User-item fairness tradeoffs in recommendations

In the basic recommendation paradigm, the most (predicted) relevant item is recommended to each user. This may result in some items receiving lower exposure than they"should"; to counter this, several algorithmic approaches have been developed to ensure item fairness. These approaches necessarily degrade recommendations for some users to improve outcomes for items, leading to user fairness concerns. In turn, a recent line of work has focused on developing algorithms for multi-sided fairness, to jointly optimize user fairness, item fairness, and overall recommendation quality. This induces the question: what is the tradeoff between these objectives, and what are the characteristics of (multi-objective) optimal solutions? Theoretically, we develop a model of recommendations with user and item fairness objectives and characterize the solutions of fairness-constrained optimization. We identify two phenomena: (a) when user preferences are diverse, there is"free"item and user fairness; and (b) users whose preferences are misestimated can be especially disadvantaged by item fairness constraints. Empirically, we prototype a recommendation system for preprints on arXiv and implement our framework, measuring the phenomena in practice and showing how these phenomena inform the design of markets with recommendation systems-intermediated matching.

Paper

Similar papers

Peer review

Reviewer t7QC3/10 · confidence 4/52024-07-07

Summary

In this paper, the authors studied the tradeoff between user and item fairness in a recommendation setting. They proposed a constrained optimization problem that imposes user fairness as its objective and incorporates item fairness as its constraints. The authors also identified that (1) when user preferences are diverse, item fairness can be easily achieved; (2) when there is mis-estimation of user preference, imposing item constraint can lead to further cost for the users. Finally, the authors illustrated their findings using arXiv data.

Strengths

- The problem of multi-sided fairness is an important problem, but most existing literature focus on studying single-sided fairness. Understanding the price of fairness is a meaningful problem for decision-makers in practice.

Weaknesses

1. The model and framework used in this paper would require significantly more justification. More specifically, - It is unclear why the authors choose to consider an optimization problem which maximizes normalized user preference as the objective subject to item fairness constraints. To me this is a rather counterintuitive choice and requires more motivation. Why not solving a dual-objective problem and treating user/item fairness in the same fashion? Alternatively, why not maximizing online platform's recommendation quality? - The assumption that user and item share the same utility $w_{i,j}$ is too strong and can hardly hold in practice. In recommendation systems, a number of factors such as pricing, rankings, utility models, etc., could impact the item/user utilities in different ways. (See prior works such as [10, 11, 32], all of which use different definitions of user/item utilities.) - The fairness notion that the authors adopt for users/items resemble a min-max type of fairness notion. This is again quite restrictive and requires more justification. Does your framework and results hold under alternative notions? 2. The theoretical results of this paper do not have sufficient technical contributions. For example, - Proposition 2 basically uses the properties of BFS in LPs. - Theorems 3 and 4 are only shown on a restrictive example (where there are only 2-3 types of users with opposing preferences in a pre-defined form). This raises the question of whether the theoretical results/insights can be extended to more general setups. Due to the restrictive setup and assumptions, the insights might also not have much practical relevance. 3. The price of fairness of multi-sided recommendations is a topic already studied in prior works. The insights provided in this work are not particularly surprising, nor distinguishable from prior works. - For example, in [11] the authors also studied price of fairness and show that the price relates to the misalignment of platform/item/user objectives and designed an algorithm that resolves the issue of having unknown user/item data. - The phenomena described in this work, such as more diverse user preferences naturally inducing item fairness, is also rather natural especially when min-max fairness is imposed, which encourages uniform item exposure. I'd expect such statement would fail to work if a different type of item fairness is considered. 4. No algorithm has been studied or introduced in this work. The empirical study of arXiv data merely serves to evaluate Problem (1) on arXiv data, but the model and framework itself already raises questions and requires extensive justification. Overall, I think this work would require significantly more work in both its model and results and would recommend rejection.

Questions

See weaknesses.

Rating

3

Confidence

4

Soundness

2

Presentation

2

Contribution

1

Limitations

See weaknesses.

Reviewer yH5N2024-08-06

Dear reviewer: I am another reviewer and an engineer who work in industrial RS. I would like to answer your question. "Why not solving a dual-objective problem and treating user/item fairness in the same fashion? Alternatively, why not maximizing online platform's recommendation quality?" In recommendation system, the key goal is to serve most user the item she want. The "platform's recommendation quality" could be evaluated in two perspectives: 1) is the users as a whole group satisfied? namely the "user utilities" defined in line 135. 2) is every individual satisfied? In other world, maximize the user utilities of the least satisfied user, ie, "minimum user utilities" defined in this paper. In industrial system, we care about both the overall recommendation quality (point 1) and the recommendation quality for each user (point 2). Actually, if the least satisfied user is satisfied, the overall recommendation quality should be ok, although not optimal. While whether an item get enough chance to be seen, is not the key concern of commercial RS. Bad item should naturally be less shown to users. Thus, "treating user/item fairness in the same fashion" does not make sense. A reasonable formulation is to treat recommendation quality as objective, while item fairness as constrain.

Reviewer rEVK7/10 · confidence 3/52024-07-09

Summary

The paper works on the relationship between user fairness and item fairness in recommender system settings. A theoretical framework is proposed and some theoretical results and intuitions are provided based on the framework. The main results are the tradeoffs between fairness and 1) uncertainty and 2) diversity where theorems are proved together with some discussions. --- I have read the rebuttal and other reviews. My rating is unchanged.

Strengths

I like this paper. The theoretical framework is clear and useful. I expect more results can be derived out of this framework in the future e.g. assuming the utility w_{ij} is an estimator with certain bias and variance, or in practice one may not achieve the global optimal \rho^* but with a gap. The main results lie in theorem 3 and 4. Both results seem valid and intuitive. They illustrate how the price of fairness is changed when other factors like diversity/uncertainty are part of the consideration.

Weaknesses

The empirical part on arxiv is quite light. I see more details in the appendix but the main body part needs more information to make it self-contained. In the theoretical framework part, I would make the setting more comprehensive but put additional assumptions in later sections in order to support those theorems e.g. I think the current setting is for a single-item recommendation problem and all users are independent. There will be correlation terms for different items when multiple items are shown to users together. In addition, the paper focuses on individual utility.

Questions

I think the platform-level utility is missing from the discussion? In practice the platform controls the recommendation algorithm and therefore they may favor their utility as the primary objective. The default for cold start user may not be the average of existing users, but selecting top-performing items. Suggest to consider this aspect in the revision process.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

n/a

Reviewer 3QdJ6/10 · confidence 3/52024-07-12

Summary

This paper investigates the trade-off between user fairness and item fairness in recommender systems. The authors develop a theoretical framework to characterize the user-item fairness trade-off by analyzing the recommendation strategy optimization problem. The following phenomena are found: 1. The more diversified the user preferences are, the smaller the user-item fairness trade-off is. 2. Inaccurate estimation of user preferences exacerbates the fairness trade-off, especially for new users. 3. In real data, moderate item fairness constraints have a small effect on user fairness but very strong constraints can significantly reduce user fairness. Overall, the theoretical derivation part of this study is brilliant and provides theoretical assurance for the framework. It is a worthy study to explore the fairness problem of recommender systems in depth.

Strengths

1. The trade-off relationship between user fairness and item fairness is systematically analyzed for the first time. 2. A new theoretical framework is proposed to simplify complex optimization problems. 3. the theoretical analysis is rigorous and provides in-depth mathematical proofs.

Weaknesses

1. Limitations of the definition of fairness: the paper focuses mainly on minimized fairness indicators. Have other indicators of fairness been considered? 2 No specific methodology is given for actually achieving the three balances 3. The thesis assumes that the utility of users and items is symmetric, which may not always hold in reality.

Questions

1. The paper assumes that the utility of users and items is symmetric, which may not always hold in reality. How does this assumption affect the generalisability of the results? 2. In this paper, fairness is quantified as the minimum normalized utility. This definition is supposed to follow Rawlsian fairness. it is not based on group fairness? fairness in this paper is measured using individuals. I am concerned about this because there are a large number of users in a recommender system and individual users may not be representative. I would like the authors to explain my concerns. 3. The selection of scenarios in the experimental section is limited, and the experiment on arXiv recommender systems, while meaningful, may not be sufficiently representative of all types of recommender systems. The generalisability of the results of this experiment could be discussed, or additional experiments in other domains could be considered.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

1. Simplification of model assumptions: It assumes symmetric user and item utilities, which may not always hold in reality. It only considers the case of a single-item recommendation, whereas real systems usually recommend multiple items. 2. Limitations of the fairness definition: it focuses mainly on minimized fairness metrics and does not consider other possible fairness measures (e.g. group fairness).

Reviewer yH5N7/10 · confidence 3/52024-07-25

Summary

This paper develop a theoretical framework to analysis the trade-off between user fairness and item fairness. From the theoretically analysis, we understand that diverse user population benefits the recommendation, and users whose preferences are misestimated can be disadvantaged by the constraints on item fairness. The conclusion makes sense and is useful.

Strengths

Clear definition on user fairness, item fairness, item utility constrained user utility, price etc. Good analysis on the two conclusion.

Weaknesses

Relatively easy setting, e.g., only one item is recommended to a user. While the author honestly pointed out these weaknesses in the last paragraph, these weaknesses are acceptable.

Questions

.

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

4

Limitations

.

Reviewer yH5N2024-08-12

I have read the rebuttal

Reviewer rEVK2024-08-13

Thanks

I have read the rebuttal and other reviews. Still, I think this is a solid work and will leave my rating (7-accept) unchanged

Reviewer 3QdJ2024-08-14

Thanks to the authors for the response. I keep my positive score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC