PaperlayerView exact revision on arXiv

Source-linked AI summary

FairRec: Two-Sided Fairness for Personalized Recommendations in Two-Sided Platforms

Gourab K Patro, Arpita Biswas, Niloy Ganguly, Krishna P. Gummadi, Abhijnan Chakraborty

arXiv:2002.10764v2cs.AIcs.GT

TL;DR

Two-sided platforms must balance customer satisfaction with producers’ exposure and well-being, since customer-centric recommendations can concentrate opportunities among few producers. FairRec maps recommendation to fair allocation, providing two-sided fairness with theoretical guarantees and marginal recommendation-quality loss.

  • Problem

    Customer-centric recommendations may undermine producers’ well-being by concentrating exposure, while producer-centric designs may become unfair to customers.

  • Method

    FairRec maps fair recommendation to fair allocation and is agnostic to the data-driven relevance model, supporting scalability and adaptability.

  • Results

    FairRec exhibits two-sided fairness with theoretical guarantees and empirical evidence from extensive real-world evaluations while causing a marginal loss in recommendation quality.

  • Takeaways & Limitations

    FairRec can be applied to offline fair recommendation scenarios such as mass promotions through email, app, and web notifications.

  • Takeaways & Limitations

    The formulation treats products and producers interchangeably; for multi-product producers, product-level fairness is intended to ensure producer fairness through exposure proportional to portfolio size.

Abstract

from arXiv · show

We investigate the problem of fair recommendation in the context of two-sided online platforms, comprising customers on one side and producers on the other. Traditionally, recommendation services in these platforms have focused on maximizing customer satisfaction by tailoring the results according to the personalized preferences of individual customers. However, our investigation reveals that such customer-centric design may lead to unfair distribution of exposure among the producers, which may adversely impact their well-being. On the other hand, a producer-centric design might become unfair to the customers. Thus, we consider fairness issues that span both customers and producers. Our approach involves a novel mapping of the fair recommendation problem to a constrained version of the problem of fairly allocating indivisible goods. Our proposed FairRec algorithm guarantees at least Maximin Share (MMS) of exposure for most of the producers and Envy-Free up to One item (EF1) fairness for every customer. Extensive evaluations over multiple real-world datasets show the effectiveness of FairRec in ensuring two-sided fairness while incurring a marginal loss in the overall recommendation quality.

1 INTRODUCTION

Two-sided platforms must balance personalized customer recommendations with fair producer exposure. FairRec maps this problem to constrained fair allocation to provide producer exposure guarantees and customer fairness.

  • Personalized top-k recommendations prioritize customer satisfaction but can create substantial disparities in producer exposure.
  • Unequal exposure can reduce producers’ economic opportunities and eventually limit customer choice and platform experience.
  • Fair recommendation should distribute producer exposure fairly while distributing losses in customer utility fairly among customers.
  • FairRec maps recommendation fairness to constrained indivisible-goods allocation, targeting minimum producer exposure and EF1 customer fairness.
  • The algorithm is model-agnostic, provides theoretical guarantees, and is evaluated extensively on multiple real-world datasets.

2 BACKGROUND AND RELATED WORK

Prior work studies fairness in multi-stakeholder platforms and fair allocation of indivisible goods. This paper distinguishes its focus on individual fairness for both producers and customers.

  • Fairness in Two-Sided Platforms: Research on two-sided platforms has examined group biases involving hosts, workers, advertisers, and marketplace producers.
  • Fairness in Two-Sided Platforms: Existing studies report visibility and popularity disparities among producers, including disadvantages for newer producers and popularity bias.
  • Fairness in Two-Sided Platforms: Some prior work addresses group fairness for producers and customers or two-sided fairness in matching problems.
  • Fairness in Two-Sided Platforms: This paper addresses individual fairness for both producers and customers and connects it to long-term platform sustainability.
  • Fair Allocation of Goods: Fair allocation research developed EF1 and MMS for indivisible goods when classical EF or proportional fair share may be infeasible.

3 PRELIMINARIES

The paper models products and producers within a two-sided recommendation setting, where relevance determines customer utility and uniform attention determines exposure.

  • 3.1 Producers and Products: Products and producers are treated interchangeably, with product-level fairness generalizable to producers with multiple products through portfolio-proportional exposure.
  • Each customer receives a set of k products from the producer set.
  • Relevance V_u(p) represents the likelihood that customer u would like product p and serves as a proxy for utility.
  • Top-k relevant products maximize the customer’s utility under the paper’s additive relevance formulation.
  • 3.5 Producer Exposure: Producer exposure is the total attention received across customers to whom the product is recommended, under a uniform attention model.

4 NEED FOR TWO-SIDED FAIRNESS IN PERSONALIZED RECOMMENDATIONS

Customer-centric recommendations can concentrate producer exposure, while producer-centric balancing can harm customer utility and create unequal customer losses. These trade-offs motivate two-sided fairness.

  • Standard personalized recommendation learns relevance scores and recommends the top-k products for each customer.
  • 4.1 Datasets and Models: Experiments use latent-factorization and dataset-specific relevance models, including Google Local’s ratings and distance-based scoring.
  • The basic setting assumes no position bias and similar customer attention across the k recommended products.
  • 4.2 Adverse Impact of Customer-Centric Recommendation: For top-k recommendations, the least-exposed 50% of producers receive 32%, 5%, and 11% of total exposure in GL-CUSTOM, GL-FACT, and LF, respectively.
  • 4.3 Pitfalls of Naive Solution: The poorest-k strategy nearly equalizes producer exposure but reduces overall customer utility and increases disparity in individual customer utility.
  • Fairness Properties: Producer fairness seeks minimum exposure guarantees, while customer fairness seeks envy-free recommendations up to one item.

5 RE-IMAGINING FAIR RECOMMENDATION AS FAIR ALLOCATION

The paper reframes fair recommendation as constrained fair allocation of indivisible goods, representing products as items, customers as agents, and relevance scores as valuations. The resulting allocation must balance producer exposure guarantees with EF1 fairness for customers.

  • 5.1 Notions of Fairness in Allocation: EF1 relaxes envy-freeness by allowing each customer’s envy to disappear after removing one item from another customer’s bundle.For every pair of agents, some item can be removed from the envied bundle to satisfy the comparison.
  • 5.1 Notions of Fairness in Allocation: MMS fairness requires each agent to receive at least the value she could guarantee by partitioning all items into |U| bundles and taking the least-valued bundle.This threshold is agent-specific and is denoted MMS_u in the paper.
  • 5.2 Fair Recommendation to Fair Allocation: Fair recommendation maps products to items, customers to agents, and relevance scores to valuations in a constrained allocation problem.The recommendation task becomes allocating product copies to customers subject to system constraints.
  • 5.2 Fair Recommendation to Fair Allocation: The construction uses an exposure guarantee E, with E ≤ MMS, to determine the number of copies of each product before allocating them among customers under EF1.Once the product set is fixed, the remaining allocation task targets customer fairness.
  • 5.3 Extending the Conventional Fair Allocation Problem: Producer fairness is modeled by creating at least ℓ copies of each product and allocating every copy, while preventing any customer from receiving duplicate copies.The extension also requires exactly k items to be allocated to each customer, producing hierarchical feasibility constraints.

6 FAIRREC: AN ALGORITHM TO ENSURE TWO-SIDED FAIRNESS

FairRec is a polynomial-time two-phase algorithm that uses modified Greedy-Round-Robin allocation to distribute products while respecting relevance and feasibility constraints. Its first phase seeks EF1 and producer exposure, while its second phase completes each customer’s recommendation set without losing EF1.

  • 6 FAIRREC: AN ALGORITHM TO ENSURE TWO-SIDED FAIRNESS: FairRec operates under k < |P| and |P| ≤ k · |U|, excluding trivial and infeasible recommendation regimes.The algorithm recommends exactly k products per customer in its intended setting.
  • 6 FAIRREC: AN ALGORITHM TO ENSURE TWO-SIDED FAIRNESS: The algorithm runs in two phases: the first seeks EF1 and minimum producer exposure, while the second fills customers to exactly k products while maintaining EF1.The second phase addresses unequal bundle sizes left by the first phase.
  • 6 FAIRREC: AN ALGORITHM TO ENSURE TWO-SIDED FAIRNESS: The first phase initializes each product with ℓ available copies and uses feasible sets to prevent assigning the same product twice to one customer.The copy vector tracks remaining product copies during allocation.
  • 6 FAIRREC: AN ALGORITHM TO ENSURE TWO-SIDED FAIRNESS: The second phase checks the next customer’s bundle size and applies a cyclically rotated customer order when additional allocation is needed.It stores the remaining number of items in T and invokes Greedy-Round-Robin again.
  • 6 FAIRREC: AN ALGORITHM TO ENSURE TWO-SIDED FAIRNESS: FairRec returns an allocation together with residual feasible sets and the last allocated customer index.The underlying Greedy-Round-Robin procedure takes customer and producer counts, copy availability, relevance scores, feasibility sets, and an ordering as inputs.
  • 6 FAIRREC: AN ALGORITHM TO ENSURE TWO-SIDED FAIRNESS: Both phases use modified Greedy-Round-Robin, selecting customers cyclically and assigning each selected customer her most desired feasible unallocated product.Feasibility is maintained through available-copy counts and customer-specific feasible product sets.

7 THEORETICAL GUARANTEES

Theoretical results show that FairRec provides customer-side EF1 while assigning exactly k items per customer, and producer-side non-zero exposure with MMS guarantees for many producers. The algorithm runs in polynomial time, with stronger MMS coverage when k is a small fraction of n.

  • Customer-side guarantees: FairRec’s Greedy-Round-Robin allocation is EF1 and satisfies the cardinality constraint that each customer receives at most one item from each producer.These are among the structural properties used in the fairness proof.
  • Customer-side guarantees: FairRec returns an EF1 allocation among m customers while assigning exactly k items to each customer when k < n ≤ mk.This is the customer fairness guarantee stated in Lemma 7.2.1 and Theorem 7.2.
  • Producer-side guarantees: FairRec ensures non-zero exposure for all n producers and MMS fairness for at least n − k producers.The guarantee holds under k < n ≤ mk, and the second phase preserves the exposure thresholds established in the first phase.
  • Producer-side guarantees: When k ≤ βn for β ∈ (0, 1), FairRec guarantees MMS fairness for at least (1 − β)n producers.Thus, the stated MMS coverage increases as the recommendation size becomes smaller relative to the number of producers.
  • Complexity: FairRec has worst-case time complexity O(mnk).Across two phases it allocates mk items and finds a maximum feasible producer in at most O(n) time per allocation.

8 EXPERIMENTAL EVALUATION

FairRec is evaluated against customer-centric, random, mixed, and producer-centric baselines using producer- and customer-side fairness, utility, disparity, and exposure-loss metrics. With an MMS guarantee, it improves producer fairness while remaining close to top-k in customer utility; higher guarantees expose a fairness–utility trade-off.

  • Producer-Side Results: FairRec and PR-k achieve the best producer satisfaction, while top-k and mixed-k perform worst.FairRec and PR-k seek larger producer exposure, whereas top-k and mixed-k prioritize customer preferences.
  • Producer-Side Results: PR-k has the lowest exposure inequality, while FairRec and random-k perform similarly or slightly worse.Top-k and mixed-k perform worst because they are highly customer-centric.
  • Producer-Side Results: 20% is the maximum reported exposure loss caused by FairRec relative to top-k.Random-k and PR-k incur the highest exposure losses, while mixed-k causes smaller losses.
  • Customer-Side Results: FairRec produces very low customer envy and performs almost at par with top-k in customer utility.Random-k and PR-k generate the highest envy and substantial utility loss because they neglect customer preferences.
  • Customer-Side Results: FairRec and mixed-k show relatively low disparity in customer utilities, whereas random-k and PR-k show large disparities for larger recommendation sizes.Top-k has zero standard deviation because it provides maximum utility to all customers.

9 CONCLUSION

The work presents a scalable, adaptable algorithm for two-sided fairness with only a marginal loss in recommendation quality. It supports offline recommendations and can extend to active-customer online settings, while fully online robustness and position bias remain future work.

  • FairRec provides scalable and adaptable two-sided fairness with a marginal loss in overall recommendation quality.The approach is supported by theoretical guarantees and extensive evaluations on real-world datasets.
  • The offline recommendation setting can extend to online recommendations by limiting recommendations to active customers.A completely online realization remains future work.
  • Future work includes developing a more robust completely online mechanism and attention models for position bias.The latter addresses customers paying more attention to top-ranked products than lower-ranked ones.
Loading 2002.10764v2…