Trading-off price for data quality to achieve fair online allocation

We consider the problem of online allocation subject to a long-term fairness penalty. Contrary to existing works, however, we do not assume that the decision-maker observes the protected attributes -- which is often unrealistic in practice. Instead they can purchase data that help estimate them from sources of different quality; and hence reduce the fairness penalty at some cost. We model this problem as a multi-armed bandit problem where each arm corresponds to the choice of a data source, coupled with the online allocation problem. We propose an algorithm that jointly solves both problems and show that it has a regret bounded by $\mathcal{O}(\sqrt{T})$. A key difficulty is that the rewards received by selecting a source are correlated by the fairness penalty, which leads to a need for randomization (despite a stochastic setting). Our algorithm takes into account contextual information available before the source selection, and can adapt to many different fairness notions. We also show that in some instances, the estimates used can be learned on the fly.

Paper

References (59)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer aPcw7/10 · confidence 3/52023-07-05

Summary

Existing works in the fair online allocation space assume that the decision-maker has access to protected attributes. In contrast, this paper studies the choice of a data source as a multi-armed bandit problem in conjunction with the primary allocation problem. This paper introduces a computationally efficient algorithm with sublinear regret in the general case. The contributions of this paper are exclusively theory-based, including regret bounds for various types of regret penalties.

Strengths

This is the first paper to consider optimizing over data sources in the problem space of fair online allocation. This paper contributes an algorithm with proven regret bounds, which is significant. I believe that other researchers are likely to build off of this work, and that it advances research in a demonstrable way. This paper is excellently written. It is clear, well organized, and adequately informs the reader. In particular, I am happy to see good explanations of the reasoning, limitations, and assumptions made in this work.

Weaknesses

This paper does not contribute code or empirical results; both of which would be of interest to the research community.

Questions

N/A

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

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Yes

Reviewer uY3q5/10 · confidence 2/52023-07-07

Summary

The authors formulate the online fair allocation problem, where a decision maker may choose an information source $k_t$ based on public context $z_t$ and generate an allocation decision $x_t$ with a reward $1$ if $x_t=1$. The fairness penalty is some convex function that is minimized when the decision $x_t$ is uncoupled with the private type $a_t$. The authors developed a virtual-value-based algorithm that achieves $O(\sqrt{T})$ online regret.

Strengths

The fairness problem is of interest to the community, especially when the private attributes are unseen.

Weaknesses

It seems that the source selection part is simply the exponential weight algorithm with some virtual values which can be updated by the EXP3 algorithm. The allocation is based on the dual value $\lambda_t$ which can then be updated by the dual gradient descent. In my opinion, the authors did not make enough effort to explain the intuitions and justified the use of these algorithms, let alone the novelty of the work. Also, I'm not sure about the technical involvement, since it seems that the only challenge here is having multiple sources of information.

Questions

What is the theoretical lower bound for this problem?

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

4 excellent

Presentation

2 fair

Contribution

2 fair

Limitations

No concerns here.

Reviewer Gv3t6/10 · confidence 3/52023-07-07

Summary

This paper studies an online allocation problem with long-term fairness constraints where, unlike prior works, the protected attributes of online users (e.g., gender, ethnicity) are not observed, but they can be purchased from $K$ different sources for some additional cost. The goal is to maximize the overall utility of the users minus the cost of additional information that was bought minus a fairness penalty that is characterized by a convex function $R(\cdot)$. The authors study this problem under an i.i.d. assumption on the users and propose Algorithm 1 with sublinear regret for the problem.

Strengths

- Addressing fairness constraints in online allocation is a very well-motivated problem and the authors' assumption that protected attributes are not observed is a more realistic model of the applications. Therefore, the contributions of this work are of potential interest to a wide audience. - Simultaneous source selection and dual gradient descent in Algorithm 1 is an important contribution that shows the naive approach of performing these two tasks separately is not good enough for this problem. - The proposed algorithm is compatible with various notions of group fairness.

Weaknesses

- The problem setup is a bit confusing. I'm not sure how the public context $z_t$, protected attributes $a_t$, and the purchased piece of information $c_{tk_t}$ are related to each other. On line 152, the authors make the assumption that $c_{tk_t}$ contains $z_t$ and I'm not quite sure if I understand what that means. - The authors have distinguished their paper from prior works based on the fact that protected attributes are not observed and could only be purchased. However, a comparison of the paper with prior works in terms of techniques and proof ideas is missing. It appears that both primal and dual updates in Algorithm 1 are quite standard and the main challenge is to combine these two steps. This point needs to be highlighted and clarified. - Given the vagueness of the problem framework, explaining the setting in the context of a few motivating applications is very helpful. In particular, the process of purchasing data from data sources needs clarification.

Questions

- Can you provide a few motivating applications that could be cast into your framework and explain what data sources are and how they relate to each other? - Can you explain the connection between $a_t$, $z_t$, and $c_{tk_t}$ in the context of a motivating application? How do these quantities relate to each other? What is the dimension of each quantity? How is $\mathbb{E}[a_t|c_{tk_t}]$ estimated? - Assuming that $\sum_{t}x_t>0$, the function $(\sum_t x_t)R(\frac{\sum_t a_tx_t}{\sum_t x_t})$ is simply the perspective of the convex function $R(\cdot)$ and therefore, it should be convex in $x=(x_1,\dots,x_T)$. However, on lines 329-330, it is claimed that the opposite is true. Can you explain why convexity doesn't hold? - Is it possible to use Algorithm 1 to obtain results for the adversarial setting? What would be the main challenges? - On line 131, why is $c_{tk}$ enough to estimate $u_t$ and $a_t$? How about the public context information $z_t$, is it not helpful for estimation? -------------------------------- I've read the authors' rebuttal, thanks for addressing my questions and concerns.

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

Discussion of potential negative societal impact is not necessary.

Reviewer 2Yz95/10 · confidence 2/52023-07-11

Summary

This paper studies the problem of online resource allocation with long-term fairness constraints. Different from mainstream literature, the authors assume that the fairness context is not known to the player in advance. However, the player does have the option of paying to buy fairness context information from different sources to make more fair decisions. The algorithms consists of two parts, the first part focuses on source selection for buying protected fairness information and the second part focuses on the allocation decision. The regret of the proposed algorithm is analyzed and it is shown that it is $O(\sqrt(T))$. Last, there are some additional discussions and results regarding some other notions of fairness.

Strengths

++ It is interesting that the authors consider the setting where protected attributes (such as demographics) are unknown to the online decision maker, but the decision maker should still be fair. It makes sense that prior works on fair allocation would only consider the case where these are known, but the reviewer is not familiar enough with the literature to know whether that’s true. ++ The assumption that a user’s protected attributes can be estimated from the online context and/or additional data which is purchased is an interesting addition to the setting. ++ It’s nice that it can model multiple notions of fairness as discussed in the last section of the paper.

Weaknesses

-- The fairness in their setting is a penalty rather than a hard constraint. This should be clarified since the claim of long-term hard fairness constraints is mentioned multiple times in the paper. -- I was expecting to see more discussion about the trade-off between paying for additional information and paying for fairness — i.e., when fairness doesn’t cost much, how does that impact the allocation? Maybe the regret bounds still hold in either case, but it doesn’t tell us qualitatively how the algorithm actually creates fair allocations. Especially since this “trade-off” is explicitly mentioned in the title of the paper… -- The assumption that users are drawn i.i.d. from a probability distribution enables this second assumption that the conditional expectation of protected attributes given context is known (learned from past data). However, in practice, past data is often unfair, which means I think these estimates may be inaccurate, particularly for underrepresented groups, and that intuitively should impact fairness here. It seems that there is no discussion of distribution shift or, in general an acknowledgment that the past data might exhibit this property. They do mention in the conclusion that the estimate of protected attributes could be learned online.

Questions

see above.

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

n/a

Reviewer 2Yz92023-08-14

response

Thank you for the detailed response and additional experiments. I have increased my score.

Reviewer Gv3t2023-08-15

Thank you for your comprehensive response. My main confusion was about the relationship between various pieces of information and I think the motivating examples help clarify this point. I encourage the authors to maintain a running example throughout the paper and explain their model assumptions in the context of the example.

Authorsrebuttal2023-08-18

We thank the reviewer for their reply and relevant suggestion, we will integrate the examples mentioned above throughout the paper to give better intuition for the model.

Reviewer uY3q2023-08-16

I appreciate the authors' efforts in addressing my concerns. In my opinion, it would be beneficial for the authors to incorporate these clarifications and discuss in more detail on relations/differences to previous work in their paper. Furthermore, I suggest considering a formal discussion of the challenges inherent in the problem. This is particularly important as the distinctions in the setup might be nuanced, and even a minor alteration could give rise to significant technical complexities.

Authorsrebuttal2023-08-18

We thank the reviewer for their reply and the constructive feedback. We will make the distinction with previous works clearer in the introduction near the contribution section, as well as better highlight the technical difficulties and challenges inherent to this problem in Section 3.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC