After reading the paper, reviews, and rebuttal, and after discussing with the reviewers, I solicited a fourth reviewer to weigh in. The consensus is that the core contribution is an adaptation of hyperbolic neural networks to handle structured (graph) inputs, with some subtleties in terms of how the constraints of hyperbolic space are maintained. The paper is quite similar to recent ICLR '19 work on hyperbolic attention networks (see R4 comments); this is the related work R3 was referring to, but unfortunately did not explicitly reference in his original review. Still, it is quite surprising that this paper is not referenced in the original submission since it has been on arXiv and OpenReview for many months and would have turned up in a cursory Google Scholar search. The authors must make clear the connections to this work in the final proceedings. R4 does identify some points of departure from the ICLR paper, and suggests that there is enough novelty in this submission to warrant publication in NeurIPS.