TFROM: A Two-sided Fairness-Aware Recommendation Model for Both Customers and Providers

1 INTRODUCTION

• Fairness approaches that protect one stakeholder can disadvantage the other, motivating joint consideration of customer quality and provider exposure.Prior work reports that guaranteeing fairness in customers’ recommendation quality can make provider exposure unfair [29].Source: §1 · p. 1
• Higher-ranked products attract more attention and can bring providers more orders and revenue, while provider fairness is presented as important to a healthy market.Source: §1 · p. 1
• Fairness measures can lower customer satisfaction by diverting recommendations from customers’ preferences, making fairness and personalization difficult to maintain together.Source: §1 · p. 1
• Prior work [29] treats each item as a provider and ignores rank-dependent exposure, whereas TFROM introduces exposure metrics that account for item position.The paper notes that providers often offer multiple items, unlike the one-item-per-provider setting in [29].Source: §1 · p. 2
• TFROM formulates two-sided fairness, measures customer individual fairness and provider group fairness, and reports better fairness with higher recommendation quality than comparison algorithms on three real-world datasets.The paper describes offline and online versions of the model.Source: §1 · p. 2

2 RELATED WORK

• Prior work separates customer-side fairness from provider-side fairness, with examples including protected-attribute mitigation for customers and fair exposure or access for providers.Examples include reranking for gender or age bias, adversarial filters [8], exposure amortization [5], and fair taxi routing.Source: §2 · p. 2
• Two-sided fairness has been studied for driver income and user waiting time, user fairness, item fairness, and diversity [36], and fair customer recommendations and provider exposure [29].The paper distinguishes its setting from [29] because each provider can supply one or more items.Source: §2 · p. 2
• Fairness of recommendations is categorized as individual or group fairness, while mechanisms are classified by when they work in the system as pre-processing or in-processing.Individual fairness aims for similar users to be treated similarly; group fairness aims to give disadvantaged groups the same opportunities as advantaged groups [7] [32] [4] [14].Source: §2 · p. 2
• This study targets individual fairness among customers and group fairness among providers, using post-processing to adjust existing recommendation results.Source: §2 · p. 3

3 PROBLEM FORMULATION

• The system includes customers receiving recommendations and providers supplying one or more recommended items or services.It assumes an existing algorithm supplies predicted ratings and original lists, which the proposed algorithm uses to produce fairer results aligned with preferences.Source: §3 · p. 3

3.1 Notations

• The notation distinguishes customer, item, and provider sets, provider-specific item sets, and original versus final recommendation lists.The original lists are based on the predicted rating matrix; the final lists are the outputs to customers.Source: §3.1 · p. 3

3.2 Exposure of items and providers

• Provider exposure depends on item rank, and a provider’s total exposure aggregates the exposures of the items it supplies.Lower-ranked items receive lower weights because attention drops quickly at first and then stabilizes.Source: §3.2 · p. 3

3.3 Fairness in Providers’ Exposure

• The model treats provider fairness as dispersion in exposure, considering both equal exposure and exposure proportional to item quality or provider catalog size.Uniform fairness allocates exposure proportional to the number of items offered; quality-weighted fairness uses their summed quality scores.Source: §3.3 · p. 3

3.4 Quality of Recommendation

• Recommendation quality is the match between the output list and the original list, measured with DCG and NDCG.NDCG is normalized against the original list; a value of 1 indicates a full match to customer preference.Source: §3.4 · pp. 3–4

3.5 Measuring the Fairness of Recommendation for Customers

• Customer-side fairness means customers receive equal NDCG, with variance across customers measuring overall fairness.The goal is to distribute any recommendation-quality reduction as evenly as possible.Source: §3.5 · p. 4

3.6 Trade-off between Customer Benefits and Provider Benefits

• Improving provider exposure fairness can reduce recommendation quality, and customer-side fairness may decrease or remain unchanged as quality falls or provider fairness rises.The original recommendation list gives customers NDCG of 1 and equal quality; distributing quality loss evenly can preserve customer fairness, though this special case may not always exist.Source: §3.6 · p. 4
• The optimization strategy sacrifices some recommendation quality to improve provider exposure fairness while distributing that loss as evenly as possible among customers.The design follows the stated trade-off among recommendation quality, customer fairness, and provider exposure fairness.Source: §3.6 · p. 4

4 A TWO-SIDED FAIRNESS-AWARE RECOMMENDATION MODEL (TFROM)

• The fairness problem reduces to a knapsack problem, and TFROM uses heuristics for offline batch recommendations and online requests.The paper characterizes the knapsack problem as nondeterministic polynomial complete and notes that adding recommendation quality makes the optimization more complicated.Source: §4 · p. 4

4.1 TFROM for Offline Scenario

• Provider fair-exposure benchmarks use providers’ item counts, while customer fairness spreads recommendation-quality losses and prioritizes relevant items.The benchmarks guide exposure balancing; the text notes that exact equality with ideal exposure may be impossible because list positions provide fixed exposure values.Source: §4.1 · pp. 4–5
• Offline recommendations select items in ranked positions while respecting provider exposure benchmarks, then reallocate skipped positions using items from the least-exposed providers.Customers are ordered by recommendation quality for later positions, and the original recommendation lists guide item selection.Source: §4.1 · p. 5

4.2 TFROM for Online Scenario

• Online TFROM targets long-term fairness by accumulating provider exposure and customer recommendation quality, with customer quality averaged over recommendation-service count.Provider fair exposure must be recalculated as request volume changes; when filling vacancies, the algorithm prioritizes preference scores to avoid sharply reducing the current customer’s recommendation quality.Source: §4.2 · p. 6

4.3 Time complexity

• TFROM-offline has worst-case time complexity O(knm^2 log(m)), while TFROM-online has worst-case complexity O(2kn).Offline selection sorts customers and may traverse their original lists; online processing handles one customer at a time.Source: §4.3 · p. 6

5.1 Datasets and Metrics

• The evaluation covers Ctrip flight, Google local-business, and Amazon review datasets spanning travel, local living, and online shopping.The authors state that their code and datasets are released on Zenodo.Source: §5.1 · pp. 6–7

5.1.1 Ctrip Flight Dataset.

• The Ctrip dataset contains 3,814 customers, 6,006 ticket items, and 25,190 orders on the Shanghai–Seoul route from 2017 to 2020.Tickets are grouped by airline, class, and departure time, with airlines treated as providers; original recommendation lists and a preference matrix are generated using collaborative filtering [16].Source: §5.1.1 · p. 7

5.1.2 Google Local dataset.

• The filtered California Google dataset contains 3,335 users, 4,927 businesses, and 97,658 reviews, with each business treated as a provider offering one item.Businesses and customers with fewer than 10 reviews are excluded; the one-item provider setting is described as a special case of the paper’s problem.Source: §5.1.2 · p. 7

5.1.3 Amazon Review dataset.

• The filtered Amazon Clothing, Shoes and Jewelry dataset contains 1,851 users, 7,538 items, 161 providers, and 24,658 reviews.Items are grouped randomly into providers of 1–100 items because provider information is unavailable in the dataset.Source: §5.1.3 · p. 7

5.1.4 Metrics.

• Provider fairness is measured by exposure variance and normalized exposure-to-relevance variance, while customer fairness and overall quality use NDCG variance and sum, respectively.Lower customer NDCG variance indicates fairer results, while a greater NDCG sum indicates less recommendation-quality loss.Source: §5.1.4 · p. 7

5.2.1 Top-k.

• Top-k recommends the highest-ranked k items from each customer’s original list, corresponding to maximizing recommendation quality.Source: §5.2.1 · p. 7

5.2.2 All random.

• All random selects k items at random from each customer’s original recommendation list.Source: §5.2.2 · p. 7

5.2.4 FairRec.

• FairRec [29] uses a greedy strategy for customer fairness and selects items from the least-exposed providers to pursue uniform provider exposure.The described approach sets minimum provider exposure.Source: §5.2.4 · p. 7

5.2.5 An ILP-based fair ranking mechanism.

• The ILP-based fair-ranking mechanism [5] minimizes the absolute difference between cumulative values while constraining recommendation quality.It is designed to ensure quality-weighted fairness of provider exposure.Source: §5.2.5 · p. 7

5.3.1 Results for the offline situation.

• Less than 10% quality loss in most offline cases; TFROM-offline spreads the loss across items, while Top-k maximizes recommendation quality.The reported quality loss may be imperceptible to customers.Source: §5.3.1 · p. 7, Figure 1 · p. 8, Figure 2 · p. 8
• TFROM-offline achieves high customer fairness, while Minimum exposure’s apparent fairness is associated with very low recommendation quality.The reported customer-fairness results cover all three datasets.Source: §5.3.1 · pp. 7–8, Figure 1 · p. 8, Figure 2 · p. 8
• TFROM-offline-Uniform, Minimum exposure, and All random maintain uniform provider exposure fairness as k grows; TFROM-offline-Quality-Weighted provides better, more stable fairness across datasets.Top-k’s provider inequality increases with k; FairRec is less effective when items are aggregated by provider.Source: §5.3.1 · p. 8, Figure 1 · p. 8, Figure 2 · p. 8

5.3.2 Results for the online situation.

• Online recommendation quality grows roughly linearly with requests; TFROM-online incurs a small loss, larger for Uniform fairness, while Top-k achieves the highest quality.The Uniform-fairness loss accumulates across requests but is described as acceptable for a single recommendation.Source: §5.3.2 · p. 9, Figure 4 · p. 9, Figure 5 · p. 9
• Online customer fairness first decreases and then increases as requests arrive; TFROM-online reaches a good fairness level once recommendations accumulate.The reported results indicate that fairness stabilizes after all customers receive at least one recommendation.Source: §5.3.2 · p. 9, Figure 4 · p. 9, Figure 5 · p. 9
• Without exposure interventions, Top-k’s provider unfairness continues to rise online; TFROM-online-Quality-Weighted provides more consistent and fair results than comparison algorithms.The reported pattern is consistent with the offline results.Source: §5.3.2 · p. 9, Figure 4 · p. 9, Figure 5 · p. 9

• On three real-world datasets, TFROM provides better two-sided fairness than comparison algorithms while losing only a little recommendation quality.Source: §6 · p. 10
• TFROM uses post-processing heuristic algorithms for offline and online scenarios, making the approach applicable to existing recommendation systems.Source: §6 · p. 10
• The model ensures provider fairness at the group level and customer fairness between individual customers while aiming to maintain high personalization.Source: §6 · p. 10