NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:5521
Title:Same-Cluster Querying for Overlapping Clusters

Reviewer 1


		
This is a well-written theory paper about a novel problem that has not been addressed in this form in earlier literature. The main novelty is the combination of pairwise queries and overlapping clusters. The results are solid. A summary (table, etc) of the main complexity results for different oracle types would make a nice addition. About related work: The authors might also be interested in an earlier HCOMP 2015 paper by Zou et al where a very similar problem using relative comparison queries is considered. Also, Bonchi et al (2013) discuss ideas to combine correlation clustering with overlapping clusters that may also be of interest to the authors. (Note that I am *not* asking to cite these papers, but merely pointing them out as potentially interesting related work.) References: Francesco Bonchi, Aristides Gionis, Antti Ukkonen: Overlapping correlation clustering. Knowl. Inf. Syst. 35(1): 1-32 (2013) James Y. Zou, Kamalika Chaudhuri, Adam Tauman Kalai: Crowdsourcing Feature Discovery via Adaptively Chosen Comparisons. HCOMP 2015: 198- _______________ Update after rebuttal: Thank you for the response! I have nothing to add at this point.

Reviewer 2


		
The authors propose the problem of clustering with overlapping under the semi-supervised framework, or more specifically, using same cluster queries. Flat clustering with same cluster queries has received much attention recently. The authors take a step forward to the case of clustering with overlapping, which is more general comparing to the original problem. I think the importance of this new problem is fairly stated in the paper. The related works section is well-written which gives a clear idea how this work is different from prior works. The main idea of all the algorithms in this paper is to find a (possibly small) set of representatives first using all possible pair-wise queries from a larger set drawn uniformly at random. Then recover the memberships for the rest elements depending on the result of the first stage. Although the similar idea has appeared in prior works (i.e. [MS17]), which is also cited by the authors, applying it for the case of overlapping clusters with modification for numbers of settings is still novel. The theoretical analysis for this paper seems to be correct, although I did not read all the details in the supplement. The first contribution of this paper is providing the condition of uniqueness of optimal solution in various settings. I think this is an important step for the follow up works tackling this problem. Various settings for this problem are discussed in this paper. Both upper bounds along with algorithms and lower bounds on query complexity are provided, which is another significant contribution. However, the authors do not make comparison for their upper and lower bounds. It would be more clearly for readers if some remarks of this comparison can be made. Moreover, the computational complexity of these algorithms is not directly stated. The computational complexity is also a critical attribute when we want to judge algorithms. Finally, although the authors give results for both worst-case and model-based, they make neither comparison nor discussion on the query complexity. Intuitively the query complexity for the worst-case scenario should be much higher than the model-based case. However, the results do not seem to match this intuition if \alpha is some constant. The authors should have some discussion on the scale of \alpha and compare the query complexity of worst-case and model-based case. For the experiment, the authors show that on real-world data, their algorithms require much less queries then their upper bound. Nevertheless, no other method is compared. For example, the authors mention in the introduction that their algorithm should work better then naively apply the low-rank matrix completion, but no experiment support this statement. Also, from the synthetic data in the supplement, there seems to be a huge gap for the upper and lower bound. Since the authors take a log-scale in y-axis it is hard for me to tell whether the upper bound or lower bound is tight or not. It would be great if some discussion can be made for this point. At last, in the reproducibility response the authors claim that they provide the source codes, but I can not find it in the supplement. [MS17] A. Mazumdar and B. Saha. Clustering with noisy queries. In Advances in Neural Information Processing Systems (NIPS) 31, 2017. ==================================================== Update after rebuttal: I thank the authors for their response. It clarify all my concerns pretty well. Hence I slightly raise my score to 7 for this submission. Hope that all these nice explanation in the response can be seen in the final version of this paper.

Reviewer 3


		
The authors extend the study of clustering with queries to the scenarios where the clusters may overlap. There has been previous work regarding the query complexity of clustering using same-cluster queries. However, this is the first work that deals with situations that a point can belong to multiple clusters. Multiple types of oracles has been considered including + Direct response oracle where given two points, the oracle outputs the number of clusters that the two points belong to simultaneously (also called the similarity of the two points). + Quantized oracle where the output is just whether there exist a cluster where the two points belong to it simultaneously. Noisy versions of this oracle is also considered where the oracle's output is flipped with some fixed probability. + Dithered oracle where the oracle first adds a Gaussian noise to the similarity of the two given points and then outputs the quantized version. The authors propose methods that recover the true clustering (up to a permutation) with high probabilities. In order to avoid the identifiability issues, the authors consider assumptions under which knowing the similarity matrix is enough to reconstruct the clustering assignments (at least with high probability). I think the paper is well-written and the setting is very interesting. Adding short discussion regarding the tightness of the upper bounds for each scenario (and the possible room for improvement) would help as there are various variables involved in the bounds. Some questions. + In Theorem 5, the given query complexity does not depend on Delta but intuitively it should be higher for larger Delta. Can you clarify? + In Theorem 5, T is logarithmic in k and n...so in the special case where the clusters are non-overlapping we seem to be able to recover the clustering using less than nk queries...which should be impossible. What am I missing? + Theorem 2 has exponential dependence on Delta. Is that really necessary? + I did not fully follow the role of |S| in the experiments; is there an intuition why the theoretical value does not work as well? + In the last experiment (unquantized query responses) a particular heuristic has been used (e.g., selecting 5 movies, etc.). Why do we do that (and not follow the proposed algorithm)? Will the result be different if we did so? how much? == After reading other reviews and the authors' response I see no reason to change my score.