Online Ad Allocation with Predictions

Display Ads and the generalized assignment problem are two well-studied online packing problems with important applications in ad allocation and other areas. In both problems, ad impressions arrive online and have to be allocated immediately to budget-constrained advertisers. Worst-case algorithms that achieve the ideal competitive ratio are known, but might act overly conservative given the predictable and usually tame nature of real-world input. Given this discrepancy, we develop an algorithm for both problems that incorporate machine-learned predictions and can thus improve the performance beyond the worst-case. Our algorithm is based on the work of Feldman et al. (2009) and similar in nature to Mahdian et al. (2007) who were the first to develop a learning-augmented algorithm for the related, but more structured Ad Words problem. We use a novel analysis to show that our algorithm is able to capitalize on a good prediction, while being robust against poor predictions. We experimentally evaluate our algorithm on synthetic and real-world data on a wide range of predictions. Our algorithm is consistently outperforming the worst-case algorithm without predictions.

Paper

Similar papers

Peer review

Reviewer eG657/10 · confidence 3/52023-06-30

Summary

This paper is for the online matching problem in the display advertising domain. Essentially there are shopper/supply side requesting coming in in a sequential order, and static advertiser side demand with budget constraint. The goal is to do some sort of global optimization (as opposed to greedy strategy to maximize value for each request t). The paper proposed a learning-augmented algorithm and demonstrated its effectiveness via a theorem as well as simulations. Score changed from 6 to 7 after seeing the rebuttal from the authors.

Strengths

1. The robustness against low-quality prediction. This is a very nice property to have for an algorithm to be practical. It seems the benefit doesn't vanish thought it decays. Given that the prediction quality is always something with uncertainty and affects the cumulative value for any algorithm, what is shown in this paper is already good. 2. Convincing simulation/experiment results using real-world dataset. 3. I particularly like line 170 - 191 which has done a nice job explaining the intuition of algorithm 1.

Weaknesses

1. Don't know if the #impression over #advertiser ratio will significantly affect the simulation results, but I think this ratio are similar in those tested datasets. It's worth a little more discussion. I think for real-world scenario, the actual ratio won't be in the scale of 100. 2. I think the gain beyond the greedy algorithm that maximizes the single-impression utility is usually more important than the regret compared to OPT. Worth having related info in the paper. 3. How to handle the targeting constraint? This is very common in display advertising that an advertiser is only interested in a subset of impressions relevant to their own business. It effectively means one additional constraint in the primal problem. 4. The level of contribution beyond what is done in existing work. I didn't get the idea that Bama 2020 work forms this as an open problem though indeed they are relevant.

Questions

see pros/cons

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

3 good

Contribution

2 fair

Limitations

N/A

Reviewer 9H3m6/10 · confidence 3/52023-07-06

Summary

This paper studied Display Ads and the generalised assignment problem (GAP) where ad impressions arrive online and are allocated immediately to advertisers on the offline side. The difference between the two problems is that in Display Ads each impression takes uniform size of advertisers’ budgets, while in GAP the size is non-uniform. The paper considered the setting that incorporates predictions for allocation, which can be accessed by the algorithm upon arrival of each impression. The paper proposed a primal-dual algorithm based on Feldman et al. (2009a), with new ingredients to deal with predictions. Firstly, the new algorithm decides whether to allocate an impression to the predicted advertiser or to one with max discounted gain based on comparison involving a robustness-consistency trade-off parameter. Secondly, the algorithm updates the dual variable with an exponential mechanism involving the trade-off parameter. The paper provided theoretical analysis for the robustness and consistency factors, as well as experimental results on synthetic and real world datasets.

Strengths

1. Online matching and related problems have been extensively studied in the literature and found their applications in online business. This paper studied Display Ads and GAP, which are generalized version of online bipartite matching and Ad Words, that could capture more complicated scenarios in practice. Worst-case guarantees can be restrictive, while external or historical information may be available for making online decisions. This paper proposed an efficient and effective algorithm in this setting. 2. The paper proposed a solid primal-dual approach with theoretical performance guarantee that incorporates predictions that is accessible in general forms. The performance of the algorithm improves over worst-case and random mixture baselines in experiments. 3. The paper is clearly written. The main text provides necessary intuition and discussion about the algorithm and analysis, as well as comparison with related work.

Weaknesses

1. The paper only presented the proposed approach for Display Ads in the main text, but put the whole GAP part to Appendix. It would be helpful to briefly discuss in the main text how to extend the approach to GAP and take care of general impression sizes. 2. It is not clear if consistency is a good metric for evaluating the performance of algorithms with predictions. Unlike robustness, the consistency factor relies on a varying value of solution PRD to be compared against. For instance, in the upper left of Figure 3, the consistency value is high when PRD itself has low competitive ratio, while it is below 1 when PRD=OPT. It is hard to relate the absolute value of consistency to the quality of ALG. 3. The paper mentioned that the algorithm incorporates machine-learning predictions in abstract and introduction, which seems misleading. The actual algorithm receives a prediction PRD(t) for each impression t, which is presented in an abstract form and is not necessarily related to machine learning.

Questions

1. In Theorem 1 the robustness R is a decreasing function with respect to \alpha, but in experiments the robustness has a trend of (slightly) increasing w.r.t. \alpha. What is the cause of this phenomenon? 2. The limit of R(\alpha) in Theorem 1 tends to be 1 when \alpha tends to be 0. How does R(\alpha) compare with the worst-case ratio 1-1/e?

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

3 good

Contribution

2 fair

Limitations

The paper briefly discussed limitations of the proposed algorithm.

Reviewer 2o364/10 · confidence 3/52023-07-10

Summary

The paper discusses the problem of ad allocation and its generalization, the generalized assignment problem (GAP), which are two well-studied online packing problems with important applications in ad allocation and other areas. The paper presents an algorithm for both problems that incorporate machine-learned predictions and can thus improve the performance beyond the worst-case. The algorithm is based on the work of Feldman et al. (2009a) and similar in nature to Mahdian et al. (2007) who were the first to develop a learning-augmented algorithm for the related, but more structured Ad Words problem. The paper’s contributions are that it designs the first algorithms that incorporate machine-learned predictions for Display Ads and GAP. The two problems are general online packing problems that capture a wide range of applications. The algorithm follows a primal-dual approach, which yields a combinatorial algorithm that is very efficient and easy to implement. It is able to leverage predictions which can be learned from historical data. Using a novel analysis, the paper shows that the algorithm is robust against bad predictions and able to improve its performance with good predictions. In particular, it is able to bypass the strong lower bound on the worst-case competitive ratio for these problems. The paper experimentally verifies the practical applicability of its algorithm under various kinds of predictions on synthetic and real-world data sets. Here, it observes that its algorithm is able to outperform the baseline worst-case algorithm and the random-mixture algorithm.

Strengths

* Originality: lies in its use of machine-learned predictions to improve the performance of online packing algorithms for ad allocation, i.e., Display Ads and GAP. * Quality: presents a rigorous analysis of its algorithm and provides experimental results that demonstrate its effectiveness. * Significance: addresses an important problem in online advertising and provides a novel solution that can improve the performance of existing algorithms.

Weaknesses

1. The paper does not provide a detailed analysis of the limitations of its algorithm. For example, it is not clear how the algorithm would perform in situations where the predictions are inaccurate or where there are other sources of uncertainty. 2. The paper does not provide a comparison of its algorithm with other state-of-the-art algorithms for online ad allocation. This makes it difficult to assess the relative performance of the proposed algorithm and to determine whether it is truly novel and effective. 3. The paper does not provide a detailed discussion of the practical implications of its algorithm for real-world ad allocation systems. This could be addressed in future work by providing a more detailed analysis of the computational requirements and scalability of the proposed algorithm, as well as by conducting experiments on real-world data sets.

Questions

See the Weakness section, especially Weakness 2.

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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

See the Weakness section, especially Weakness 1.

Reviewer 3Ccp7/10 · confidence 4/52023-07-10

Summary

This paper considers the online display ads and generalized assignment problems under free disposal in the "algorithms with predictions setting". In the display ads problem there are offline advertisers $a$ with budgets $B_a$. A sequence of impressions arrive online with differing values to each advertiser (impression $t$ has value $w_{at}$ to advertiser $a$). The algorithm must either allocate each impression to an advertiser, earning its value and consuming one unit of the advertiser's capacity, or reject the impression entirely. Decisions are made irrevocably, but in the free disposal model the algorithm may assign more impressions to an advertiser than their capacity allows but the algorithm only gains the value of the most valuable impressions that fit within the capacity. The generalized assignment problem generalizes the display ads problem by allowing impression $t$ to instead consume $u_{at}$ units of advertiser $a$'s capacity. In the "algorithms with predictions" setting the online algorithm also has access to potentially noisy predictions about the optimal solution. In particular, this paper assumes that the algorithm may access a prediction $PRD(t)$ when impression $t$ arrives which gives a hint about which advertiser to allocate impression $t$ to in an optimal solution in hindsight. The prediction may be incorrect, so two quantities are analyzed: consistency and robustness. Consistency is given by the worst-case ratio of $ALG/PRD$, where $ALG$ is the algorithms solution quality and $PRD$ is the predictions solution quality. Robustness is given by the worst -case ratio of $ALG/OPT$, where $OPT$ is the optimal solution in hindsight. Theoretically, this paper gives algorithms with non-trivial tradeoffs between robustness and consistency. The algorithm is based off of the primal-dual algorithm used by Feldman et al. 2009, but modified to account for the predictions. Proving their consistency bound needed new techniques since a local analysis as is typically used in the primal-dual method is insufficient here. The authors complement the theoretical results with an experimental analysis on both real and synthetic data of their proposed algorithm using various predictions as input. The experiments compare to the worst-case baseline of Feldman et al. 2009 and the simple random-mixture algorithm they discuss in the introduction.

Strengths

This paper considers online allocation problems motivated by advertising applications in the algorithms with predictions setting. Prior work has considered related problems in this setting. This paper considers more general versions of these problems and also considers the free-disposal setting, which has not yet been considered in the algorithms with predictions literature. The paper gives strong theoretical guarantees for their algorithm and complements this with a thorough experimental analysis.

Weaknesses

- Some of the presentation of the results in the experiment section could be clarified/improved. See comments/questions below. - The improvement over the worst-case baseline using "more realistic" predictions is somewhat narrow (although in practice this smaller improvement could matter in some cases). - Theoretically, it is not clear if the trade-off between robustness and consistency is tight. Giving tight guarantees has become of interest recently, see the two references below: - Wei, A., Zhang, F., "Optimal Robustness-Consistency Trade-offs for Learning-Augmented Online Algorithms." NeurIPS 2020. - Jin, B., Ma, W., "Online Bipartite Matching with Advice: Tight Robustness-Consistency Tradeoffs for the Two-Stage Model." NeurIPS 2022. (this paper may also be good to discuss in the related work).

Questions

## Questions - Can you - Please clarify what is meant by "prediction competitiveness" in the plots in Figure 4. - What was observed experimentally for the competitive ratio of the worst case baseline? - In Figures 3 and 5 the average competitive ratio across 5 trials is reported in parentheses for different predictors. Since there is a hypereparameter $\alpha$ that is varied can you clarify which value of $\alpha$ was used to produce these numbers? ## Comments - For the plots in figure 2, it might be helpful to visually compare with the consistency/robustness guarantees that would be guaranteed by the random-mixture algorithm as well as the algorithm due to Mahdian et al. 2007.

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

I believe the authors have adequately addressed potential limitations and there is not a high potential for negative societal impacts from this work.

Reviewer 3Ccp2023-08-16

Thank you for your response, I will keep my score the same (accept).

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC