NeurIPS 2020

Fast Adaptive Non-Monotone Submodular Maximization Subject to a Knapsack Constraint

Review 1

Summary and Contributions: In this paper, the authors study the problem of maximizing a submodular function subject to a knapsack constraint. They present a simple randomized greedy algorithm that achieves a 5.83 approximation. They also study the stochastic version of this problem. In experimental evaluations, they evaluate the performance of their algorithms on real and synthetic datasets.

Strengths: 1. It is an important problem. 2. The paper is well-written. 3. Experimental evaluations are shows the effectiveness of the proposed algorithms. The code is provided and the results are reproducible. === After author response === I have read the author's feedback and the other reviews. Given the marginal contribution of this work, I would keep my score unchanged.

Weaknesses: a. The main reason for my score is the level of the contribution of this paper. The contribution of this paper is marginal due to two factors: 1. The algorithm for the monotone function is extensively is studied although it is not the state of the art. 2. The ides of sub-sampling to convert an algorithm that is designed for monotone functions to non-monotone function is also well investigated: [3], [FKK], and [AEFNS]. Given these previous works, combining these two steps seems straightforward. Furthermore, the extension to the adaptive case is somewhat straightforward given the result of [25]. b. The authors do not use the sate of the art problem for maximizing a monotone submodular function subject to a knapsack constraint. [YZA] provides a tighter result. I think merging the idea of sub-sampling with the result of [YZA] improves the approximation guarantee. c. The idea of reducing the computational complexity by lazy evaluations is a direct consequence of the result of [36]. [YZA] Grigory Yaroslavtsev, Samson Zhou, and Dmitrii Avdiukhin. "“Bring Your Own Greedy”+ Max: Near-Optimal 1/2-Approximations for Submodular Knapsack." In International Conference on Artificial Intelligence and Statistics, pp. 3263-3274. 2020. [FKK] Moran Feldman, Amin Karbasi, and Ehsan Kazemi. "Do less, get more: Streaming submodular maximization with subsampling." In Advances in Neural Information Processing Systems, pp. 732-742. 2018. [AEFNS] Naor Alaluf, Alina Ene, Moran Feldman, Huy L. Nguyen, and Andrew Suh. "Optimal streaming algorithms for submodular maximization with cardinality constraints." In 47th International Colloquium on Automata, Languages, and Programming (ICALP 2020). Schloss Dagstuhl-Leibniz-Zentrum für Informatik, 2020

Correctness: I checked the proof for the non-adaptive case and the results are correct. As far as I checked, for the adaptive case, the proofs are correct. The empirical methodology supports the theoretical results.

Clarity: The paper is well written and easy to follow.

Relation to Prior Work: The relation to the previous work is clear. It is not discussed and is not clear to me why this result is not the direct consequence of the previous works. The reference to the state of the art algorithm for monotone submodular maximization subject to a knapsack constraint is missing. I believe that result could improve the guarantee provided by this paper too.

Reproducibility: Yes

Additional Feedback: 1. Please explain the difficulty of proving the guarantee, given the existing works on using the idea of sub-sampling for non-monotone submodular maximization. 2. Use the algorithm of [YZA] for the main building block of your algorithm.

Review 2

Summary and Contributions: The paper considers the problem of maximizing a (not necessarily monotone) submodular function subject to a knapsack constraint in the offline and adaptive settings. It provides 5.83 and 9 approximations for these two settings, respectively, using simple and efficient algorithms.

Strengths: The problem studied by the paper is natural and important, but for some reason, did not receive much attention from the research community. All the papers that refer to this problem do so only as an afterthought, while focusing mainly on a different setting. Thus, it is nice to see a paper that puts this problem in the center of the stage, and therefore, is likely to encourage more work on this somewhat forgotten problem. It should also be mentioned that the empirical results of the paper present the algorithms suggested by it in a good light, but naturally can compare these algorithms only with the state-of-the-art. Since the program has been partially forgotten, this state-of-the-art is not too impressive, and thus, being better than it is, unfortunately, not very impressive.

Weaknesses: The paper is quite incremental with respect to the techniques. In other words, it basically presents the results one can almost immediately get by throwing standard techniques of submodular maximization at the problem considered. Hence, the value of the paper is mostly in promoting the study of its problem, and providing baseline results for it on which one can improve with future research. In addition, the paper seems to make one or two assumptions that are a bit hidden, and significantly weaken the result it presents for the adaptive setting (see the detailed comments section).

Correctness: Naturally, most of the proofs have been deferred to the additional material, which I only skimmed over. However, based on the explanations given in the paper itself, it looks like everything should work out.

Clarity: Yes. Reading the paper is quite enjoyable.

Relation to Prior Work: Yes, the the relation with the previous work is discussed to a sufficient degree.

Reproducibility: Yes

Additional Feedback: - The response was read. The main question is about the novelty of the techniques used by the paper. Following the response, I took a look at the supplement material to see if there is indeed anything surprising there. However, I unfortunately failed to find such a thing, and thus, I keep my score. - On Page 2, a 2-approximation for the monotone version of the problem is attributed to [45]. I did not find in [45] such a claim, and moreover, the algorithm described does not have such a good approximation ratio (it provides about 2.58-approximation). - Section 2 is the first place mentioning the assumption in the adaptive setting that the objective function is a distribution over submodular functions. This assumption should appear already in the statement of the results. Furthermore, it is unclear from the description in Section 2 whether the paper also assumes that the objective function depends only on the state of the elements in the evaluated set. This should be clarified. Furthermore, if that is the case, then this assumption should also appear already in the statement of the results. - The experiments section assumes that the algorithm of the paper are run 5 times. The quality of the output in this mode of execution is interesting, but it would also be interesting to see the expected value of a single run and the standard deviation. Furthermore, this mode of execution does not seem to make sense in the adaptive setting. - According to the graphs in Section 5, FANTOM exhibit a super-linear dependence of the query complexity on the input size, while the algorithms of the current paper do not exhibit such a behavior. However, this seems to contradict the theoretical running time analysis presented in the paper, and thus, some discussion of this apparent contradiction will be in place. - The results of Greedy are omitted for some reason in Figure 1(b). This should be either fixed, or explained. - The conclusion expresses hope that the sampling technique will be used in the streaming settings in a future work. However, this was already done by a paper named “Do Less, Get More: Streaming Submodular Maximization with Subsampling” (NeurIPS 2018).

Review 3

Summary and Contributions: This paper is about maximizing a non-monotone submodular function subject to knapsack constraint. First of all, the paper proposes a randomized greedy algorithm which achieves a 5.83-approximation ratio and runs in O(nlogn) time (number of oracle evaluation). Although the approximation ratio is not state-of-the-art which is e=2.72, the running time is much faster and can be used in the practice. Secondly, it transfers the algorithm into a stochastic version in order to solve the adaptive submodular maximization, and obtains 9-approximation ratio. Furthermore, the paper gives the experimental comparison results in several instances like Video Recommendation, influence-and-Exploit Marketing, and Maximum Weighted Cut to show that the given algorithm is indeed practical and performs well. ==ADDED== In the response, the comparison about the ENV paper is clean. It is not practical compared to the result in this paper.

Strengths: 1. Provide a practical algorithm for non-monotone submodular maximization subject to knapsack constraint which has constant approximation ratio. The algorithm is quite simple and easy to implement. 2. Provide experimental results for their algorithm which proves that their algorithm can run well in practice. They also compare it with Fantom’s algorithm.

Weaknesses: 1. The approximation ratio is not good. 2. “To the best of our knowledge, nothing nontrivial is known for non-monotone functions and a knapsack constraint.”: There is a paper “Submodular Maximization with Matroid and Packing Constraints in Parallel” in STOC 19 which is not mentioned in this paper. In that paper, they provide a 1/e-ratio approximation algorithm for non-monotone submodular maximization under packing constraint while knapsack constraint is a special case of packing constraint. I believe the STOC paper has worse query complexity than this paper since it has heavy dependency over epsilon. But the STOC paper has very small (roughly O(log^2 n)) adaptive complexity (number of sequential rounds of independent value oracle calls). This means it is possible that this algorithm is practical in the sense that it may be feasible to run in large distributed system. Since that result considers more general constraint and matches the state-of-the-art approximation ratio which is much better than the ratio in this paper, it is worth to compare the results with that result, both in query complexity and in adaptive complexity.

Correctness: In my understanding, the theoretical results are correct.

Clarity: Overall, the structure of the paper is clear, and allows readers to clearly understand the author's intentions. One point that may cause misunderstanding is the definition of adaptive submodular in the title. The adaptive submodular here is based on the stochastic of element acquisition, not the adaptive complexity (number of sequential rounds of independent value oracle calls) in the literature. Some comments: 1. Line 51: it should be “density greedy algorithm”. 2. Line 251: “n=5 iterations”, but in the other part in the paper, n is used to represent the number of items.

Relation to Prior Work: The paper should compare their results with “Submodular Maximization with Matroid and Packing Constraints in Parallel” in STOC 19.

Reproducibility: Yes

Additional Feedback:

Review 4

Summary and Contributions: This paper considers non-monotone adaptive and non-adaptive submodular maximization under a knapsack constraint. The proposed algorithms achieve (3+2*sqrt(2))-approximation and 9-approximation for the non-adaptive and adaptive setting, respectively.

Strengths: The largest contribution of this paper is that it proposes a simple approach to non-monotone submodular maximization under a knapsack constraint. The current best approximation ratio 1/e was obtained by using the continuous greedy algorithm, which is not so practical. This paper proposes a practical, easy-to-implement greedy algorithm based on the sampling greedy algorithm by Feldman, Harshaw, and Karbasi (2018), whose approximation ratio (3+2*sqrt(2)) \approx 5.8 is best among the existing combinatorial algorithms. Moreover, the authors apply a similar approach to the setting of adaptive submodular maximization with a slightly worse approximation guarantee, which is the first algorithm for this problem. This paper seems to be a useful contribution to submodular maximization research.

Weaknesses: In my opinion, a weakness of this paper is its theoretical novelty. The analyses for the non-adaptive setting are a good combination of Feldman, Harshaw, and Karbasi (2018) and the standard technique for knapsack problems that takes the maximum of the largest singleton and the greedy solution. They are cleverly modified so that the adaptive setting can be dealt with, but there is an existing study (Gotovos, Karbasi, and Krause (2015)) on non-monotone adaptive submodular maximization that uses a similar approach as the authors mentioned. Both proof techniques are neat but not surprisingly novel. Another weakness is its practical side. To make the approximation guarantees hold, the sampling probability p should be set to the value specified in the theorems such as 0.41 or 1/6, but in the experiments, the authors set the sampling probability p in [0.9, 1], which make the proposed algorithm quite similar to the naive greedy algorithm. The authors claim that the proposed algorithms empirically perform well, but it is difficult to say the theoretical results are supported by the experimental results due to the artificial value of p.

Correctness: The technical content of the paper appears to be correct.

Clarity: The paper is generally well-organized and clearly written.

Relation to Prior Work: This paper mentions sufficiently many existing studies.

Reproducibility: Yes

Additional Feedback: In line 49 of the appendix, S \cup_{r=j*1}^\ell Q_r might be replaced with S \cup \bigcup_{r=j*1}^\ell Q_r. In line 180 of the appendix, a right parenthesis is missed. # Update after the author response I read the author response. The authors responded to each of my two concerns. (1) On technical novelty. Though applying the sub-sampling technique to the non-monotone submodular maximization under a knapsack constraint is a nice contribution, I don't think this paper develops a technically new idea. (2) On practical performance. The authors claim that they can make an artificial example in which small p works, but small p does not work well for practical objective functions. However, to show the practical advantage of the proposed algorithm, the authors should find a practical example which p smaller than 1 works. Due to these reasons, I keep my score.