --dataset-str="cora" --dw=1 --epochs=200 --walk-length=5 --window-size=3 --number-walks=5 --lr_dw=0.01

• Supported models are "gcn_vae" and "gcn_ae"
• Supported datasets are "cora" and "citeseer"
• dw, whether to use regularization or not (0: no regularization, 1: yes)
• if dw = 0, then all the remaining params are useless
• refer to gae/train.py for other program arguments

Results on CORA test set

Link Prediction results:

Model	ROC	AP
GAE	0.91	0.92
VGAE	0.914	0.926
GAE (our impl)	0.91430	0.92585
VGAE (our impl)	0.921715	0.927751
ARGE	0.924	0.932
ARVGE	0.924	0.926
DW-GAE	0.924	0.918
DW-VGAE	0.926	0.918

Clustering results:

Model	Acc	NMI	F1	Precision	ARI
GAE	0.596	0.429	0.595	0.596	0.347
VGAE	0.609	0.436	0.609	0.609	0.346
GAE (our impl)	0.526	0.42	0.508	0.530	0.308
VGAE (our impl)	0.590	0.445	0.563	0.578	0.351
ARGE	0.640	0.449	0.619	0.646	0.352
ARVGE	0.638	0.450	0.627	0.624	0.374
DW-GAE	0.669	0.464	0.618	0.629	0.417
DW-VGAE	0.685	0.455	0.668	0.685	0.417

Results on Citeseer test set

Link Prediction results:

Model	ROC	AP
GAE	0.895	0.899
VGAE	0.908	0.92
ARGE	0.932	0.919
ARVGE	0.924	0.93
DW-GAE	0.921	0.915
DW-VGAE	0.913	0.908

Clustering results:

Model	Acc	NMI	F1	Precision	ARI
GAE	0.408	0.176	0.372	0.418	0.124
VGAE	0.344	0.156	0.308	0.349	0.093
ARGE	0.573	0.350	0.546	0.573	0.341
ARVGE	0.544	0.261	0.529	0.549	0.245
DW-GAE	0.616	0.344	0.585	0.605	0.343
DW-VGAE	0.613	0.338	0.582	0.595	0.336

Results on Pubmed test set

Link Prediction results:

Model	ROC	AP
GAE	0.964	0.965
VGAE	0.944	0.947
ARGE	0.968	0.971
ARVGE	0.965	0.968
DW-GAE	0.947	0.947
DW-VGAE	0.953	0.952

Clustering results:

Model	Acc	NMI	F1	Precision	ARI
GAE	0.697	0.33	0.69	0.72	0.322
VGAE	0.608	0.219	0.612	0.613	0.195
DW-GAE	0.726	0.355	0.714	0.729	0.37
DW-VGAE	0.736	0.346	0.725	0.736	0.381

Runs in 2-3 mins for cora dataset on cpu. The code currently doesn't support GPU.