Reviews: Variational Graph Recurrent Neural Networks

This paper studies a Graph RNN model for dynamic graphs. Cardinalities of nodes and edges can be time-varying. Especially the proposed VGRNN is made for highly variable graph sequences. The hidden state h_t, which is tracked via RNN function, governs the prior of latent variables and the sampled latent variable controls the generation of time-varying adjacency matrices. Such hierarchical modeling allows the proposed VGNN to fit to highly time-variable graph sequences.