Bicriteria Approximation Algorithms for the Submodular Cover Problem

In this paper, we consider the optimization problem Submodular Cover (SCP), which is to find a minimum cardinality subset of a finite universe $U$ such that the value of a submodular function $f$ is above an input threshold $\tau$. In particular, we consider several variants of SCP including the general case, the case where $f$ is additionally assumed to be monotone, and finally the case where $f$ is a regularized monotone submodular function. Our most significant contributions are that: (i) We propose a scalable algorithm for monotone SCP that achieves nearly the same approximation guarantees as the standard greedy algorithm in significantly faster time; (ii) We are the first to develop an algorithm for general SCP that achieves a solution arbitrarily close to being feasible; and finally (iii) we are the first to develop algorithms for regularized SCP. Our algorithms are then demonstrated to be effective in an extensive experimental section on data summarization and graph cut, two applications of SCP.

Paper

Similar papers

Peer review

Reviewer Swbd6/10 · confidence 4/52023-07-03

Summary

This work focuses on designing approximation algorithms for the submodular cover problem. In this problem, we are given a submodular function $f$ over subsets of a ground set $U$, and the goal is to find the smallest subset $S$ such that $f(S)$ is greater than or equal to a given threshold. The authors present a fast algorithm for monotone functions with similar approximation guarantees as greedy, while being almost a factor of $n$ faster. They improve the approximation guarantee with an algorithm that runs in exponential time with respect to the size of the optimal solution. They extend their results to non-monotone functions and achieve a high-quality solution with exponential running time. Furthermore, they consider the regularized submodular cover problem, which is similar to the submodular cover problem minus an additional modular function. They are the first to study this problem and present bicriteria algorithms for this problem. Moreover they compare their algorithms with baselines in multiple experiments.

Strengths

The authors' work makes significant contributions to the field of submodular cover optimization. Their algorithms are simple and easy to understand, and the results are explained clearly. The authors also provide a fair comparison with previous works. The experimental section is well-organized and the datasets used are commonly used in submodular maximization publications in ML conferences. The results show the strengths and weaknesses of this work, and support the theoretical guarantees.

Weaknesses

First, the novelty of the work is not clear to me. The algorithms and proof techniques are similar to previous works and are well-known in the literature. Second, exponential time algorithms are not practical in most cases, and I do not agree with the reasons provided in this work. Third, the results for Regularized SCP are good, but less interesting compared to the rest of the results in this work.

Questions

Please provide details if the weaknesses mentioned above are not accurate.

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

yes

Reviewer AFAK4/10 · confidence 3/52023-07-06

Summary

The paper studies bicriteria approximation algorithms for sub-modular cover problem (SCP). In the problem, we are given an oracle to some sub-modular function f, and a threshold tau. Our goal is to find the smallest set X such that f(x) >= tau. The paper is not well-written at all. There are critical typos in the theorem statements, making it hard to understand what the main results are. Only after reading the rebuttal and the supplementary materials carefully, I was able to put all the pieces together. Two main results of the paper are For the monotone SCP problem (MSCP), the result is an algorithm that achieves O(ln(1/eps))-approximation with 1-epsilon violation on the value, using O_eps(n log n) queries. This improves the previous query complexity of O(n^2). However, the algorithm is called threshold-greedy-c, described in Algorithm 5 in the supplementary material. But the theorem statement says the algorithm is threshold-bi, which is never defined. The second result is for SCP. It says there is an algorithm that achieves 1/epsilon approximation with 1-epsilon violation on the value. The result uses theorem 2, which converts a randomized algorithm for the dual problem SMP into a randomized algorithm for SCP. The algorithm needs to guess a budget g, and repeat the algorithm for SMP multiple times in order to convert the expectation guarantee into high probability guarantee. However, in the theorem statement, gamma should be epsilon as pointed by the authors in the rebuttal, and alpha is never defined. From the supplementary material, I learnt that alpha is the parameter controlling the multiplicative step size of the guessed size budget g, and the term log |OPT| in the running time should really be log_{1 + alpha} |OPT|. I am raising my score to borderline reject. There are so many typos in the theorem statements. Yes, they are just a few typos in the whole paper, but I think mistakes in the main theorems, which give the formal descriptions of main results, are intolerable. They are the first things that people will read, and they only take a few lines in the paper. If I needed to judge the contribution of the results, without taking the typos into consideration, I would say the first result makes a fair improvement, but I do not see too much technical contribution from the second result. It is based on the dual relationship between SCP and SMP, and conversion from expectation to high probability comes from repeating the algorithm multiple times.

Strengths

The paper improved the query complexity for the MSCP algorithm from O(n^2) to O(n log n).

Weaknesses

The paper is not written well. There are critical typos in the statements of the main theorems.

Questions

No questions.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

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

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

None.

Reviewer DEGK8/10 · confidence 3/52023-07-07

Summary

This paper proposes several bicriteria approximation methods for solving the standard monotone submodular cover problem (SCP). Additionally, the authors also propose new variants of the problem: removing the monotone assumption, producing nearly feasible solutions, and adding regularization via a modular cost function. The general methodology involves novel conversions from common bicriteria algorithms for the (dual) submodular maximization problem (stochastic greedy, threshold greedy, distorted greedy) into ones for SCP.

Strengths

- Significance: This paper uses a powerful framework which leads to not only novel algorithms but also new problem formulations. Expanding the scope of SMP/SCP conversion is of independent interest and has high potential impact - Originality/Clarity: Novel combination of previous work and new ideas (discussed and cited clearly)

Weaknesses

- Experiments: The impact of the regularized SCP problem would be more significant if the paper included experimental results comparing convert-reg and distorted-greedy-bi. The paper would be improved if it provided additional motivations/applications for each SCP formulation. When is SCP preferable to SCM? - Typo: distorted algorithm is distorted-greedy-bi in the main paper and distorted-bi in the appendix

Questions

- What are some applications of regularized SCP? - Are any bicriteria approximations or query complexity results optimal, or can they be improved further? ----- EDIT: I have read the author rebuttal, and it addressed my questions sufficiently.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

No discussions of limitations or broader impact

Reviewer fbw17/10 · confidence 2/52023-07-08

Summary

The paper considers the Submodular Cover Problem (SCP), where one is given a submodular function f through an oracle that returns the value f(A) for each subset A of the underlying set U (over which the function f is defined). The goal is to find a minimum cardinality subset X of a set such that the value f(X) is >= a given threshold \tau and |X| is a minimum. Since this problem is NP-hard in general, researchers have considered bicriteria approximation algorithms, where the size of the size of the returned subset X is within a small factor of the size of a minimum subset and the value f(X) is at least a suitable fraction of the given threshold. While some such approximation algorithms are available (using a dual problem), they may use quadratically many queries to the oracle. The focus of the paper is to develop methods whose approximation guarantees are close to known results but which use significantly fewer queries to the evaluation oracle. Results are provided for general SCP as well as some restricted versions.

Strengths

(1) The paper presents new methods for SCP with significantly smaller numbers of queries to the function evaluation oracle without significantly affecting the performance guarantees. To this reviewer's knowledge, this is the first work that achieves this goal for SCP. Given the importance of submodular optimization in ML these results represent a useful advance. (2) The paper nicely summarizes prior work on the topic.

Weaknesses

A (very) minor weakness is that it takes a fair amount of time to understand the details regarding the algorithms. (This is due to the nature of the problem.)

Questions

(1) In the results for the general SCP (lines 50--54, page 2), it is mentioned that the algorithm does not run in polynomial time in general. Is there a known hardness result which precludes a polynomial time algorithm for this case (under some well accepted hypothesis in complexity theory)? (2) Lines 98--100 in Section 1.2 discusses a result from {Crawford, 2023]. This result (as described) says that one "cannot have an algorithm for SCP which can guarantee that f(X) >= \tau/2 using only polynomially many queries". Does this result rely on an underlying complexity hypothesis (such as P != NP)? Some minor suggestions to the author(s): (a) In stating the results for RCSP (lines 56--62 on page 2), please consider indicating the number of queries to the oracle. (b) Line 65: "make a large" ---> "provides a large" (c) Line 120 (page 3): "about queries" ---> "with respect to the number of queries"

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

None

Reviewer AFAK2023-08-11

After reading the paper and supplementary materials more carefully, I am able to put all the pieces together. Other than the results for regularized SCP and the empirical study, there are two main results given in the paper. The first result (i) is for MSCP. It says there is an algorithm that achieves O(ln(1/eps))-approximation with 1-epsilon violation on the value, using O_eps(n log n) queries. This improves the previous query complexity of O(n^2). The algorithm is threshold-greedy-c, described in Algorithm 5 in the supplementary material. But the theorem statement says the algorithm is threshold-bi, which is never defined. The second result is for SCP. It says there is an algorithm that achieves 1/epsilon approximation with 1-epsilon violation on the value. The result uses theorem 2, which converts a randomized algorithm for the dual problem SMP into a randomized algorithm for SCP. The algorithm needs to guess a budget g, and repeat the algorithm for SMP multiple times in order to convert the expectation guarantee into high probability guarantee. However, in the theorem statement, gamma should be epsilon as pointed by the authors in the rebuttal, and alpha is never defined. alpha controls the precision of the guessed size budget g, and the running time should have the term log_{1 + alpha} |OPT|.

Authorsrebuttal2023-08-11

We do have a number of different results on several problems, and so we plan on including a table in the next version of the manuscript that makes each contribution and how they fit in more clear. We would like to clarify the main contributions of the paper as follows: (1) We have three main results for the monotone submodular cover problem (MSCP) in Section 2.1. These results only apply for MSCP. The first is the algorithm threshold-greedy-c (Algorithm 5 in the appendix) and its guarantees in Theorem 1. The second is the converting method convert-rand that takes randomized algorithms for monotone submodular maximization and converts them into ones for MSCP (Algorithm 6 in the appendix) and its guarantees in Theorem 2. The third is the algorithm stoch-greedy-c (Algorithm 1 in the main text) and its guarantees in Theorem 3. The theorem statements will be edited in order to make it more clear that they apply only to monotone submodular functions (it currently says monotone both in the text and section headings, but some of the theorem statements are unclear). We believe our most interesting result from this section is stoch-greedy-c, which is inspired by the stochastic greedy algorithm for monotone SMP [Mirzasoleiman et al., 2015a] but we use a method of adaptively guessing the size of the optimal solution throughout the algorithm to propose a new sample-efficient algorithm for MSCP. (2) We have one main result for the general submodular cover problem (SCP), where $f$ is submodular but not necessarily monotone. This is the algorithm stream-c (Algorithm 2) and its guarantees in Theorem 4. (3) Finally, we have a result for the regularized monotone submodular cover problem (RSCP), which does not fall under the setting of SCP since the objective may take on negative values. We propose a method of converting algorithm for the regularized monotone submodular maximization problem, convert-reg (Algorithm 3 in the main text) and its theoretical guarantees are in Theorem 5. We then propose the algorithm distorted-bi for regularized monotone submodular maximization that produces different approximation guarantees compared to existing ones in the literature (see Section 2.3 in the paper for more details on this), in order to be used by convert-reg to produce an algorithm for RSCP. In addition, "threshold-bi" is meant to say threshold-greedy-c. We will fix this additional typo in the next version of the manuscript. The alpha in Theorem 2 is an input parameter which we also mentioned in the corresponding pseudocode convert-rand (Algorithm 6 in the supplementary). We will clarify it more clearly in Theorem 2. The reviewer is correct that the term $\log(|OPT|)$ in Theorem 2 should say $\log_{1+\alpha}|OPT|$. We want to point out that most of the typos and ambiguity is coming from Theorem 2 and its corresponding algorithm convert-rand. We believe part of the problem is that much of the information on convert-rand is in the supplementary material. We plan to fix the typos as well as add more discussion about convert-rand and its theoretical guarantees in the main paper.

Reviewer AFAK2023-08-12

I raised my score to weak reject. There are so many typos in the statements of the main theorems, and I think they are intolerable.

Reviewer fbw12023-08-12

I have gone through the rebuttal. My questions/concerns have been addressed satisfactorily.

Reviewer DEGK2023-08-14

Reply to Rebuttal

The authors have addressed my questions sufficiently.

Reviewer Swbd2023-08-17

I thank the authors for their response.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC