A Comparative Study of Representation Learning Techniques for Dynamic Networks

Representation Learning in dynamic networks has gained increasingly more attention due to its promising applicability. In the literature, we can find two popular approaches that have been adapted to dynamic networks: random-walk based techniques and graph-autoencoders. Despite the popularity, no work has compared them in well-know datasets. We fill this gap by using two link prediction settings that evaluate the techniques. We find standard node2vec, a random-walk method, outperforms the graph-autoencoders.