Abstract

Representation learning of static and more recently dynamically changing graph-structured data has gained noticeable attention, especially in the unsupervised case. The majority of prior research work has focused extensively on modelling the evolutionary dynamics of individual nodes with applications, such as link prediction and node classification. However, there is a lack of methods for unveiling more global patterns of dynamic graph evolution. Thus, in this work, we present a generative model which is able to learn evolving node and community representations. We can use these dynamic representations to investigate the community transition patterns for individual nodes, or to study macro-trends, such as the evolution of graph communities over time. The proposed co-evolutionary model is trained via variational inference by optimizing the ELBO of the observed edges in the dynamic graph. We demonstrate competitive or superior performance of our method against state-of-the-art baselines on dynamic link prediction. We then analyze the evolution of interpretable communities in the DBLP bibliography, and also the community transition probabilities of individual authors to showcase the capabilities of our co-evolutionary model as a tool to extract dynamic graph insights.