Variational Graph Auto-Encoders

Variational Graph Auto-Encoders (VGAE) bring the variational auto-encoder framework to graph-structured data for unsupervised learning. The model introduces latent variables to learn interpretable latent representations of undirected graphs, using a graph convolutional network as the encoder and a simple inner product decoder to reconstruct graph structure.

On a link prediction task in citation networks, VGAE achieves competitive results. A key advantage over most existing unsupervised graph and link-prediction models is that it can naturally incorporate node features, which significantly improves predictive performance on several benchmark datasets. By combining latent-variable modeling with node features, VGAE offered a simple yet effective framework for unsupervised learning and link prediction on graphs.