NeurIPS 2020

The Potts-Ising model for discrete multivariate data

Meta Review

Four expert reviewers placed this paper near the borderline. All reviewers recognized that the basic idea to define a model that strikes a middle ground between the Potts model and Ising model was simple and elegant, and that the results were technically correct. R1 and R4 especially liked the simplicity and elegance, and all reviewers were generally very positive about the idea being motivated by a real-world application. Overall, the reviewers were not surprised by the technical results given the similarity to existing models, but acknowledged their novelty. There were a number of questions/concerns/suggestions, especially relating to baselines, evaluation metic, and diversity of data sets in the experiment as well as dense and non-standard notation. The authors provided new experimental results, proposed a new evaluation metric, and compared to the Copula Multinomial baseline suggested by R2 in the rebuttal; these improvements convinced two reviewers to raise their scores. While some lesser questions/concerns remain (e.g. R3 is surprised/skeptical that an apparently impoverished model will work in a broad range of settings), all reviewers recognized positive aspects of the paper and two were enthusiastic. The authors are strongly encouraged to take the very thoughtful reviewer comments into account while revising the paper, especially relating to notation.