NeurIPS 2020

Online Algorithm for Unsupervised Sequential Selection with Contextual Information

Review 1

Summary and Contributions: In this paper the authors propose an extension of the Unsupervised Sequential Selection problem to contextual input. They introduce the notion of Contextual Weak Dominance, and show that it is sufficient to solve the problem. They propose an algorithm, USS-PD, and show that it is able to achieve logarithmic regret under the additional hypothesis xi >0. Finally, they evaluate the performance of USS-PD on both real and synthetic datasets.

Strengths: Overall I think that this paper presents some significant contributions to USS : the authors prove that their condition, the contextual weak dominance, is sufficient to identify the best arm, and that in this case it is sufficient to estimate the disagreement between arms to achieve this objective. Their analysis of the regret of USS-PD is interesting and the empirical evaluation is convincing.

Weaknesses: - The formulation of some results are mildly confusing : for instance the result of Lemma 1 should be equality, and does not require j > i, as shown by the authors in their proof. It is unclear to me why the authors chose to only report a weaker result. - Technically it is possible to satisfy CWD and have \inf ( \xi) = 0. I think the paper could benefit from some insights of the authors regarding this particular case. - While this setting, Contextual USS, is clearly new (to my knowledge), most of the tools and results presented in this article are derived from existing results in USS and GLM. However, their extension to the contextual USS setting appears to be non trivial to me, so this work contribution is significant.

Correctness: All claims and theoretical results presented are proven, and the empirical methodology is correct.

Clarity: The paper is well written, and the proofs are detailed enough to be easy to follow.

Relation to Prior Work: The novelty of this work is clearly discussed in the paper.

Reproducibility: Yes

Additional Feedback:

Review 2

Summary and Contributions: [Please find the update in the "Additional feedback section"] The authors study a problem of selecting a classifier for making predictions from the collection of classifiers with various cost and reliability, called "Contextual Unsupervised Sequential Selection". This work elaborates on the recent study of Unsupervised Sequential Selection (without taking into account contexts). To model context-conditional correlations between arms the authors adopt the idea of GLM bandits.

Strengths: The paper is well-organized and the problem is clearly stated. The work responds to the call from the prior art regarding studying contextual version of the USS problem, so it covers a blank space in this research area.

Weaknesses: The idea might look incremental: what are the main challenges solved by this work apart the combination of USS and GLM bandits? The main weakness is experimental section that lacks comparison with baselines and other SOTA methods.

Correctness: The methodology is correct in general, but I have not checked math. details

Clarity: The paper is well written.

Relation to Prior Work: The prior art is clearly discussed.

Reproducibility: No

Additional Feedback: Update after rebuttal: Based from the authors feedback and other reviews I can see the paper is not that straightforward from the mathematical viewpoint, but I am not a deep specialist in bandits theory to acknowledge these contributions. However, I still think the paper lacks experiments. "To the best of our knowledge, we are first to consider the contextual USS problem, so there is no state-of-the-art (SOTA)" > There are methods for non-contextual USS problem, but none of them was considered as a baseline. Only weak baselines were provided. Therefore I will slightly increase the overall score of this submission.

Review 3

Summary and Contributions: This paper studies Contextual Unsupervised Sequential Selection (USS) problem, which is a new variant of the stochastic contextual problem. Compared with vanilla-USS (Verma et al., 2019), the decision maker can observe an additional context during each round, which is a new setting not studied before. The goal of the problem is to learn a decision rule such that the total expected loss is minimized. A new algorithm with sub-linear regret under the assumpsion that the problem instance satisfies `Contextual Weak Dominance` is proposed.

Strengths: An algorithm suffers O(\log T) regret is proposed and it is validated by experiments on both synthetic and real datasets.

Weaknesses: No lower bound is presented.

Correctness: I am not an expert in this area. From the presentation, the claims look reasonable to me.

Clarity: The paper is well written and the theorems are clearly stated.

Relation to Prior Work: This paper clearly states the difference between previous contributions.

Reproducibility: Yes

Additional Feedback: Overall the paper is well written with good presentation. People in multi-armed bandits fieid may be interested in this problem. One drawback of the paper is that no lower bound is provided and it is hard to tell whether the variables hidden from big-O in corollary 1 can be optimized. In equation (2), \lambda_{I_t} is missing and it is explained in line 126 why \lambda_{I_t} is missing. It is better if it can be put before equation (2). Typos: Line 85: generate -> generates Line 89: denote -> denotes ==== After rebuttal ==== The author feedback and the reviews of the other reviewers did not change my positive score about the paper.

Review 4

Summary and Contributions: In this paper, the contextual unsupervised sequential selection problem is studied. Different from the classical contextual bandit algorithm, the learner cannot observe the losses when the arms are pulled. Instead, an ordered relationship is assumed among the arms which are named the contextual weak dominance property. Relying on this assumption, the true loss can be equivalently reduced to the disagreement among arms. A context-based learning algorithm is proposed under this condition, whose theoretical guarantee is proved and empirical performance is verified through experiments.

Strengths: - The paper studies a new setting for contextual bandit learning, which is a topic with theoretical and practical importance. - The theoretical results are grounded with rigorous analysis. The proposed approach is verified through proper experimental results.

Weaknesses: - The contribution seems to be incremental comparing to the previous works (see detailed discussions below)

Correctness: As far as I can see, the claims and methods are correct both in theoretical and algorithmic perspectives.

Clarity: The paper is overall clearly written. The contributions are clearly described, and the analysis and algorithms are detailed introduced.

Relation to Prior Work: As far as I can see, the literature survey is thorough, and the relationships to previous studies are clearly explained.

Reproducibility: Yes

Additional Feedback: - The main concern for me is the technical novelty comparing to previous studies, especially [1]. In [1], the exactly same weak dominance property is also introduced, and the Theorem 1 in [1] is the same to Theorem 1 in the paper. By this theorem, extending classical GLM algorithm to the current setting seems to be somehow straightforward. I think it is useful to discuss the main technical challenges the proposed approach overcomes in the paper. - To make the setting more novel to previous study, it may be interesting to refine the assumptions on models to consider new constraints. For example, to consider the setting in which the context received in each round contain private information which needs to be protected. - I wonder if the feedback y can be real values instead of only binary values. This can make the setting more general. [1] Online Algorithm for Unsupervised Sensor Selection, AISTATS’19. --- after rebuttal My concern lies in the technical significance of extending the USS problem to the GLB setting. After reading all the reviews and the author feedback, I still find the assumption similar to the previous work as pointed out in my review. Even though I do think the paper studies an important real problem, I keep my score unchanged.