NeurIPS 2020

Stochastic Deep Gaussian Processes over Graphs

Meta Review

This paper shows good empirical results on graph learning. While some reviewers would have preferred to see a larger contribution on the technical part (see e.g. useful suggestions by R3), it was deemed that the demonstration of a DGP working convincingly on graphs was significant enough without complicated technical derivations such as those often seen in DGP papers (which focus on new inference methods rather than new application areas). Importantly, the motivation and practical merit of combining DGPs with graphs has been made clear: it allows to perform deep learning on graphs (a recent, successful trend in the field) with added and well-analyzed qualitative advantages, namely obtained variance and ARD relative importance.