Clustering with Non-adaptive Subset Queries

Recovering the underlying $k$-clustering of a set $U$ of $n$ points by asking pair-wise same-cluster queries has garnered significant interest in the past few years. Given a query $S \subset U$, $|S|=2$, the oracle returns"yes"if the points are in the same cluster and"no"otherwise. For adaptive algorithms, the query complexity is known to be $\Theta(nk)$, while non-adaptive algorithms are extremely limited: even for $k=3$, such algorithms require $\Omega(n^2)$ queries, matching the trivial upper bound. However, non-adaptivity is highly desirable since it allows queries to be asked in parallel. To break the quadratic barrier for non-adaptive queries, we study a natural generalization of this problem to subset queries for $|S|>2$, where the oracle returns the number of clusters intersecting $S$. Previous work obtained an $O(n)$ query adaptive algorithm, but the realm of non-adaptive algorithms remained completely unknown. In this paper, we give the first non-adaptive algorithms for clustering with subset queries. Our main result is a non-adaptive algorithm making $O(n \log k \cdot (\log k + \log\log n)^2)$ queries, improving to $O(n \log \log n)$ when $k$ is constant. In addition to non-adaptivity, we make other practical considerations, such as enforcing a bound, $s$, on the query size. We show $\Omega(\max(n^2/s^2,n))$ queries are necessary and obtain algorithms making $\smash{\widetilde{O}(n^2k/s^2)}$ queries for any $s \leq \sqrt{n}$ and $\smash{\widetilde{O}(n^2/s)}$ queries for any $s \leq n$. Finally, we obtain improved upper bounds when the clusters are roughly balanced, and when the algorithm is allowed two rounds of adaptivity.

Paper

References (48)

Scroll for more · 36 remaining

Similar papers

Peer review

Reviewer hP147/10 · confidence 2/52024-06-30

Summary

**[Setting]** This paper studies the problem of clustering $n$ items into $k$ clusters using an oracle that can tell how many ground-truth clusters are represented in any given subset $S$ of items. The goal is to develop non-adaptive algorithms (where all queries are chosen before the oracle answers anything) and study their sample complexity. **[Contributions]** 1. Randomized non-adaptive algorithms for constant k that make: 1. $O(n k \log \log n \log k)$ queries when subsets of any size can be queried 2. $O(n k \log n \log\log n)$ queries when $|S| = O(\sqrt{n})$ 2. Randomized non-adaptive algorithms for general k that make: 1. $O(n \log n \log k (n/s + \log s))$ queries of size at most $s$ 3. An $\Omega(\min(n^2 / s^2, n))$ lower bound for any non-adaptive algorithm that is allowed to query subsets of size at most $s$. The developed algorithms run in polynomial time. Additionally, the paper also presents results for the case when : 1. Clusters are roughly balanced (i.e., their sizes are within constant factors of each other). Here, the algorithm makes $O(n \log k)$ queries when $k = O(n/\log^3 n)$ and $O(n \log^2 k)$ queries for any $k$. 2. An adaptive round is allowed. This improves dependency on logarithmic factors. For these results, the details are only in the appendices.

Strengths

1. The paper is fairly comprehensive and presents query complexity results in a wide range of settings. The proposed algorithm are optimal up to logarithmic factors. 2. The algorithms are well motivated. I especially appreciate the high-level intuition because of the clarity it adds to the paper. 3. While I have some concerns about the problem setting (see Weaknesses), I believe the presented results are a very good first step towards filling an interesting research gap - non-adaptive algorithm for clustering with subset queries.

Weaknesses

1. My primary concern is with the feedback model. The oracle returns the number of clusters in the subset $S$, which implies that it knows the correct grouping in $S$. Why then would it only return the number of clusters in $S$ in practice and not the clusters themselves? Perhaps there are bandwidth constraints on response in some applications? But wouldn't one require way fewer queries if the oracle returns the clusters in $S$, which alleviates the bandwidth concerns, especially in the non-adaptive setting? 2. This is a relatively minor concern, but the paper defers a lot of details to the appendix without even including high-level ideas. I recommend including some details in the main paper, perhaps at the cost of a few technical details from Section 2. For example, 1. What changes to the algorithm allow the query complexity to improve for the roughly balanced case? 2. What is the "extra round of adaptivity"?

Questions

1. Please address point 1 under weaknesses. 2. It seems to me that both Algorithm 1 and 2 work for any value of k. Can the authors elaborate on what they mean by constant $k$ and general $k$? Do you have two algorithms because one of them has a better dependence on $k$ ($O(\log k)$ instead of $O(k \log k)$)? On a similar note, how should one go about choosing between Algorithm 1 and 2 in practice? 3. Your lower bound is based on the idea of converting subset queries to pairwise queries. Unfortunately, this hides the dependence on $k$. Are you aware of any lower bounds in the pairwise case that depend on $k$ as well?

Rating

7

Confidence

2

Soundness

4

Presentation

3

Contribution

4

Limitations

While it is true that all theorems have their assumptions listed, it would help the readers if the authors include a dedicated paragraph pointing out some of the research gaps that are left open by their work.

Reviewer WJzB7/10 · confidence 3/52024-07-09

Summary

The problem of cluster recovery via membership queries asks to assign each point in a dataset to its $k$ ground-truth cluster by using as few queries as possible to an oracle. A well-studied oracle is the same-cluster oracle that answers, given two points, whether the points are in the same or in different clusters. The complexity of adaptive algorithms for this problem is $\Theta(nk)$, but non-adaptive algorithms require $\Omega(n^2)$ for $k \geq 3$. The authors of this paper consider subset queries to surpass this barrier. Given a set of points $S$, a subset query returns the number of clusters the points in $S$ belong to. It is known that the complexity of an adaptive algorithm is $O(n)$. The authors consider non-adaptive algorithm and obtain the following results. For general subset queries, they obtain an algorithm that makes $O(n \log^2 n \log k)$ queries for unbalanced clusters, and $O(n \log k)$ queries for balanced clusters. For constant $k$, they obtain an algorithm that makes $O(n \log \log n)$ queries. If the allowed subset size of the queries is bounded by $s \leq \sqrt{n}$, their algorithms need an essentially optimal $\tilde{O}(n^2 / s^2)$ number of queries for constant $k$, and $\tilde{O}(n^2 / s)$ for arbitrary $k$. Finally, they show that 2-round adaptivity can improve on the log factors in the complexity. The algorithms follow two algorithmic approaches. For constant $k$, the authors observe that one can use subset queries to identify an isolated point representing its cluster in a subset and explore the cluster from this point, or one can use subset queries that contain two points from the same cluster to sample intra-cluster edges. The two strategies have a reverse complexity with respect to the cluster size, and a combination of them give the final complexity trade-off. For general $k$, the authors related subset queries to Boolean OR queries, and use these to recover clusters one by one.

Strengths

The paper provides a rather rich collection of results on foundational variants of the non-adaptive cluster recovery problem with near-linear complexity for unbounded subsets.

Weaknesses

Unbounded subset size may be a rather strong assumption, and the results for bounded subset size of non-adaptive algorithms are not near-linear and almost trivial for constant subset size. One may be interested in few-round adaptive algorithms with bounded subset size.

Questions

/

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

/

Authorsrebuttal2024-08-08

Thank you for seeking clarification. We apologize if there were some ambiguities in the previous response. Our result is optimal for all values of $s \leq \sqrt{n}$ within a $\log n \log \log n$ factor. In particular, our algorithm in Theorem A.1 (or the simplified version, 1.4) achieves $O(\frac{n^2}{s^2} k \log n \log \log n)$ non-adaptive queries for any $s \leq \sqrt{n}$. Thus for constant k (which is the regime for Theorem A.1 and 1.4), our bounds are optimal for all values of $s \leq \sqrt{n}$ within the $O(\log n \log \log n)$ factor since $\frac{n^2}{s^2}$ is a lower bound for non-adaptive queries. So to summarize, it is not correct that the bound is optimal only for "the largest possible" $s \in o(\sqrt{n})$. Instead, the bound is optimal for all values of $s \leq \sqrt{n}$.

Reviewer WJzB2024-08-14

Sorry for the confusion, that was my bad, and thank you for the clarification! Edit note for review: Taking the whole rebuttal into account, I think this paper should be accepted (change score 6 -> 7).

Reviewer SHzL8/10 · confidence 4/52024-07-10

Summary

The paper studies the clustering problem with non-adaptive subset queries. The problem formulation is as follows. Suppose we are given an oracle $q: \mathcal{V}\rightarrow \mathbb{R}$ such that for any query on a subset $S\subseteq V$ of vertices, the oracle returns how many clusters are in $S$ in the optimal clustering. The goal is to recover the optimal clustering with subset queries that are as small as possible. The problem setting is similar to the clustering with same-cluster query problem, where the oracle returns whether two vertices $(u,v)$ are in the same cluster in the optimal clustering. Both problems are motivated by scenarios where we could use semi-supervised learning to obtain such types of oracles. However, in that setting, the number of necessary queries becomes $\Omega(n^2)$ when $k$ is at least $3$. As such, to circumvent this strong barrier, it is natural to look into the new model with extra power for the queries. In the new model, we can easily obtain an algorithm $O(n\log{k})$ queries by binary search. However, such an algorithm requires at least $\Theta(\log{k})$ rounds of adaptivity. The paper focused on the non-adaptive setting, where the queries have to be made *independent* of the results of previous queries. The main results include $i).$ an algorithm that makes $O(n \log\log{n})$ queries when $k=O(1)$ (linear dependency on $k$), and; $ii).$ an algorithm that makes $O(n\log^{2}{n} \log{k})$ queries for general $k$. The paper also studied the case when the maximum size of $S$ is bounded and obtained tight bounds when ${|S|}\leq \sqrt{n}$. **Techniques.** The two main algorithms used different ideas. For the algorithm with constant $k$, the key observation is that after recovering the biggest cluster(s) in the queried set, there are non-trivial probability that the remaining of the queried set has exactly one or two points. For the two cases, we can recover clusters via subset queries for some specific sizes. As such, we can repeat enough times, and use a ‘valid’ set to recursively recover the clusters in the reconstruction phase. For the algorithm with general $k$, the paper reduces the problem to solving the OR queries of boolean vectors. In this way, the paper provides some new results for OR queries and obtains the desired bound.

Strengths

I have a positive view of the paper. In my opinion, the problem setting is well-motivated, and it is a natural extension of the well-studied same-cluster query model. The paper is well-written, and the authors did a nice job explaining the ideas in the high-level overview. Therefore, although the techniques are non-trivial, I managed to follow the ideas well. Given the page limits of the conference, the paper still managed to do fairly well in terms of organization.

Weaknesses

I don’t see any major issue in the paper. Some minor issues to address: - It might be good to further clarify what ‘non-adaptive’ means in your paper. I think in your algorithm, you could still adaptively ‘add on’ to clusters *after* making the queries. For a moment I thought your algorithm could conduct $\log {n}$ rounds *in parallel* to recover all clusters, which will be even stronger in the ‘non-adaptive’ notion. - The paper doesn’t contain experiments. I’m personally OK with it, but I think for NeurIPS you might want to justify why this is not an issue. I think AKB [NeurIPS’16] (‘clustering with same-cluster queries’) gives a nice motivation for the same-cluster oracle model, and maybe this paper should provide some similar justifications. - The dependence on $k$ should be mentioned after theorem 1.2. I think $O(nk)$ is really not that bad – I thought it was something like $2^k$.

Questions

Most of the questions are asked in the ‘weakness’ part. Two MISC questions: - Line 165: should it be $p\leq \log\log{n}$? - If we just want the clustering of an $\alpha$ fraction of the vertices, can we use an even smaller number of queries?

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

I do not see any potential negative societal impacts.

Reviewer MjFn6/10 · confidence 4/52024-07-12

Summary

The paper gives results on clustering a dataset using "subset queries." This is a generalization of "same-cluster queries." The same cluster query asks whether two given elements belong to the same optimal cluster. The subset query asks how many different clusters the elements of a given subset span. There has been recent interest in clustering using same-cluster queries. This paper is an extension of this line of work. Clearly, O(n^2) queries are sufficient to cluster all the points. We would want to explore the possibility of making a much smaller number of queries. There are two sub-models to consider when discussing subset queries: 1. Adaptive queries 2. Non-adaptive queries. Adaptive queries allow the query algorithm to adapt new queries based on the response for the earlier queries, whereas, in the non-adaptive model, all subsets should be predefined before making any queries. Adaptive queries are more powerful since O(nk) adaptive same-cluster queries are sufficient, whereas O(n^2) are required in the non-adaptive setting. This paper gives results in the non-adaptive setting: - An O(n log^2{n} log{k}) query algorithm for general values of k. This is using a reduction to a known problem in group testing. This is improved to O(n log{k}) for the balanced case (when cluster sizes are within a constant factor of each other). - An O(n log log{n}) algorithm for constant values of k. This is using a simple randomized method.

Strengths

- From a purely theoretical point of view, the results are interesting. The discussion is elaborate, with various special cases discussed. Except for logarithmic factors, the bounds are tight. - The writing is good. The intuition is clearly stated within the main write-up, and the proofs are in the appendix.

Weaknesses

- The motivation for considering subset queries is not immediately clear. The motivation for same-cluster queries was easier to understand since one can set up a crowdsource platform and ask users to check whether two objects (say images) belong to the same class. Setting up subset queries for large subsets may be tricky since answering such queries is not simple.

Questions

- Subset queries may not be easy to answer accurately, specifically when the subsets are large. However, answering such queries approximately (within some margin of error) may be possible. It may be interesting to explore whether some interesting clustering can be obtained using a small number of such "inaccurate approximate queries."

Rating

6

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

This work is theoretical, and there are no potential negative societal impact.

Reviewer WJzB2024-08-08

Thank you for your response! If I understand correctly, the details you provided agree with my understanding at large, and in particular, the bound is essentially optimal for "the largest possible" $s \in o(\sqrt{n})$ only (and not all $1 \leq s \ll \sqrt{n}$). If that's not the case, please follow up and feel free to let me know.

Reviewer SHzL2024-08-09

Thanks for the clarifications and the response. I do not have any further questions.

Reviewer hP142024-08-12

Thank you taking time to respond to my questions. I appreciate the added motivation for subset queries and the clarification regarding constant-vs-general k. All the best for your submission !

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC