Replicable Clustering

We design replicable algorithms in the context of statistical clustering under the recently introduced notion of replicability from Impagliazzo et al. [2022]. According to this definition, a clustering algorithm is replicable if, with high probability, its output induces the exact same partition of the sample space after two executions on different inputs drawn from the same distribution, when its internal randomness is shared across the executions. We propose such algorithms for the statistical $k$-medians, statistical $k$-means, and statistical $k$-centers problems by utilizing approximation routines for their combinatorial counterparts in a black-box manner. In particular, we demonstrate a replicable $O(1)$-approximation algorithm for statistical Euclidean $k$-medians ($k$-means) with $\operatorname{poly}(d)$ sample complexity. We also describe an $O(1)$-approximation algorithm with an additional $O(1)$-additive error for statistical Euclidean $k$-centers, albeit with $\exp(d)$ sample complexity. In addition, we provide experiments on synthetic distributions in 2D using the $k$-means++ implementation from sklearn as a black-box that validate our theoretical results.

Paper

Similar papers

Peer review

Reviewer t2BD7/10 · confidence 2/52023-07-05

Summary

This paper studies the design of clustering approximation algorithms in the context of statistical clustering under the notion of replicability. In replicable clustering problem, it requires that with high probability, the output of the algorithms should have the exact same partition of the sample space after two executions on different inputs drawn from the same distribution. Given some black-box $\beta$-approximation algorithm to the $k$-clustering problem, this paper gives an approximation framework that with constant probability, the output of the algorithm can achieves an $O(\beta)$-approximation satisfying $\rho$-replicable property.

Strengths

This paper is the first to initiate the study of replicable clustering, which is a recent hot topic in the field of machine learning. In this paper, a new framework is introduced for handling replicable clustering requirements. This paper also establishes sampling complexities for clustering problems in different norms with different metrics. In order to achieve replicability, a novel tree decomposition method called Replicable Quad Tree is designed such that replicable estimation can be obtained based on the tree decomposition constructed. The proposed Replicable Quad Tree is interesting and is quite different from the well-known HST-based tree decomposition method, which could have great potential for designing approximation algorithms for other replicable problems.

Weaknesses

1. The sample complexity has exponential dependence on the dimension $d$, which could be the main weakness for this paper when $d$ is large. Although in Euclidean metrics, this can be avoided by dimensionality reduction techniques, for general norms, there is no data-oblivious dimensionality reduction scheme. 2. The main proofs of this paper are hard to follow. It is better to provide intuitive ideas in each section before the proofs such that one could easily follow the main ideas of the proofs.

Questions

1. Can the authors offer some intuitive insights into how the proposed Replicable Quad Tree ensures replicable estimation?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

Since this is a theoretical paper, I don't think this paper has any potential negative societal impact.

Reviewer nt3r6/10 · confidence 4/52023-07-06

Summary

This manuscript focuses on the concept of replicability in statistical clustering algorithms. It introduces the notion of replicability, which refers to the ability of an algorithm to produce consistent results when executed multiple times on different inputs from the same distribution. Replicability is seen as crucial for ensuring the validity and reliability of scientific findings. The manuscript proposes replicable algorithms for three common clustering problems: k-medians, k-means, and k-centers. These algorithms utilize approximation routines for their combinatorial counterparts. The replicable algorithm for statistical Euclidean k-medians achieves a replicable O(1)-approximation with polynomial complexity. Similarly, the algorithm for statistical Euclidean k-centers achieves replicable O(1)-approximation with an additional O(1)-additive error but with exponential sample complexity. The manuscript also discusses the importance of clustering algorithms in unsupervised learning and the lack of a universally agreed-upon definition for the quality of clustering solutions. It highlights the challenges and trade-offs that algorithm designers face due to factors such as random initialization, similarity measures, noise in measurements, and outliers in the dataset. These issues contribute to non-replicable results in clustering algorithms. Furthermore, the manuscript provides related work on norms and parameters used in clustering and presents experimental results on synthetic distributions to validate the proposed replicable algorithms.

Strengths

The manuscripts' innovation lies in proposing replicable clustering algorithms for statistical clustering problems. It introduces the concept of replicability and highlights its significance in ensuring the validity and reliability of scientific findings. The document aims to address concerns about replicability in clustering algorithms, which is a crucial topic in the subfields of machine learning and data science. It presents replicable algorithms for k-medians, k-means, and k-centers problems and outlines their theoretical results. Overall, the manuscripts' innovation lies in its contribution to the development of replicable algorithms for statistical clustering.

Weaknesses

There are many theorems that are not well understood, and then there are relatively few experimental parts.

Questions

1] How does the proposed replicable algorithm compare to existing clustering algorithms in terms of performance and accuracy? 2] Are there any limitations or assumptions in the proposed replicable algorithms that could affect their applicability in real-world scenarios? 3] How generalizable are the experimental results presented on synthetic distributions in 2D? Can the replicable algorithms be applied to other types of datasets? 4] How sensitive are the replicable algorithms to the choice of parameters such as the choice of exponent in the cost function or the measure of similarity of the data? 5] Does the study discuss any potential implications or future research directions regarding the replicability of clustering algorithms?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

In the experiment, real-world data sets are not used, which is not very convincing.

Reviewer fyTj6/10 · confidence 3/52023-07-07

Summary

This paper studies the concept of replicability in clustering. This topic is important because replicability is what allows for other researchers to reproduce the results of a study to verify their correctness. This is a big problem these days, 50% of scientists saying that there is a replicability crisis. In the interest of improving the scientific verification process, designing replicable algorithms (i.e., algorithms that function the same when executed twice with the same random seed). This differs from the related notion of clustering stability in that its success is more dependent on the algorithmic design than structures in the data. In their setting, points are sampled from a random distribution over the d-dimensional unit ball according to some given metric. It models many norms, including \ell_p norms. They use Impagliazzo et al.’s notion of rho-replicability: The probability (over every two sampled sets of points and a single random seed) that the algorithm on each set but with the same random seed yields the same output is at least rho. Outputs to these clustering problems are functions that, given a point, assigns it to a cluster (with no explicit center, just an identifier). The main results are new algorithms for statistical k-means and k-medians, with extended improvements when using the Euclidean metric. Their approximation factors both near the approximation factor of a given black-box vanilla algorithm, and their sample complexity on general metrics has an exponential dependence on the dimension (it is sub-exponential in the Euclidean version). They also show an approximation algorithm for the replicable statistical k-centers algorithm. Given an approximation of the form a*OPT+b for k-centers, they give an approximation of the form c*OPT+d, where c and d are linear in both a and b. The query complexity is exponential in the dimensionality. Their algorithm uses a quadtree decomposition, which creates a tree structure over all points. The leaves (which correspond to points) are then mapped to cluster centers. So when you receive a query for a point, you traverse the tree to find the leaf which contains it (as it is a partition of the space), and you map it to the corresponding cluster. However, in order to adequately depict the distribution of the space with their decomposition (which they are not given), they require access to an exponential number of queries in d. In the Euclidean case, they can first reduce the dimensionality via Johnson-Lindenstrauss, and map an epsilon-net from the original space to the low-dimensional space. There still is the difficulty of mapping the clustering between these two spaces, but it seems most of those details are in the appendix. For replicability, they form a grid, sample points in the space, and map them to the centers of cells to estimate probability mass of each cell. On high-mass cells according to a random threshold, and then they apply a vanilla algorithm. Finally, they compare k-means++ and k-means++ run on the core produced for their replicability results. It is pretty clear that their approach leads to much greater replicability.

Strengths

This paper is written extremely clearly. I really appreciate how slow and careful the introduction and preliminaries are, which I believe could be followed by someone unfamiliar with the field. For instance, while many of us take for granted that a replicable clustering algorithm optimizes for some utility notion, they take the time to explain this and how it is distinct from their replicability goal (which is a bigger focus of their preliminary section). They also take the time to point out to the reader that their metric function and the ell_p parameter p are different, and give an in-depth but understandable explanation of the intuition behind ell_p for different values of p. The informal presentation at the beginning of results in section 3 is also very nice. The results are nice and interesting, but I wouldn’t say they are groundbreaking. The techniques have clear foundations in previous methods, but it also seems unique how they combined them together. I wasn’t sure by reading the paper how their replicability results compare to previous literature.

Weaknesses

I found it very odd that the main focus of the paper introduction and title was replicable k-centers, but they spend most of their time on (nonreplicable) k-centers. It seems like if the highlight of the paper was replicability, they should have focused on the methods used to achieve this. I feel the details of this algorithm were insufficient for a reader invested in replicable k-centers. Additionally, I wasn’t sure how their results compared to any baseline results that exist, so it is difficult to judge the novelty.

Questions

1. I was a bit confused by the sentence “We emphasize that even though we only observe a sample from P we aim to solve the problem on the whole support of the distribution.” Can you clarify exactly what this means? Is it that the algorithm works on an actual given dataset P but needs to be able to work on all possible datasets? I would think that would be assumed. 2. How do your algorithms compare to existing algorithms? Is there any baseline implied by any other work?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

There was no limitations section, though it did not seem needed (but would be preferred).

Reviewer Jm7K6/10 · confidence 2/52023-07-18

Summary

The topic of this paper is replicable clustering which in high-level asks for design of an algorithm that given two runs of the algorithm on different samples from the “same” input distribution. Replicability is a notion introduced recently in a work by Impagliazzo et al. [2022]. The algorithm for statistical clustering with centroid-based objective with $p$-norm cost. For the case of $k$-means and $k$-medians, via corresponding centroid-based $k$-clustering on “finite” size (referred to as combinatorial clustering) in a black-box manner and design $O(1)$-approximation with poly($d$) sample complexity (but exponential in k?). However, the case of $k$-center is more complicated as it depends on the $\ell_\infty$-norm and very sensitive to the largest distance. The results on $k$-center has worse approximation guarantee and sample complexity and requires stronger assumptions. In particular, for the case of $k$-center, the guarantee is bicriterion.

Strengths

At high-level, the notion is interesting and related to several other important topics too. The technical contribution of the paper seems to be satisfactory. However, I'm less familiar with the literature and the technicality of the paper might not be significant for someone working on this area or differential privacy.

Weaknesses

The paper is following the notion of [Impagliazzo et al., 20]; however, I still not quite convinced why the shared randomness assumption is meaningful. A related question is whether an assumption such as Assumption F. 2 can be sufficient for achieving replicability? The result seems to heavily depend on the coreset construction of Frahling and Sohler [2005]. Can you elaborate why more recent coreset construction can be applied here? Theorem 3.1 and Theorem 3.2 provide upper bounds but there is no lowerbound. E.g., Any hope to show exponential dependence on k (or d) is needed (d, for the general metric space)? How the algorithm in this paper is different from those for differentially private clustering? Can you elaborate on the proof of Corollary E. 9? I particular, how does the dimension show up?

Questions

Please see the questions in the weakness section.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

There is no discussion of limitation by the authors. Though, the paper does not seem to have potential negative societal impact.

Reviewer RdwG7/10 · confidence 4/52023-07-31

Summary

This paper initiates the study of formal replicability for clustering algorithms. Replicability is defined to be a property of an algorithm for a statistical clustering problem, requiring that fixing the internal randomness of the algorithm while resampling the input data will yield exactly the same (representation of a) mapping from points to clusters. Upper-bounds are given for the statistical k-means, k-medians, and k-centers problems, with poly(d) sample complexity proven via dimensionality reduction in the case of euclidian distance. The theoretical results are also empirically evaluated on synthetic data.

Strengths

This paper addresses an interesting problem by extending recently introduced formal notions of replicability to the task of statistical clustering. Clustering seems like a setting where replicability is particularly well-motivated, as there are many reasonable desirable properties for clustering solutions, and being able to ensure that solutions which demonstrate a particular balance of these properties are in fact a result of the algorithm generating solutions, and not just a fluke of the sample, is a natural objective.

Weaknesses

The writing was very readable, but it would have been nice to have a more thorough comparison of this work to the related work. What is the cost of replicable clustering compared to non-replicable algorithms for the same problems? Does this notion of replicability solve any of the issues that motivated the "stability of centers" condition for selection of k that is mentioned in Section 1.1? I also found Section 4 a bit difficult to parse. I think some of the notation used in Algorithm 4.1 isn't defined, and though it's reasonably intuitive, using more precise language in the high-level exposition and fully defining all variables/notation would definitely improve the readability of this section.

Questions

See weaknesses.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

Yes, the authors address all limitations.

Reviewer RdwG2023-08-18

I thank the authors for their detailed response. I will keep my score.

Reviewer t2BD2023-08-19

I would like to thank the authors for their response and clarification. After looking at other reviews and corresponding rebuttals, I have no further questions.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC