NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2695
Title:Spectral Modification of Graphs for Improved Spectral Clustering

This paper presents an interesting approach of finding cut-similar graphs that have more desirable spectral properties if one to were, e.g., run spectral clustering. All of the reviewers found some very positive aspects of the paper theoretically. On the other hand, the experiments are considerably underwhelming and the authors did not make much of an effort to address this with the author feedback. Since there is a really nice contribution here, I am recommending this paper for acceptance. However, I would really like to encourage the authors to add more meaningful numerical experiments for a camera version. The current experiments in Section 4.1 can be removed and replaced with some of the experiments suggested by, e.g., Reviewer 3 on some moderately sized (maybe ~100,000 edges), more heterogeneous graphs.