Learning-Augmented Dynamic Submodular Maximization

In dynamic submodular maximization, the goal is to maintain a high-value solution over a sequence of element insertions and deletions with a fast update time. Motivated by large-scale applications and the fact that dynamic data often exhibits patterns, we ask the following question: can predictions be used to accelerate the update time of dynamic submodular maximization algorithms? We consider the model for dynamic algorithms with predictions where predictions regarding the insertion and deletion times of elements can be used for preprocessing. Our main result is an algorithm with an $O(poly(\log \eta, \log w, \log k))$ amortized update time over the sequence of updates that achieves a $1/2 - \epsilon$ approximation in expectation for dynamic monotone submodular maximization under a cardinality constraint $k$, where the prediction error $\eta$ is the number of elements that are not inserted and deleted within $w$ time steps of their predicted insertion and deletion times. This amortized update time is independent of the length of the stream and instead depends on the prediction error.

Paper

Similar papers

Peer review

Reviewer a2JE7/10 · confidence 3/52024-07-12

Summary

Authors consider submodular maximization with cardinality constraint in dynamic setting: Algorithm sees a series of $n$ insertions and deletions of elements and has to maintain a subset of active elements maximizing given submodular function. Best known algorithms for this problem by Lattanzi et al., Monemizadeh, and Banihashem et al., achieve approximation ratio 0.5-eps with amortized update time polylogarithmic in $n$ and $k$, where $n$ is the length of the input stream and $k$ denotes the cardinality constraint. In this paper, authors propose an algorithm which receives predictions about insertion and deletion time of each element beforehand. This allows the algorithm to precompute an update strategy assuming that the predictions are correct. Given a parameter $w$, authors define the prediction error $\eta$ as the number of elements whose predicted insertion or deletion time is not within $w$ time steps from the real one. Their main result is an algorithm with approximation ratio 0.5-eps whose amortized update time is polylogarithmic in $\eta$, $w$, and $k$. With $\eta = o(n)$, their algorithm achieves an assymptotic improvement over the existing algorithms which do not use predictions.

Strengths

* result seems strong and requires introduction of new ideas as well as proving new properties about existing algorithms

Weaknesses

* the predictions used are quite verbose: having all predictions ahead of time is quite a restrictive requirement. Authors pose utilization of predictions which come one by one as an open problem.

Questions

* where does dependence on epsilon appear in your bounds in Theorem? * is the algorithm of Lattanzi the only existing algorithm which is threshold-based, or is this property more common among existing techniques?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Clearly explained.

Reviewer T4wq6/10 · confidence 4/52024-07-12

Summary

The authors studied monotone submodular maximization under a cardinality constraint in a dynamic model where predictions of insertions and deletions are given. In submodular maximization, a ground set of elements and a function assign a value to any subset of these elements. A function is submodular if adding an element to a smaller set contributes more value than adding it to a larger set. A function is monotone if its value always increases with the set size. In the dynamic model, elements are inserted and removed, and only elements that are inserted but not deleted can be selected. The error $\eta$ is defined as the number of elements whose actual insertion or deletion time differs from the prediction by at least ww. The goal is to find, after each update, a subset of size at most k from the current elements that maximizes the value of the monotone submodular function while minimizing the number of query calls. Previous works on dynamic monotone submodular maximization under cardinality constraints (without predictions) achieve a 1/2 approximation factor using O(polylog(n)) or O(k⋅polylog(k)) query calls. There are also works on dynamic monotone submodular maximization with predictions under matroid constraints, which is a generalization of cardinality constraints, achieving a ∼0.32 approximation factor. In this paper, the authors design an algorithm for dynamic monotone submodular maximization under cardinality constraints using predictions that achieve a 1/2 approximation factor with only $O(poly(log(\eta),log(w),log(k)))$. When the predictions are poor, the algorithm’s complexity is as bad as the algorithm without predictions, having O(polylog(n)) query complexity since $\eta$ would become $\theta(n)$. They use precomputation to compute a solution based on predictions for each time and leverage previous works on the dynamic model and delete-robustness model to update their solution according to real insertions and deletions. They also run experiments, comparing their query complexity with the dynamic algorithm. The results show that when prediction errors are small, their algorithm requires significantly fewer query calls. However, when the predictions are poor, the query complexity is comparable to that of the dynamic algorithm.

Strengths

Recently, many learning-augmented algorithms have been developed for dynamic and online models. They are interesting because, in real applications, we sometimes have an idea of how elements will change, raising the question of how we can improve our algorithms by leveraging these predictions. In this work, the authors improved the query complexity, eliminating the dependency on n and reducing the dependency on k to only logarithmically. It is interesting how query complexity can be improved using predictions, and it would be even more interesting if the approximation factor could also be improved.

Weaknesses

There were different algorithms for this problem without using predictions: one using O(polylog(n)) (first algorithm) and the other using O(k⋅polylog(k)) (second algorithm) query calls. Since the authors' algorithm has polylog(k) in its query complexity, if k is as large as $\theta(n^\epsilon)$ where $\epsilon$ is a constant, their algorithm is not better than the first algorithm. If k is small, the second algorithm is not that bad. Still, in the second case, the authors' algorithm is slightly better, but we should note that they also have O(n⋅polylog(n)) for precomputation. ## Comments for the authors: - Line 37: cited [30] twice - Some sentences are too long and make it hard to read. For example, look at the sentences from lines 60 to 63, and the next sentence.

Questions

I see that you mentioned Chen and Peng [10] proved that a dynamic algorithm for this problem needs poly(n)poly(n) query complexity, but does their approach apply when we have good predictions? I understand this question may be difficult and don’t expect the authors to answer it, but if they can answer affirmatively, it makes their result much stronger.

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

N/A

Reviewer A7Fx8/10 · confidence 4/52024-07-13

Summary

The paper studies the monotone dynamic submodular maximization problem under a cardinality constraint $k$ in the framework of algorithms with predictions. The authors consider a prediction model where the insert and delete times of the elements are predicted at time 0, and for any window size $w$, the prediction error $\eta$ is defined to be the number of elements whose actual insertion or deletion times differ from the predicted insertion or deletion times by more than $w$. Their main result is an algorithm that produces a $(1/2-\epsilon)$-approximate solution at each time step with expected amortized update time $O(\text{poly}(\log \eta, \log w, \log k))$ and preprocessing time $\tilde{O}(n)$.

Strengths

* The problem studied is important and interesting to the NeurIPS community * The result is strong, and the prediction model and the notion of error are natural. The algorithm is robust in the sense that it can handle an arbitrary number of elements with low prediction error (that fall within the $w$ window of their actual insertion/deletion time) and a reasonable number of elements with high error (whose predictions are off by more than $w$, counted by $\eta$). Moreover, their algorithm can handle elements that are not predicted to arrive but actually show up in the input sequence (these elements contribute to $\eta$). The performance of the algorithm also degrades gracefully as the prediction error increases. * While the subject is inherently complicated as there are lots of parameters involved, the authors did a good job keeping everything clear and precise so that it is relatively easy to follow. The notation is also good. Moreover, the warm-up algorithm helps the reader understand some of the main ideas of the final algorithm.

Weaknesses

* In the case where $k = o(\log n)$ and the prediction error $\eta = \Omega(n)$, the update time of the algorithm is worse than the update time $O(k \cdot \text{polylog}(k))$ achieved in reference [7]. Thus, in some cases, the algorithm performs worse than a worst-case algorithm without predictions.

Questions

* Why is not the algorithm with update time $O(k \cdot \text{polylog}(k))$ presented in [7] compared to your learning-augmented algorithm theoretically or empirically?

Rating

8

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors state their theoretical results formally, describing all assumptions.

Reviewer L8zN4/10 · confidence 3/52024-07-15

Summary

This paper studied the dynamic submodular maximization problem with predictions. The goal of the problem is to maintain a high-quality solution in the presence of insertions and deletions. The main contribution is leveraging predictions, in the form of the pattern of insertions and deletions, to accelerate the update time of dynamic submodular maximization algorithms.

Strengths

1. The paper is well-written and easy to read. The proposed solution and analysis, especially the connection to the robust submodular maximization problem, are technically sound. 2. The results improve upon existing ones when the prediction is accurate.

Weaknesses

1. The paper does not provide enough details on how predictions are obtained in real life. For example, it lacks specific information about machine learning algorithms that could potentially be used to make the predictions. This is concerning because the current notation of prediction error is defined in a worst-case manner. If there exists one prediction with poor accuracy, it might dramatically hurt the results. To fully benefit from their results, a very robust and stable prediction algorithm is required. 2. The tightness of their results is unclear. As mentioned in the introduction, if the prediction is 100% accurate, there exists a trivial algorithm that requires constant update time. However, even with zero prediction error, their bound does not reduce to a constant. 3. A more natural benefit of predictions is achieving an enhanced approximation ratio. While the authors mention existing studies along this direction, it is worth exploring whether predictions can improve the approximation ratio, using the same update time as algorithms without predictions. 4. Generally, incorporating predictions often sacrifices worst-case performance, which is especially true if the prediction is inaccurate. It would be beneficial for the authors to comment on this, specifically addressing to what extent bad predictions might hurt the update time of their algorithms. 5. The last comment is more of a clarification question instead of a comment. In the dynamic setting, why not simply use the state-of-the-art offline algorithm, such as the classic greedy algorithm, to solve the maximization problem in each round? I understand that this might incur higher update time, but if we consider a special case of the dynamic setting where there is only one round, it seems that the update time cannot be lower than the state-of-the-art offline algorithm. Is there any assumption regarding the length of the time horizon, for example, the number of rounds must be at least O(n)?

Questions

Please refer to the weakness section.

Rating

4

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Authorsrebuttal2024-08-12

Dear reviewer L8zN, Thank you again for your helpful suggestions. Since the author-reviewer discussion period is almost over, we wanted to see if our response addressed all your concerns. We would be happy to answer any follow-up questions you might have.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC