Putting Gale & Shapley to Work: Guaranteeing Stability Through Learning

Two-sided matching markets describe a large class of problems wherein participants from one side of the market must be matched to those from the other side according to their preferences. In many real-world applications (e.g. content matching or online labor markets), the knowledge about preferences may not be readily available and must be learned, i.e., one side of the market (aka agents) may not know their preferences over the other side (aka arms). Recent research on online settings has focused primarily on welfare optimization aspects (i.e. minimizing the overall regret) while paying little attention to the game-theoretic properties such as the stability of the final matching. In this paper, we exploit the structure of stable solutions to devise algorithms that improve the likelihood of finding stable solutions. We initiate the study of the sample complexity of finding a stable matching, and provide theoretical bounds on the number of samples needed to reach a stable matching with high probability. Finally, our empirical results demonstrate intriguing tradeoffs between stability and optimality of the proposed algorithms, further complementing our theoretical findings.

Paper

Similar papers

Peer review

Reviewer ct3v5/10 · confidence 5/52024-07-11

Summary

This paper investigates the bandit learning problem in matching markets. It introduces a critical perspective that the objective of regret minimization does not align with achieving market stability. The study explores the sample complexity required to find a stable matching. To address this issue, the authors propose two algorithms: the Uniform Sampling DA algorithm and the Arm Elimination DA algorithm, and provide guarantees based on the corresponding envy-set size.

Strengths

1. This paper raises a significant issue in the literature of bandit learning in matching markets, highlighting that the objective of regret minimization may not align well with the goal of finding a stable matching. 2. The investigation of sample complexity in this paper delves into the detailed structure, specifically examining the envy-set size with respect to a particular stable matching, thereby deepening the understanding of the learning task's difficulty.

Weaknesses

1.Although some existing works consider stable regret, their theoretical guarantees also imply guarantees for the sample complexity of reaching (player-optimal/pessimal) stable matchings, as demonstrated by Liu et al. [2020], Kong and Li [2023], and Zhang et al. [2022]. However, this paper does not compare its results with the induced sample complexity of these existing works. Consequently, it is difficult to determine how this work improves upon existing studies in terms of sample complexity and the objective of achieving stable matchings, specifically player-optimal and player-pessimal stable matchings. 2.The technical novelty is limited. The uniform-sampling DA algorithm is standard in the literature, and the arm elimination DA algorithm appears similar to the ODA algorithm in Kong and Li [2024], although they consider a more general many-to-one setting. Please verify this.

Questions

Please see the last part.

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors adequately addressed the limitations.

Reviewer ct3v2024-08-10

Thanks for your detailed response. I am happy to increase my score. I also want to further ask how to derive an $O(K\log T/\Delta^2) $ regret for the AE-arm-DA? As its current result depends on $|ES(\underline{m})|$ which is of order $O(NK)$.

Authorsrebuttal2024-08-10

Thank you for taking the time to read our response and seek clarification. We are glad to clarify your question. Please note that the definition of player-pessimal stable regret is the regret for only ONE agent, while the definition of sample complexity in our paper is how many samples ALL agents need collectively to reach a stable matching. Therefore, uniform arm-DA algorithm has player pessimal stable regret depending on $K$ and sample complexity depending on $NK$; AE arm-DA algorithm has sample complexity depending on $|ES(\underline{m})|$. We didn't do regret analysis for AE arm-DA in this paper. Note that the ordering we wrote for player-pessimal stable regret is based on the empirical observations.

Reviewer ct3v2024-08-10

Thanks for your response. I wonder whether there is some misalignment between the stable regret and sample complexity. Consider the global preference example where all agents' preference rankings are $b_1>b_2>...>b_K$, and all arms' preference rankings are $a_1> a_2> ... >a_N$. In this case, the agent $a_1$ will sample all arms until it identifies $b_1$ is optimal. Then the agent $a_2$ will sample arms except for $b_1$ until it identifies $b_2$ to be optimal. The agent $a_N$ needs to wait for all of the other agents to sample arms. This part of the samples contributes not only to the regret of the agents with higher priority but also $a_N$'s own regret. However it does not contribute to the sample complexity repeatedly. So the sample complexity/N is not equal to a single agent's stable regret. Is this right?

Authorsrebuttal2024-08-10

Right, it is not true that sample complexity/N is regret. In fact, as we saw, sample complexity is a function of probability budget $\alpha$, and regret is a function of $T$. Observe that we define regrets based on the difference between the sampled arm with the player-optimal (or pessimal) stable arm. When an agent does not sample an arm, it does not incur extra regret. In your example, by our AE arm-DA algorithm, when agent $a_1$ samples all arms, other agents are not sampling so they should have 0 regret.

Reviewer ct3v2024-08-11

I think it is not the traditional stable regret definition as it does not consider the collision scenario in the multi-agent problem. When a player samples nothing, it receives 0 rewards. So the regret is the mean of the (optimal or pessimal) stable arm, instead of 0. This is the definition in all of the existing works [Liu et al. 20; Kong and Li. 2023, 2024].

Authorsrebuttal2024-08-11

Yes, We agree with the reviewer that our setting is a little different from the traditional setting and we do not consider regrets in our theory. Instead, we consider sample complexity, which focuses on minimizing the amount of samples to reach a stable matching.

Reviewer ct3v2024-08-11

Overall, I appreciate the contribution of studying the pure exploration problem in matching markets and the use of the envy-set notion to measure its complexity. However, since both the uniform sampling and arm elimination algorithms are similar to existing approaches, I am adjusting my score to 5. I encourage the authors to discuss the relationship between their sample complexity and the standard regret, as well as their algorithmic difference between arm elimination and ODA in the next version.

Reviewer ormW7/10 · confidence 3/52024-07-15

Summary

The paper studies stability in matching problems with learned preferences on one side. The two sides are agents and arms. The agents learn their preferences over the arms through sampling while the arms know their preferences over agents. Then a DA algorithm is run with either arms or agents proposing. The paper gives bounds on the probability that the matching found is unstable, as well as sample complexity results, with uniform sampling and for a more refined arm-elimination algorithm. The paper is complemented by numerical simulations.

Strengths

The problem studied, the stability of the matching found by learning, is interesting and novel. Indeed, much of the literature rather focused on the regret achieved but I have not seen many works on looking at the probability of blocking pairs. The bounds proved are relevant and well behaved, and relatively intuitive. The comparison of uniform sampling and arm elimination (which basically removes a factor K) is nice. The paper is easy to follow.

Weaknesses

1. The bounds proved contain a term in |ES(m)|. This term is natural but may be quite large. It would be good to give an order of magnitude for some reasonable preferences. 2. The theoretical tools to prove the main results are relatively standard from bandits. This is not a weakness in itself, I am mentioning it to justify that I do not see the contribution as a particularly intricate theoretical analysis. 3. Is it possible to provide lower bounds to know whether the analysis is tight? And to do better than arm elimination? 4. The motivation for the problem studied could be further developed to justify the setting introduced (sampling, then DA). For instance, a reasonable alternative would be to have a more dynamic setting where different agents get matched at different times, e.g., when they are confident that they are done exploring (and then the pair leaves the system). Of course, this would be a different problem and I am not suggesting the authors should have studied this one instead, but it would be nice to strengthen the motivation for the given problem studied.

Questions

See above, in particular 1. and 3.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer wAZr4/10 · confidence 4/52024-07-16

Summary

The paper studies the sample complexity of finding a stable matching under the probably approximately (PAC) framework. In the model, $N$ agents are to be matched with $K \ge N$ arms. Each arm has an (a priori unknown) utility to each agent, inducing the agents’ preferences; Arms also has strict preferences over the agents. Agents learn about their rewards by pulling the arms and observing stochastic rewards with means equal to the utilities. The objective is to find a matching that is stable with the desired probability with least possible samples (pullings of arms). The paper proposes an arm-proposing deferred acceptance (DA) algorithm with action elimination (AE). The algorithm implements the arm-proposing DA, with the modification where, in face of uncertain preferences over a new proposal and the current match, the agent uses AE to reach a decision of rejection. The sample complexity of the algorithm is analyzed and simulations are used to evaluate the algorithm empirically.

Strengths

The paper combines the online learning from stochastic observations with the classic literature on stability of matchings, a novel and important direction. While intuitive, the algorithmic combination of arm-proposing DA with AE is novel to my understanding. The paper is overall clearly written, with consistent notations and clear definitions. I find it easy to follow. The claims seem reasonable, although I did not go over all the proofs.

Weaknesses

While the results may be new and interesting, I do not find them sufficiently significant. From the theoretical perspective, the algorithm proposed and its analysis both seem fairly standard. The algorithmic improvement also seem marginal, as Theorem 5 only seem to provide a weak improvement on existing results, and nor do the simulations (which are run in a very restrictive setting with a specific form of preferences) seem conclusive on the superiority of the main algorithm. [Edit: From the discussion, I understand that there is some valid rationales behind the simulations (the choice of utility distribution in particular). I think it is worth more discussion and justification.] On the empirical front, the paper would benefit from additional motivating examples on why the results can be impactful in real world applications. As will be mentioned in one of my questions, I am not sure how the proposed algorithm can be applied to settings such as school admission, ride share, etc. - in particular, the assumption that rewards are iid and that an arm can be pulled (sampled) repeatedly in an arm-proposing fashion seems hard to interpret in these settings. Without sufficient practical motivation, the results seem only a nice yet marginal theoretical exercise.

Questions

- Can the authors elaborate on the assumption that $N \le K$? Note that, since the main algorithm is an arm-proposing DA variant, the two sides are not symmetric and such an assumption requires justification. When $N > K$ we know that the entire set of stable matchings can exhibit strikingly different behaviors than when $N\le K$ [1]. - Is there any reason why the experiment is only carried out on the specific utility model with utilities 1 through 20? What if they come from other distributions, e.g., exponentially or polynomially growing? The additional SPC setting is also quite restrictive in my understanding. - Have the authors considered relaxing the notion of stability in such stochastic settings (e.g., [2,3])? Ultimately, strict stability is a rather restricted solution concept and often beyond hope, and an approximate notion may be both more realistic and also more tractable. - How should one reason about the minimum preference gap? In general, I would expect this quantity to scale down as the market size increases; realistically speaking, at most $1/n$. - Further, indifference can be common in real world applications, and I would hope that the model and the algorithm should allow for such scenarios - pressing a student to form a precise preference over schools that are otherwise entirely comparable seems unnecessary. Maybe approximate stability could be one way to resolve this? E.g., maybe there is a better measure of “average stability” that accounts for utility gap rather than just the probability of stability (Line 324, in my understanding). [1] Itai Ashlagi, Yash Kanoria, and Jacob D. Leshno. 2017. Unbalanced Random Matching Markets: The Stark Effect of Competition." Journal of Political Economy. [2] Yannai A. Gonczarowski, Noam Nisan, Rafail Ostrovsky, and Will Rosenbaum. 2015. A Stable Marriage Requires Communication. In Proceedings of the 2015 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA). [3] Itai Ashlagi, Mark Braverman, and Geng Zhao. 2023. Welfare Distribution in Two-sided Random Matching Markets. In Proceedings of the 24th ACM Conference on Economics and Computation (EC '23). Minor comments: - Line 135: typo. “… leads to unique …” => “… leads to a unique …” - Line 359-360: “such variations often introduce computational complexity with known preferences” - I might be missing something obvious, but in the hope to understand this claim better, I hope the authors could clarify what “computational complexity with known preferences” they are referring to here. In my understanding, with known preferences, the vanilla DA should work fine with ties and is efficient. - When the problem is formulated as agents pulling arms, I find it hard to reasonable arms proposing. To me, this means that the side that bears the burden of exploration in fact does not control their very own action? Is there a better setup or application that motivates the study?

Rating

4

Confidence

4

Soundness

2

Presentation

3

Contribution

1

Limitations

Limitations are discussed in Section 7: Conclusion and Future Work. The paper is mostly of a theoretical nature.

Authorsrebuttal2024-08-12

Thank you for your responses, increasing the score, and taking the time to read our responses. 1.⁠ ⁠We discuss about ties in limitation. 2.⁠ ⁠⁠As we mentioned, since our choice of utilities are similar to experiments in previous papers [Liu 2021], we did not discuss this. We can add it briefly in revision. 3.⁠ Please note in this line of bandit learning in matching markets, the true utility is given and we do not have any assumption about the utility distribution. Thus our use of envy-set (Theorem 5) exploits the structure of the input preference. Since the utilities are part of the input, we cannot discuss about the average case but we provided the best case and the worst case improvements over the uniform DA algorithms in Lemma 3. For an example of practical scenario, please see our labor market example in rebuttal.

Reviewer Uabi6/10 · confidence 4/52024-08-05

Summary

This paper studies the unexplored question of stability in the study of two-sided market matching problems where preferences of one side are unknown and have to be progressively learned through a bandit learning mechanism. The paper makes a significant contribution in this area, which has hitherto only considered the regret cost aspect. Yet it still neglects the fairness aspect of the ensuing matching, which is a weakness in most solution ignoring balance cost and sex equality cost. It would be interesting to consider the cost of eliciting preferences and solving the problem in the classic manner, and illustrate the benefit of the proposed solution over that approach.

Strengths

S1. Solid improvement in terms of probability to yield stable solutions leading to Arm Elimination arm-DA algorithm. S2. Derivation of bounds on samples needed to reach stable matching w.h.p. S3. Experimental study among proposed variants.

Weaknesses

W1. No consideration of fairness aspects of the ensuing matching. W2. No illustration of benefit vs. solution that elicits all preferences.

Questions

What is benefit of the proposed methods over one that pays the cost to elicit all preferences?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer ormW2024-08-08

Thank you for the responses, I am happy to keep my score as is.

Reviewer wAZr2024-08-12

I appreciate the very detailed response from the authors. I have taken the discussion into consideration and adjusted my evaluation accordingly to reflect my current judgment: 1. I agree that the incompatibility with indifference in preferences is, unfortunately, common to prior works in this line of literature. I invite the authors to discuss such limitation in the paper and the ideas to overcome it. 2. It seems that the authors are thoughtful with the choice of utility distribution in the evaluation. While I still don't think it is necessarily the most convincing one, the additional argument now makes it more appealing to me. I might suggest including additional discussion on the rational of such a choice, as summarized in the rebuttal comment. 3. Overall, I see the contribution from this work but personally I still think it falls short of the bar for NeurIPS. In a very short summary, I did not find the "at best infinite" argument along with the simulation convincing enough (maybe some average case notion of improvement can be more helpful?), and I still have reservations over the practical significance of the results.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC