Generative Model for Node Generation

We present a generative model applied to graph-structured data for node generation by incorporating the graph convolutional architecture and semi-supervised learning with variational auto-encoder. This idea is motivated by successful applications of deep generative models for images and speeches. However, when applied to graph-structured data, especially social network data, existing deep generative models usually don't work: these models can not learn underlying distributions of social network data effectively. In order to address this problem, we construct a deep generative model, using architectures and techniques that prove to be effective for modelling network data in practice. Experimental results show that our model can successfully learn the underlying distribution from the social network dataset, and generate reasonable nodes, which can be altered by varying latent variables. This provides us a way to study social network data in the same way we study image data.