Blocked Collaborative Bandits: Online Collaborative Filtering with Per-Item Budget Constraints

We consider the problem of \emph{blocked} collaborative bandits where there are multiple users, each with an associated multi-armed bandit problem. These users are grouped into \emph{latent} clusters such that the mean reward vectors of users within the same cluster are identical. Our goal is to design algorithms that maximize the cumulative reward accrued by all the users over time, under the \emph{constraint} that no arm of a user is pulled more than $\mathsf{B}$ times. This problem has been originally considered by \cite{Bresler:2014}, and designing regret-optimal algorithms for it has since remained an open problem. In this work, we propose an algorithm called \texttt{B-LATTICE} (Blocked Latent bAndiTs via maTrIx ComplEtion) that collaborates across users, while simultaneously satisfying the budget constraints, to maximize their cumulative rewards. Theoretically, under certain reasonable assumptions on the latent structure, with $\mathsf{M}$ users, $\mathsf{N}$ arms, $\mathsf{T}$ rounds per user, and $\mathsf{C}=O(1)$ latent clusters, \texttt{B-LATTICE} achieves a per-user regret of $\widetilde{O}(\sqrt{\mathsf{T}(1 + \mathsf{N}\mathsf{M}^{-1})}$ under a budget constraint of $\mathsf{B}=\Theta(\log \mathsf{T})$. These are the first sub-linear regret bounds for this problem, and match the minimax regret bounds when $\mathsf{B}=\mathsf{T}$. Empirically, we demonstrate that our algorithm has superior performance over baselines even when $\mathsf{B}=1$. \texttt{B-LATTICE} runs in phases where in each phase it clusters users into groups and collaborates across users within a group to quickly learn their reward models.

Paper

References (27)

Scroll for more · 15 remaining

Similar papers

Peer review

Reviewer HmVS6/10 · confidence 3/52023-07-05

Summary

This paper considers the online recommendation for a hard-clustered user. Problem setup summary: This problem involves $M$ users and $N$ items. the reward matrix $P \in R^{M \times N}$ is unknown. There are $T$ rounds in total, and each user receives a recommendation $\rho_u(t)$ at each round. Each user belongs to one of (unknown) clusters indexed by $\{1,2,\dots,C\}$. Each user can consume at most $B = O(\log T)$ same item. The oracle-optimal strategy is to recommend top-($T/B$) items for $B$ (max possible) times, and the regret is the loss of recommending items except for these top arms. This is multitask learning in the sense that the results are better than solving $N$-armed bandit for each of $M$ users (in that case, regret would be linear to $M$). Section 2.1 introduces a bound on the accuracy of low-rank matrix completion (Lemma 1, [16]). Based on that it describes the performance of explore-then-commit with $B=1$. Section 3 introduced the B-Lattice algorithm, its regret bound (Thm 1), and regret lower bound (Thm 2) The paper proposes B-LATTICE algorithm for online recommendation. The algorithms is phased and at each phase $l$ it obtains accuracy of $2^{-l}$. This paper is well-written. Although the algorithm described here is not truly practical in the sense it can be applied to real recommendation (see weaknesses). The contributions of the paper are enough good. My main concerns are about the solidness of the results. ##### I think the paper is above the acceptance threshold even though the exposition can have some space for improvement, so I keep current rating (WA).

Strengths

* A new scheme on the factorized recommendation with bandit settings (Tbl 1) * parametric rate of regret * Non-trivial (non-tight) regret lower bound using Fano's inequality. * Simulation results in appendix

Weaknesses

* Somewhat artificial budget constraint (top T/B items, where B is consistent over all items) * Many parts of the algorithms are in the appendix (mainly, Alg 4,5) * Alg depends on B, C that are hard-to-know before (in particular B). * Alg is very involved. * (minor) exposition on the order of elements can be improved. * For example, Lemma 1 informally describes incoherency before the main results, the main results part (Thm 1) do not include B-LATTICE detail even though it is bound on this algorithm.

Questions

* Is Lemma 1 derived from Lemma 2 of [16] (or from Lemma 5 of [16])? * Question on the applicability of Lemma1: Lemma 1 requires the sampling of $\Omega$ to be iid. I assume B-Lattice adopts the same bound, but can it be usable for such a complex sampling scheme? * Questions on the bounds: * Can the authors clarify the incoherency condition (L176: $||\bar{U}||_{2,\infty}$ - this norm undefined?), and subset strong smoothness? * What is the dependence of Thm1 in $C$ (# of clusters)? * What $B = \Theta(\log T)$ exactly states? In particular, can i be $C \log T$ where $C>0$ is independent of $T,N,M$, and # of clusters? * (minor) Are the technology used in this paper (bound on low-rank matrix completion and graphs) entirely new in this paper? How the basic technology can be compared with related work (e.g., Table 1)? * (minor) Is there any way to simplify the algorithm? For example, can explore (Alg 2) be randomly recommending an available item, and exploit be randomly recommending an item weighted on the likelihood to be good?

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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

Not applied.

Reviewer 7P267/10 · confidence 4/52023-07-06

Summary

In this paper, the authors tackle a collaborative-filtering type multi-armed bandit problem. There are many users, each facing a multi-armed bandit problem. However, there are only $C$ base multi-armed bandit instances, where each user comes from one of these instances. The authors provide a novel algorithm which only pulls each arm at most $B$ times per user, and provide information theoretic lower bounds for this new setting. The authors also provide an improved, practical version of this method.

Strengths

1. The proposed algorithm is reasonably easy to follow, and makes intuitive sense. The authors explain the process clearly. 2. The authors provide a new lower bound, using a different Fano's inequality based method. The approach and proof of the upper bound make sense, but I have not checked the details thoroughly. 3. It would be good to emphasize that $N,M\gg T$ is actually a very relevant setting; $T$ is number of interactions per user, each user will only interact with the system a moderate number of times, but the overall number of users can be very large.

Weaknesses

1. The assumptions regarding $C$ are very strong. a) that all users come from exactly 1 of $C$ instances (no within group variability), b) that $C$ is exactly known, and c) that $\tau$ is constant (it is unclear how the bound depends on $\tau$). Additional text regarding why these assumptions are reasonable, the brittleness of the algorithm to these differences, or on how these assumptions can be loosened, would strengthen the paper. 2. Empirical improvement: it does not appear that the algorithm has much better performance than the greedy baselines, unless I am misreading Figure 1? It could improve the paper to add simulations for larger $T$ to show how the asymptotic behavior compares (over Greedy). Additionally, comparing against [24] in the case where $B=T$ would be interesting. Finally, to highlight the improvement of collaboration, it could help to include a baseline of running e.g. UCB independently for each user, showing that this would yields dramatically worse performance. 3. Clarity: the work is overall quite clear and well-written, but there are some points (discussed below) that could help improve presentation. Minor: 1. Lower bound as stated is not technically correct. Classical bandit lower bounds hold when the noise is truly Gaussian (or Bernoulli), and cannot be invoked for general sub-Gaussian random variables, as the variance proxy is an upper bound (e.g. a Gaussian with variance .01 is 1-sub-Gaussian). 2. Eqs after Line 1252 and 1259 have typing errors; maybe missing a $\geq$? Typos: Line 28: unknown cluster and *the* expected reward 53: suboptimal dependence on *the* number 207: "columns comprise of at" 249: "small for e.g." Fig 1 caption: period missing before last sentence, space after PB-Lattice. Redundancy in setting $\Delta_{\ell+1}$ in line 2 of Alg 1 Line 274,278: clarify that goal is to get an entrywise $O(2^{-\ell})$ estimate. B-Lattice doesn't take in B as input, and this parameter doesn't show up anywhere in Alg 1. Alg 2 also doesn't take $B$ as input. Using $\mathbf{C}$ as a counts matrix and $C$ as the number of clusters is confusing notation. Algorithmic clarification: 1. It appears as though $\mathcal{M}^{(\ell)}$ is an *estimation* of disjoint nice subsets of users, whereas in Algorithm 1's input it is stated that they are truly nice subsets of users. 2. Line 287: each connected component is a nice subset *with high probability*? Several results need to be clarified that they're conditional on the good event in Lemma 1 (that estimation worked). 3. The algorithms defined do not appear to be "stand-alone", i.e. they have side effects involving the matrices C and D (which appear to be global variables)? A clearer description of what each algorithms requires as input, and what variables it modifies, would be helpful.

Questions

See weaknesses. My primary question is regarding the assumption of knowing $C$, and that each user has exactly 1 of $C$ different mean vectors.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

Discussed, modulo $C$ above.

Reviewer 2VtF5/10 · confidence 2/52023-07-06

Summary

The paper considers the problem of blocked collaborative bandits with multiple users that each belong to a cluster with associated bandit instance. The goal is to minimize the regret of each user while satisfying the constraint that each arm can be played at most B times for each user. The paper proposes an algorithm with sublinear regret upper bound.

Strengths

- The paper proposes a novel algorithm that uses matrix completion and clustering techniques. - Although not very surprising, a regret bound of O(sqrt{TN/M}) is proved when B=O(log T). - The paper also proves a lower bound on the regret.

Weaknesses

- The regret upper bound holds only for B=O(log T). As the papers main contribution is providing regret bounds with the existence of the budget constraint B, I believe the authors should provide the dependency of the regret bound on B. - Assumption 1 is strong. It requires the information to not be concentrated in some clusters which in my opinion is an interesting case to deal with. Moreover, the choice of assumptions and constants is unnatural; for example why is |S|=T^{1/3} in Assumption 1.3? - The regret lower bound does not match the achieved upper bound.

Questions

- It seems that the regret lower bound in Theorem 2 is linear in T while the upper bound is sublinear?

Rating

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

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

Yes.

Reviewer tct34/10 · confidence 3/52023-07-06

Summary

This paper studies the blocked collaborative bandits where each user represent a single bandit problem. The latent groups are assummed to exist among users in which all the users in a group share the same expected reward. But the constrant is placed on the number of serving time for each user to solve the challenge of lacking user-item interactions. Then, the authors introduce a complicated algorithms including the exploitation, exploration and clustering components. The theoretical analysis is provided and achieved a tigher regret bound for this setting.

Strengths

This paper provide two regret bound. The first one recover the results and follows the regret analysis is online clustering of bandits. The second authors provide a lower bound, which consider the case when T is large.

Weaknesses

The introduced algorithm is very complicated and I am worried about the efficiency of this algorithm. Authors didn't report the time cost of this algorithm and conduct ablation study for the exploration component. Because this paper tries to solve a very practical problem, the online recommendation, it only has very limited empirical evaluations, which run on the systhetic datasets. The main manuscript doesn't inlcude the empirical evaluation.

Questions

In this problem setting, it seems to not have the assumption regarding the reward gap between two clusters. The reward gap assumption is commen in the problem of online clustering of bandits. How does it influence the analysis?

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

More empirical evaluations are needed.

Reviewer tR2a7/10 · confidence 3/52023-07-12

Summary

This paper studies blocked collaborative bandits. Multiple users, clustered in a fixed number of clusters according to the rewards obtained from them are provided recommendations in each discrete time instance till T timesteps. The paper proposes algorithms that minimize the cumulative regret with the constraint that no user is recommended the same arm (item ) more than B times. The authors propose an algorithm called B-LATTICE that collaborates across users while satisfying budget constraints, and minimising regret. The authors show theoretical guarantees of the proposed algorithm and also validate their results empirically with synthetic data.

Strengths

The paper introduces an important setting with practical application. The main strength of the paper is its sublinear regret guarantee. The authors give the first sublinear regret algorithm in the proposed setting. Furthermore, the analysis and proposed algorithms are involved and nontrivial; the authors do an excellent job of providing an intuitive understanding of their proposed algorithms.

Weaknesses

The paper does a poor job of explaining the setting and its connection to Bandits. Following are my questions and (some) suggestions to the authors. 1. A crisp summary of how the problem is MAB problem is missing from the introduction/setting section. For instance, the following points are not clear in the first reading of the first 2 sections; what constitutes an arm (items?), who pulls arms (a recommendation engine or users), it is not clear that rewards are not only arm-specific but also user dependent (which in my opinion most important difference from classic MAB setting). 2. Can authors shed some light on what is known apriori to the algorithm/recommendation engine? For instance, the authors use the fact that the number of true clusters $C$ is known in Algorithm B-LATTICE, however, I don't understand why T is needed as an input. In particular. can the regret guarantee of the proposed algorithm be converted into an anytime guarantee (for any value of large enough T)? In contrast, in the Greedy algorithm (Remark 1) authors explicitly use T. 3. I don't see why the authors normalize the cumulative regret by $M$ (Equation 2). Is it crucial to consider per-user regret guarantee? Howe do the results fare without this normalization?

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

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

2 fair

Contribution

3 good

Limitations

none

Reviewer tR2a2023-08-10

Thank you for your detailed response to my questions. I maintain my original view on the acceptance of the paper.

Authorsrebuttal2023-08-13

Thanks!

We thank the reviewer for their positive assessment of our paper. Please let us know if there is anything else that we can clarify.

Reviewer 7P262023-08-10

Response to rebuttal (increase score to 7)

Thanks for the clear and well-written rebuttal. These responses address my main points, and I have increased my score to a 7. Below I respond to specific points: C: Interesting, does this "loose clustering" work off the shelf? Empirical improvement: You claim that "We believe a similar phenomenon happens in our setting (i.e., greedy with best exploration parameter is a very strong baseline with good regret guarantees)". Do you have any theoretical justification for this? If so it would be very interesting + informative to include. Additionally, regarding the new simulations: you state that "we use standard imputation techniques to fill those missing values". This seems potentially prone to "overfitting", as this means that the matrix is being filled in via an optimization procedure, and that later B-LATTICE observes a subset of entries and performs another (potentially related) optimization procedure to guess the missing values. Additional clarification regarding how the imputation was performed, and how to verify that the gains of B-LATTICE aren't due to a matching between the two procedures, would be helpful.

Authorsrebuttal2023-08-11

Thanks and Answers to Follow-up questions

Thanks so much for the positive assessment of our work. Below, we provide answers to follow-up questions: 1) **C:** Yes, the "loose" clustering is off the shelf. We can compute the estimate of C and use some multiple of that in our algorithm as input - as we mentioned before, a loose upper bound is sufficient for our algorithm. 2) **Do you have any theoretical justification for this?** - This is a great question! Note that if the sub-optimality gaps are known, we can definitely show theoretically that greedy can do much better! As the reviewer suggests, we will include this in the final version. 3) **Clarification regarding imputation of missing values** - We follow the standard procedure as outlined in Section 4 in https://arxiv.org/abs/1606.00119. Namely, we use Python package *fancyimpute* with the default settings. All the algorithms are completely agnostic to the process through which these matrices have been completed. We understand the reviewer's point regarding overfitting, But since the percentage of missing entries is very low compared to the available ground truth entries, there is not much chance of overfitting - this is the reason why for validation in real world experiments, we seek datasets with a very low percentage of missing values.

Reviewer tct32023-08-15

Thanks for authors' response and added experiments. (1) Experiments. Since the authors only compared the proposed algorithm with the greedy baseline, is there any other work as same as your problem setting? From my perspective, these works for online of clustering can be modified to the problem setting by adding the constraint. But, I understand, one week is too short to add more experiments. (2) Gap assumption. Authors discussed two cases with two different solutions. How can your algorithm differentiate these two cases if there is no gap assumption? (3) Theoretical analysis. For theorem 1, How does $C$ affect your regret bound? Because these is not $C$ in your regret bound, are you assuming $C =1$? For theorem 2, I don't find a special meaning of it when considering $C=1$. Because if $C=1$, we can regard all users as one "big" user because they all have identical expected rewards. Then, this analysis can be done by using the standard MAB. We are more interested in the case when $C > O(1)$ but $C << M$.

Authorsrebuttal2023-08-16

Further Responses

We thank the reviewer for following up with some nice questions. Below, we provide detailed answer to them: **Experiments.** To answer the reviewer's question, we compare with a strong baseline, an Alternating Minimization (AM) based algorithm introduced in [1] and modify the AM algorithm by adding the blocking constraint. Note that even without the blocking constraint, AM algorithm outperforms several baselines as reported in [1]. The AM algorithm is very general since it handles a low rank reward matrix (a special case of which is the cluster setting). In the same MovieLens setting described in the global response (with $M=200$ users, $N=200$ items and $T=60$ rounds), the modified AM algorithm, ran with the same parameters as reported in [1], obtains a Cumulative Regret of 6901.93. On the other hand, as shown in the plot, our algorithm obtains a Cumulative regret of 6543.91 while the greedy algorithm with 30 and 10 exploration rounds obtains a cumulative regret of 7302.19 and 7794.96 respectively. Thus our algorithm, outperforms the AM algorithm with the blocking constraint. However, we point out again that the AM algorithm does not have any theoretical guarantees at all (not even without the blocking constraint) while our algorithm has strong theoretical guarantees. [1] Alternating Linear Bandits for Online Matrix-Factorization Recommendation (2018) **Gap assumption. Authors discussed two cases with two different solutions. How can your algorithm differentiate these two cases if there is no gap assumption?** To get the intuition for how the algorithm can differentiate between the two cases, we first consider the phased elimination algorithm for standard Multi-armed Bandits (MAB) (see Exercise [6.8] in [2]). This algorithm doesn't require the knowledge of any gap parameters, but still achieves the optimal instance dependent regret. It does this by getting more accurate estimates of the arm rewards over phases and eliminating arms based on a pre-specified threshold that decreases exponentially with each phase. If there is huge gap in the rewards between the best arm and the rest, then the sub-optimal arms will be identified and eliminated in earlier phases. On the other hand, if the gap is small, then the sub-optimal arms will be eliminated in later phases. In both these cases, it can be shown that the algorithm achieves the optimal regret while being oblivious to the gap. Our algorithm is designed along similar lines- it runs in phases and in every phase, we estimate the user clusters and simultaneously eliminate arms within each cluster. In the beginning of the algorithm, these clusters are coarse. As the algorithm proceeds, we refine our clusters and eliminate arms based on a pre-specified threshold. If the gap between clusters is large, our algorithm can separate these clusters in earlier phases. On the other hand, if the gap between clusters is small, our algorithm can separate them in later phases where we have more data. In both these cases, we can show that our regret is minimax optimal, without actually knowing the distance between clusters. To be more precise, we have a graph based approach for clustering users. The graph based approach takes a pre-specificied parameter $\Delta_{\ell}$ (in phase $\ell$) that decreases exponentially with the phase index. Here the graph corresponds to a user-similarity graph constructed based on the estimates from our low rank matrix completion module. and subsequently, the clustering proceeds by simply considering each connected component of the user similarity graph as a cluster. If the clustering succeeds that is, the sets of users belonging to separate true clusters are identified, then we are good. But note that even if the clustering does not succeed, we show that it is because the entry-wise gap between the cluster centers must be small - i.e. the gap must be $O(\Delta_{\ell})$. Because of this, we can treat the users jointly as a single cluster of users. Hence, we will suffer bounded regret in the exploration component in the next phase. Note that we have not used any explicit knowledge of gap anywhere. Since $\Delta_{\ell}$ decreases exponentially, the regret decreases exponentially with the phase index $\ell$ (even if we are not able to distinguish between the true clusters of users). [2] https://tor-lattimore.com/downloads/book/book.pdf **Theoretical analysis. For theorem 1, How does $C$ affect your regret bound?** Of course, the setting of C=1 is trivial as the reviewer mentions. We have assumed $C$ to be a constant (i.e. it does not scale with M,N,T) - see Assumption 2. Our regret guarantee scales polynomially in $C$ (more precisely, the dependence is $O(C^3)$) - but for simplicity of results/exposition, we have hidden the dependence of $C$ within the $\widetilde{O}(\cdot)$ notation. However, if the reviewer suggests, we can make the dependence of $C$ explicit in our result.

Reviewer tct32023-08-17

Thanks for the response, but my concerns are not fully addressed. For Q2, there is no lemma that can show that the proposed algorithm can separate two closed clusters or the complexity of the number of rounds needed to separate them. For Q3, the related work such as (https://arxiv.org/pdf/1401.8257.pdf), depends on $\sqrt{C}$. $O(C^3)$ is too high. So, I keep my original assessment.

Authorsrebuttal2023-08-17

Response from Authors

**Q2 (Gap Assumption):** We believe there has been a misunderstanding. As we have mentioned before, our algorithm has the same intuitive idea as phased elimination algorithms for standard Multi-Armed bandits ( for the exact proof, we suggest the reviewer to first see Exercise 6.8 in [1]) that obtains the optimal regret without having the knowledge of any gap parameters. We emphasize again that our phased algorithm does not try to separate the true clusters - if the gap is large, the clusters are automatically separated in earlier phases. if the gap is small, users belonging to them continue to be treated as a single cluster until later phases and the regret will decrease with phases too. When phase index is ~ $\lceil log(1/gap)\rceil$ ($O(1/gap^2)$ rounds), the separation will also happen. Until this point, regret would only still be O(1/gap). This two case explanation is ONLY for the analysis and it does not figure in explicitly in the algorithm whose parameters are agnostic to the true gaps. Again, we emphasize that the high level idea is same as in phased elimination for MAB. For the exact proof, please see Corollary 2 in the Appendix where we explicitly demonstrate the entry-wise difference in rewards between two users in different clusters who have not been separated in a certain phase. Also see the detailed regret analysis in Appendix D.2 [1] https://tor-lattimore.com/downloads/book/book.pdf **Q3(Dependence on $C$)** - We emphasize that $C$ is usually a very small constant in practice. Similar poly(C) dependence have also been provided in other related works [2,3]. **Our goal was to first obtain the right dependencies on $M,N,T$ (users, items, rounds) which are the large quantities in our setting - this itself was an important open problem in our setting with blocking constraint.** Only, recently in [3], the optimal rates in $M,N,T$ were provided in our setting without the blocking constraint. We provide the first algorithm matching the rates in [3] in the blocked setting. However, theoretically speaking, we agree that improving the dependence on $C$ is important as future work. **Since the reviewer points to [4] for the dependence on $C$, we request them to please look at the detailed discussion of [4] given in Section 1.1 in [3].** We must highlight the fact that [4] solves the linear bandit problem with i.i.d contexts (without the blocking constraint). Yes, the results of [4] are applicable to our setting without the blocking constraint - however, their result gets a significantly sub-optimal dependence on $M,N$ itself (which are much larger than $C$). **In fact as discussed in [3], without the blocking constraint, [4] obtains a highly sub-optimal regret of $O(\sqrt{M^2CT}+M^3N)$** - the dependence on $M,N$ is highly sub-optimal. Again, we emphasize that getting the right dependence on $M,N,T$ is the first priority since they are much larger than $C$ . Finally, the dependence on $C$ in our results can be improved significantly using cliques instead of connected components in the algorithm (at the cost of a more complex analysis). To simplify the analysis and presentation in the paper, we did not try to optimize the dependence on $C$. We can add a remark regarding this in the paper. [2] https://arxiv.org/abs/1606.00119 [3] https://arxiv.org/abs/2301.07040 [4] https://arxiv.org/pdf/1401.8257.pdf

Reviewer HmVS2023-08-19

Thank you for your response

* Somewhat artificial budget constraints I mean, for example, why the budget is $O(\log T) = o(T^{0.01})$? (as well as the homogeneity of $B$ as replied) Is there any theoretical justification for this? The largest limitation I acknowledge is the complexity of the algorithm and the limited experimental results described in the appendix (and in the author here in response to other reviewers).

Authorsrebuttal2023-08-20

Response from Authors

We thank the reviewer for the response. We have provided our comments below: **Artificial Budget Constraint** Note that our analysis and techniques can extend to $B=\Omega(\log T)$ with some minor technical changes. However, our goal is to provide guarantees in the practically important regime of small $B$. Our theoretical guarantees for B-LATTICE holds for $B=\Theta(\log T)$- we have left the problem of extending our theoretical guarantees to $B=O(1)$ as a future work (see Remark 3 and Section 5 in paper). The $\log T$ term appears in the budget because of the number of phases (logarithmic in rounds) - we pull each arm at most once in each phase. Extending our guarantees to $B=o(\log T)$ requires handing dependent data in low rank matrix completion - this needs new techniques. Finally, we should point out that the greedy algorithm we propose (Algorithm 5 in Appendix B) works for any $B\ge 1$ (see Remark 1 in paper). **Complexity of Algorithm** Regarding complexity of our algorithm B-LATTICE, at its core, B-LATTICE has 3 components namely Exploration, User Clustering and Exploitation. These $3$ components are conceptually simple - please see the descriptions in Section 4. The main complexity comes from implementing these components in a recursive manner because of the sequential nature of the data. In the final version, we will provide more explanation regarding the implementation with more available space. **Experimental Results** Please note that we have provided detailed experiments on synthetic datasets in Appendix C in the paper. We have also provided experiments on the MovieLens dataset (real-world dataset) in the global response. We can also provide additional experiments on real datasets in the next version if the reviewer suggests so. However, we emphasize that the primary contributions of our paper are theoretical.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC