Adaptive Selective Sampling for Online Prediction with Experts

We consider online prediction of a binary sequence with expert advice. For this setting, we devise label-efficient forecasting algorithms, which use a selective sampling scheme that enables collecting much fewer labels than standard procedures, while still retaining optimal worst-case regret guarantees. These algorithms are based on exponentially weighted forecasters, suitable for settings with and without a perfect expert. For a scenario where one expert is strictly better than the others in expectation, we show that the label complexity of the label-efficient forecaster scales roughly as the square root of the number of rounds. Finally, we present numerical experiments empirically showing that the normalized regret of the label-efficient forecaster can asymptotically match known minimax rates for pool-based active learning, suggesting it can optimally adapt to benign settings.

Paper

Similar papers

Peer review

Reviewer naDi7/10 · confidence 4/52023-07-02

Summary

This paper presents an adaptive label-efficient forecasting technique for online binary prediction with expert advice. The proposed approach implements a label querying probability that is a function of the observed scenario, rather than based on pessimistic conditions. This enables the method to adapt, i.e., have lower label complexity, to benign environments while remaining robust to adversarial ones, unlike prior approaches in this label-efficient forecasting. Sharp analyses of the regret and label complexity and results on synthetic scenarios are provided in support of the method.

Strengths

* The paper is very well-written and organized. * The introduced algorithm is novel and intuitive. The regret and label complexity analyses of the approach seem sound. * The method remedies the shortcoming (lack of adaptivity) of prior label-efficient prediction approaches. This enables it to query fewer labels in benign scenarios while remaining robust to adversarial ones. * The authors present empirical evaluations that demonstrate the effectiveness of the method.

Weaknesses

* The method only applies to binary prediction tasks with zero-one loss. * Additional details on prior work on label-efficient prediction would be helpful in contextualizing the benefit (adaptivity) of the proposed approach.

Questions

1. The authors mention that the same inductive approach can be used for general loss functions and $y_t \in [0,1]$, but that the expression for the selective sampling probability would be really complicated. Do the authors conjecture that an efficiently-computable upper bound for the complicated function can be found as was done for the binary zero-one loss case? 2. The authors mention active learning as a possible application in the introduction. Could this method (or its generalization) be applied to, e.g., a simple active learning scenario with a ResNet on CIFAR10?

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Yes.

Reviewer G2NF8/10 · confidence 4/52023-07-02

Summary

The paper considers a binary prediction game on $0-1$ loss, it proposed efficient sampling scheme via an modification of the exponentiated weight forecaster, which selectively acquire labels $y_t$ based on $Ber(q_t)$, where the design of $q_t$ is correlated to the disagreement among experts’ predictions at each round for the exponentiated weight forecaster. Ultimately, the proposed algorithms achieves the regret of the exponentiated weight forecaster $O(\ln N \sqrt{ n} )$ over time horizon $n$ and $N$ number of experts, the design labelling acquisition parameter $q_t$ resultant to a labelling complexity of $O( \frac{\sqrt{n}}{\Delta^2} )$, where $0 < \Delta \le \mathbb{E} [\ell_{t,i} - \ell_{t,i^{\ast}}], \forall t \in [n], i \neq i^{\ast}$ represents the lower bound expected loss gap comparing to the optimal expert with index $i^{\ast}$, which also signifies the difficulty of identifying the optimal experts.

Strengths

The paper is well written and easy to follow. It first introduces the intuition of sampling parameter $q_t$ if there is a perfect expert, then extended to general case. The paper also provided an graphical illustration on the lower bound of $q_t$ which matches with expectation. The result is novel comparing to previous result in two folds. (1) There is no assumption on how $y_t$ is generated. The proposed algorithm is able to attain the $O(\ln N \sqrt{n})$ regret without less labels. (2) In contrast to previous work on sampling by disagreement, the label complexity can be quantified as $O(\frac{\sqrt{n}}{ \Delta^2})$. $q_t$ is easy to compute, numerical experiment with respect to time horizon $n$ shows the expected regret and expected number of labels which matches with theoretical results. Experiment with respect to number of label (number of weights update) shows labelling efficient algorithm proposed in the paper matches the minimax rate in active learning asymptotically.

Weaknesses

It seems that the assumption $0 > \Delta \ge \mathbb{E} [\ell_{t,i} - \ell_{t,i^{\ast}}], \forall t \in [n], i \neq i^{\ast}$ is required only for bounding the labelling complexity, ( in order to track how $q_{t+1}$ evolves in line 493), without this assumption the regret still holds. I am a bit concern such assumption is very strong, it generally asks the best expert $i^{\ast}$ is wining over every other expert at every round. In addition, the assumption is $\ell_{t,i} \in \{ 0, 1\}$, if we are at a specific iteration $t$ where $\ell_{t,i^{\ast}} = 1$, then what values can $\ell_{t,i}, \forall i \neq i^{\ast}$ take in order to satisfy the condition for a strictly positive $\Delta$? Whether labelling complexity can be bounded without this assumption?

Questions

Appendix C: From line 437 to line 438, whether intermediate steps could be further elaborated. I am concerned about the proof, which seems primarily confused on the relationship of combining minimum values: $\min_i (a_i + b_i)$ VS $\min_i(a_i)+ \min(b_i)$. My understanding is that from the last equation line of line 437, we have $$ L_{n |1}^{w_{.,0} } \le \frac{n \eta }{ 8 } \underbrace{ - \frac{\eta}{8 } + A_{1,1} + (1 - 2 A_{1,1})y_1 + \mathbb{E} \left[ \min_{i} \left( L_{i, n} + \frac{\ln w_{i,0}}{\eta} + \frac{q_1}{\eta} ( \ln w_{i,0} - \ln w_{i,1} ) - \ell_{i,1} \right) \right]}_{A} $$ To show the desired result claimed in the theorem, we aim to show $$ A \le \mathbb{E} \left[ \min_{i} \left( L_{i,n} - \frac{\ln w_{i,0}}{\eta} \right) \right] $$ That is to show $$ A_{1,1} + (1 - 2 A_{1,1})y_1 + \mathbb{E} \left[ \min_{i} \left( L_{i, n} + \frac{\ln w_{i,0}}{\eta} + \frac{q_1}{\eta} ( \ln w_{i,0} - \ln w_{i,1} ) - \ell_{i,1} \right) \right] \le \frac{\eta}{8 } + \mathbb{E} \left[ \min_{i} \left( L_{i,n} - \frac{\ln w_{i,0}}{\eta} \right) \right] $$ From the claim in line 438, it seems we need to show above inequality should hold even without expectation, that is: $$ A_{1,1} + (1 - 2 A_{1,1})y_1 + \min_{i} \left( L_{i, n} + \frac{\ln w_{i,0}}{\eta} + \frac{q_1}{\eta} ( \ln w_{i,0} - \ln w_{i,1} ) - \ell_{i,1} \right) - \min_{i} \left( L_{i,n} - \frac{\ln w_{i,0}}{\eta} \right) \le \frac{\eta}{8 } $$ It seems the paper used $$\min_{i} \left( L_{i, n} + \frac{\ln w_{i,0}}{\eta} + \frac{q_1}{\eta} ( \ln w_{i,0} - \ln w_{i,1} ) - \ell_{i,1} \right) - \min_{i} \left( L_{i,n} - \frac{\ln w_{i,0}}{\eta} \right) \le \min_{i} \left( \frac{q_1}{\eta} ( \ln w_{i,0} - \ln w_{i,1} ) - \ell_{i,1} \right) $$ which doesn't seem to be true. These intermediate steps are just my personal guess. I wonder whether details to understand those two lines in the manuscript can be further explained.

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

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 VgJf8/10 · confidence 4/52023-07-07

Summary

This paper proposes an interesting novel approach to prediction with expert advise. In the standard prediction with expert advise setup, the learner receives experts' predictions, commits to its own and then sees the true outcome as produces by the (possibly adversarial) nature. Suppose that obtaining the true outcome is costly; do we really need to do this all the time? Clearly, if all (or most) experts agreed on the same prediction, the value of the true outcome for adjusting our trust in them is negligible; in the standard weight-based algorithms with multiplicative update the contribution of this round will simply be eliminated by normalisation. The paper takes an algorithm from Cesa-Bianchi and Lugosi (exponential weighting with fixed $\eta$) and shows that its regret bound stands as it is if the true outcome is requested with certain probability. It is then shown that the expectation of the number of outcomes actually requested is upper bounded by $3\eta T + O(\log\log T)$, so for small $\eta$ there is linear improvement in the number of requested outcomes. If $\eta$ is chosen to minimise the regret bound using prior knowledge of $T$, we take $\eta\propto 1/\sqrt{T}$ and thus the bound reduces to $O(\sqrt{T})$. This results holds under some conditions on experts' behaviour though and they seem restrictive.

Strengths

I think this is an interesting take on the well-known problem and should be published.

Weaknesses

No obvious weaknesses.

Questions

The authors chose to present the results in the probabilistic setup. Typically predicting $0$ or $1$ in a probabilistic setup is equivalent to predicting $\gamma\in [0,1]$ under a deterministic setup; the later seems more important to me (this is my judgement of course). Will the results of the paper stand? I think the final version would benefit from a better discussion of the $O(T)$ lower bound, which is mentioned but not discussed in detail.

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

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

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Yes

Reviewer VgJf2023-08-11

Many thanks to the authors for the response. I do think it should be straightforward to extend the result to \[0,1\], but regardless of this I am happy to keep my high evaluation of the paper.

Authorsrebuttal2023-08-19

Thank you for your response! After thinking a bit more about this, we agree that it is straightforward to extend our results to expert predictions in [0,1], while still keeping the label sequence and forecaster predictions binary (in {0,1}). Extending to more general forecaster predictions and losses does not seem quite as straightforward, although it may be doable with similar techniques.

Reviewer 1FCH6/10 · confidence 3/52023-07-08

Summary

This paper investigates the PEA problem in the context of online binary classification where the cost of obtaining labels for streaming data is high, necessitating selective label collection adaptively. To this end, the authors introduce a carefully designed label collection strategy based on the classical Hedge algorithm. The resultant label-efficient forecaster has a best-of-both-worlds theoretical guarantee. The authors further demonstrate that regret of their label-efficient forecaster asymptotically reaches the minimax rates of pool-based active learning.

Strengths

* The paper is well-structured and easy to comprehend. * The theoretical analysis provided in the paper appears to be solid and sound.

Weaknesses

* The primary conclusion of the paper, Theorem 3, relies on a core assumption that there exists a unique optimal expert whose expected loss in each round surpasses all other experts by a specific margin $\Delta$. Is this assumption too strong? Are there real-world application scenarios that satisfy such an assumption? (To my knowledge, this assumption has only been used in the COLT'14 paper: A Second-order Bound with Excess Losses). * Does this specific PEA setting investigated in the paper relates to the bandit setting? Both are concerned with identifying the optimal expert (arm). How then do the setting and techniques used in this paper differ or relate to those in bandit scenarios? Does the problem studied in this paper present novel challenges in comparison to the bandit setting? * The paper states that the online prediction setting is similar to streaming active learning. If this is the case, should the numerical experiment compare the regret convergence rate of the proposed algorithm with that of the streaming active learning algorithm? * Some writing aspects could be improved: * In line 227, the symbol $\tilde l$ is undefined. * In line 264, the regret bound (7) is termed "pessimistic" without explanation.

Questions

See Weaknesses above.

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The paper should further clarify the method and theory part, see Weaknesses above.

Reviewer naDi2023-08-10

Thank you for your response. I've read the authors' rebuttal and the other reviews and will maintain my position favoring acceptance (7).

Reviewer G2NF2023-08-13

Dear Author, Thank you very much for your response. I have read through global response which addressed the condition on $\delta$, and the interpretation (the expected cumulative loss gap growth). Also, thank you for addressing the proof from line 437 to 438. After reading all responses, I elevated my score from 7 to 8, since this is the first paper being able to quantify a $O(\sqrt{T})$ label complexity under the considered assumption given the best of my knowledge.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC