Reviews: D-VAE: A Variational Autoencoder for Directed Acyclic Graphs

Originality: The main architectural pieces (VAE, message passing) have been proposed previously, but the specific focus on directed graphs using these components is new to my knowledge. Quality: The proposal was technically sound, but the proposal has undesirable properties which were not addressed by the authors, and lacks sufficient empirical evaluation: 1. Does the proposal easily (from a practical implementation perspective) allow for batching? Instead of N steps with an RNN, the proposal's decoder uses N*(1 2 ... N-1) steps. This may limit the proposal to small graphs (the authors have only evaluated on a fixed, small graph size).