Constant Approximation for Individual Preference Stable Clustering

Individual preference (IP) stability, introduced by Ahmadi et al. (ICML 2022), is a natural clustering objective inspired by stability and fairness constraints. A clustering is $\alpha$-IP stable if the average distance of every data point to its own cluster is at most $\alpha$ times the average distance to any other cluster. Unfortunately, determining if a dataset admits a $1$-IP stable clustering is NP-Hard. Moreover, before this work, it was unknown if an $o(n)$-IP stable clustering always \emph{exists}, as the prior state of the art only guaranteed an $O(n)$-IP stable clustering. We close this gap in understanding and show that an $O(1)$-IP stable clustering always exists for general metrics, and we give an efficient algorithm which outputs such a clustering. We also introduce generalizations of IP stability beyond average distance and give efficient, near-optimal algorithms in the cases where we consider the maximum and minimum distances within and between clusters.

Paper

References (25)

Scroll for more · 13 remaining

Similar papers

Peer review

Reviewer kydQ7/10 · confidence 3/52023-07-05

Summary

This paper continues the study of individual preference stability (IP stability) initiated by Ahmadi et. al. (2022). Given a set of points in a metric space, a k-clustering is said to be 1-IP stable if for every point in the set, its average distance to the points in its cluster is at most its average distance to the points in a different cluster; so, in this sense, a point prefers its own cluster to other clusters. There are datasets for which a 1-IP stable clustering does not exist, and further, it is NP-hard to decide whether a dataset admits a 1-IP stable k-clustering. In light of this, it is natural to broaden the definition to approximate IP stability: a k-clustering is \alpha-IP stable if the average distance of a point to another cluster is at most \alpha times its average distance to its own cluster. The work of Ahmadi et. al. shows that an O(n)-IP stable clustering always exists and gives an algorithm for computing one. The present work closes the gap between the known lower and upper bounds: the authors propose an algorithm that always outputs an O(1)-stable clustering (so they also prove that one always exists, a fact not already established). Interestingly, the output clustering has the additional property that it is a constant factor approximation for k-center. Indeed, the greedy k-center algorithm of Gonzalez is a phase in their algorithm, and the "ball-carving" phase of their algorithm is similar in spirit to another approximation algorithm for k-center (although there are important nuances in the algorithm in the present work that differentiate it). The fact that the algorithm doubles as an approximation algorithm for k-center, the authors note, ensures that the O(1)-IP stable clustering is in some sense meaningful; for, they demonstrate that a random k-coloring a graph will produce an O(1)-IP stable clustering (for certain k), but of course such a clustering is not meaningful in general. The authors also introduce two variants of IP stability. For Min-IP stability, in which "average distance" is replaced with "min distance" in the definition of IP stability, they show that the Single Link algorithm produces an exactly stable (optimal) clustering. For Max-IP stability (defined analogously), they show the greedy k-center algorithm of Gonzalez gives a 3-IP stable clustering. Finally, the authors show experiments comparing their algorithms to k-means ++ as a baseline. While their algorithm performs slightly worse than k-means ++ for IP stability in general, they claim its robustness by showing a hard instance for which it outperforms k-means ++. They also show that in practice k-means ++ can be up to five times worse than their (optimal) algorithm for Min-IP stability.

Strengths

- This work closes the (previously large) gap between upper and lower bounds for IP-stability. It demonstrates the existence of constant-stable clusterings and how to find them efficiently. It also provides a convincing argument that a clustering produced by the algorithm is meaningful by showing that it also gives a constant approximation for k-center, which I think is a key contribution given that the authors show, on the flip side, that a clustering can satisfy O(1)-stability but otherwise not uncover meaningful structure. - The algorithm takes inspiration from approximation algorithms for k-center but there are important subtleties in the adaptation of these algorithms. The Ball-Carving algorithm (Algorithm 1) is a more nuanced version of an approximation algorithm for k-center (in which one guesses the optimal radius, and then repeatedly cuts out balls of that radius until no points remain). In Algorithm 1, more care is taken as to how to choose the center of the ball at each iteration (instead of being chosen arbitrarily as in the approximation algorithm) as well as how to assign points to centers (instead of just carving out a ball of a certain radius). Finally, the centers from running the algorithm of Gonzalez are used to prune the number of clusters produced in the carving algorithm. - The authors repurpose existing algorithms (Single Link clustering and the greedy algorithm of Gonzalez) to give guarantees for Min- and Max- IP stability. This is interesting in its own right because it shows that the notion of stability is intimately related to other clustering objectives.

Weaknesses

- While Algorithms 1 and 2 are described very clearly, little motivation or intuition is given. For instance, it would be useful to provide examples that show the necessity of the nuances in the Ball-Carving algorithm. - A discussion of how Algorithms 1/2 differ from or build on the algorithms of Ahmadi et. al. would help highlight the novelty of the contributions. - A more specific comment: The paragraph on lines 66-70 seems inconsistent with later claims (unless I have misinterpreted something). It seems that if the parameter r is chosen to be anything, then the properties claimed in lines 66-70 hold, but potentially with more than k clusters. Only when r is chosen as a function of r_0 and then the pruning step in Algorithm 2 is performed do the properties seem to hold with k clusters. If this is the case, then I think the current wording in lines 66-70 is easily misinterpreted. Still, the stronger guarantee of uniform IP stability with r chosen as a function of r_0 is noteworthy. - It seems that Algorithm 3 is simply the well-known algorithm of Single Link clustering. This should be labelled as such and cited.

Questions

- A variant of Single-Link clustering is used in the work of Awasthi, Blum, and Sheffet (2012) to find optimal clusterings on perturbation stable instances. Given the algorithmic similarities, do you see any connections between these two notions of stability? - Do you have any intuition as to why k-center approximation algorithms in particular are useful for IP-stability algorithms, given that IP-stability (in its original form) is based on average distances? While k-median and k-center are in general different notions, it is perhaps not clear a priori which objective could help inspire an algorithm for IP-stability. Have you considered the question of finding an algorithm that produces a clustering that is O(1)-IP stable and also constant approximate for other objectives?

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

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

Not applicable.

Reviewer F6zy7/10 · confidence 5/52023-07-07

Summary

This paper considers the problem of finding stable clusterings under a specific notion of stability - individual-preference stability which roughly requires the clustering produced to have the property where the average distance of any datapoint to points within its own cluster to be smaller than the average distance to points within any other cluster. The clustering is said to be $\alpha$-stable if the average in-cluster distance to be no more than a multiplicative $\alpha>1$ factor of the average distance to points assigned to any other cluster. This is a fairly natural problem, and was recently proposed by Ahmadi et. al. (ICML 22), who gave some preliminary results for it, including NP-Hardness of deciding whether the input dataset admits a $1$-stable clustering. This paper makes three key contributions - (a) It shows that given any dataset with $n$ points in a general metric space, and any desired number of clusters $k$, there always exists a $k$-clustering that is $O(1)$-stable. (b) This $O(1)$-stable $k$-clustering can be found by a computationally efficient algorithm. Moreover, the resulting clustering produced is a constant factor approximation to the optimal $k$-center objective. (c) For min and max stability (i.e. where the average is replaced with min and max, respectively), they show that for any dataset in a general metric space, and for any choice of $k$, (i) a min-stable $k$-clustering always exists, and is achieved by the usual early-termination of Kruskals minimum spanning tree algorithm, and (ii) a 3 approximate max-stable $k$-clustering always exists and is achieved by the greedy $k$-center algorithm.

Strengths

This paper considers a very natural question of obtaining a stable clustering, and substantially expands our understanding of this problem. Overall the paper is quite well-written and easy to read, and would be a well-suited for a venue such as NeurIPS. (a) They show that given a set of any $n$ points in an arbitrary metric space, a $O(1)$-stable $k$-clustering always exists for any choice of $k$. Moreover, the clustering includes all $n$ points. The only known results prior to this work was the existence of a $1$-stable clustering for any set of $n$ points in a $1$-D Euclidean space or metrics induced by a weighted tree, or a bicriteria $O(\log^2 n/\epsilon)$ stable clustering achieved by discarding an $\epsilon$-fraction of the points from the input set. For the stricter requirement of clustering all $n$ points, the only known result was a trivial $O(n)$-stable clustering. (b) The algorithm that finds this $O(1)$-stable $k$-clustering is simple and quite computationally efficient. The earlier bicriteria approximation result was achieved by applying the HST hammer naively. The new algorithm has the additional desirable property that the resulting clustering produced achieves a constant factor approximation for the $k$-center objective.

Weaknesses

The paper has two main weaknesses, which in my opinion are not deal breakers given its other strengths. (a) The extension to $f$-stable clusterings is obvious, and the results for $f=$ min and max stable clusterings are also very elementary. It would have been far more interesting had the authors characterized properties of a generic $f$ required such that any instance always admits an $O(1)$-$f$-stable $k$-clustering. (b) The fact that the algorithm achieves a $O(1)$ approximation to the $k$-center objective seems like an afterthought. It doesnt seem like the algorithm was explicitly designed to additionally achieve good approximation guarantees for this other objective, and more like it happened to be that way.

Questions

I understand that these questions are out of scope of this paper, but im curious about the following thing - Given how you say that its only interesting to look at stability when the algorithm additionally gives good approximation guarantees for other natural clustering objective functions, I really want to know what happens when you simultaneously want to optimize for other objective functions such as say $k$-means or $k$-medians. Is it possible to get a clustering that achieves a good approximation for these standard clustering objectives, while simultaneously being approximately stable? Or are there cases where these two objectives of clustering "quality" and "stability" are at odds of each other - achieving one objective must necessarily incur a large penalty for the other objective? My other question is related to your extension of the notion of stability, and one I have raised before in the weaknesses - can you characterize properties of a generic $f$ required such that any instance always admits an $O(1)$-$f$-stable $k$-clustering?

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

I dont see any obvious limitations of this work.

Reviewer zmPi7/10 · confidence 3/52023-07-07

Summary

This paper concerns alpha-individual preference (IP) stable clusterings, meaning clusterings such that its average distance to points in its cluster is at most alpha times greater than to that in any other cluster. Previously, only O(n)-IP solutions were known, 1-IP solutions were known to not always exist, and we did not know if there existed O(1)-IP solutions. This paper answers affirmatively on metric spaces. They provide an O(1)-IP algorithm on metric spaces which additionally is a constant-factor approximation to the k-center problem. The algorithm follows the popular greedy ball technique on metrics, which selects radius-bounded balls (bounded by O(1) times the optimal k-center solution) that cover many points and are distant from each other, and then fixes unassigned points (according to the greedy cluster order) and merges some clusters according to their closest k-center centers. Additionally, they define variants Max-IP and Min-IP (where you instead look at the maximum/minimum distances, as opposed to the averages). They show Min-IP can be solved optimally by running Kruskal’s algorithm until there are k trees in the forest. Additionally, greedy k-center 3-approximates Max-IP. Finally, the experimentally validate their IP algorithm on real and synthetic (adersarily-designed) datasets. On both, their algorithm outperforms k-means++, showing its robustness to hard instances.

Strengths

This is a nice result that bridges a significant gap posed by previous works (i.e., the gap between O(n)-IP and O(1)-IP). In addition, it is complemented by the lower bound of 1-IP, and it implies a few nice future directions for this work, including finding the minimum alpha such that an alpha-IP solution exists and characterizing when 1-IP solutions exist. The extensions to Max-IP and Min-IP are also strong. For the most part, the writing was clear (though I did find some typos here and there). The methods seemed new in many ways, though they were heavily founded based off existing methods (i.e., clustering by balls on metric graphs).

Weaknesses

The biggest weakness is that, while the improvement in approximation is significant, I think the problem is somewhat narrow in scope. I noticed the authors only cited one past work on this area (presumably, the one that proposed it), and the applications were not cited. While this does certainly seem like an interesting new problem, I am not sure of what its place is within ML research literature.

Questions

1. Can you motivate this problem a bit more? Has anyone other than [1] tried attacking this problem? 2. Do you have any citations for uses? 3. I understand that IP was the main focus of the paper, but it would be nice to have an argument for the motivations of Min-IP and Max-IP. Do you know any applications they might work for?

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

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

I did not see a limitations section. It would be preferable but is not necessary for this work.

Reviewer d1bb7/10 · confidence 5/52023-07-11

Summary

In $\alpha$-individual preference (IP) stable clustering, the average distance between every data point to other points in its cluster must be at most $\alpha$ times the average distance between it to points in any other cluster. This paper gives the first polynomial time $O(1)$-IP stable clustering algorithm in general metrics, which improves [Ahmadi et.al, ICML 2022] 's $O(n)$-IP stable result. The algorithm in this paper also has more interesting features that I appreciate. Firstly, it satisfies an even stronger uniform IP stability guarantee. Precisely, there exists a global parameter $r$, such that for every data point, its average distance to points in its cluster is bounded by $O(r)$ while its average distance to point in any other cluster is at least $O(r)$. Secondly, the k-clustering it finds is also a constant factor approximation to the classical k-center clustering problem, which makes the solution much more meaningful in applications. I think this result shows that k-center clustering can admit IP stability without paying too much. Exciting news! Moreover, the algorithm presented is clean and easy to follow with its analysis building on certain clever observations. Authors also do experiments against [Ahmadi et.al, ICML 2022] to show the practical value of their algorithm. At last, the authors also study other definitions of IP-stable and obtain similar results. Overall, I recommend "accept".

Strengths

1. The first efficient algorithm to compute O(1)-IP stable approximate clustering in general metrics. Additionally, the algorithm has more important features beyond IP stability. The result significantly improves the previous works. 2. The algorithm and analysis are interesting and have some non-trivial insights. 3. The experiments also outperform the baseline.

Weaknesses

Both the constants for IP approximation and k-center approximation may be too large. While IP stability is motivated by practical applications, I would expect that the guarantee is not a quite large number such as 2 or 3. The paper's experiment already shows that the worst case may be avoided in practice and I would like to see more discussion in this direction. Anyway, as a theoretical submission, I think this is totally fine and does not lower my rating.

Questions

1. It is argued that the running time can be improved in Euclidean space. What is the exact running time that you can achieve? What about $\tilde{O}(nkd)$? (a typical running time for Euclidean $k$-clustering.) 2. Have you formally proved that your algorithm returns an $O(1)$-approximation for $k$-center in the main text? I know it can be seen from Algorithm 2 but it is better to provide formal proof and write the exact approximation ratio.

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

I do not see the potential negative societal impact.

Reviewer kydQ2023-08-18

I thank the authors for their detailed response. While the problem itself is fairly new as other reviewers point out, its formulation is natural and contributes to the growing literature on fairness and stability in clustering. The repurposing of existing algorithms is both subtle and clean, and the simultaneous guarantees for k-Center are appealing. Moreover, there are exciting follow-up questions for further research. I maintain my evaluation.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC