Source-linked AI summary
CPFair: Personalized Consumer and Producer Fairness Re-ranking for Recommender Systems
Mohammadmehdi Naghiaei, Hossein A. Rahmani, Yashar Deldjoo
TL;DR
Fairness research in recommender systems often treats consumer and producer concerns separately, despite systems operating as marketplaces. This paper proposes a joint fairness re-ranking method and reports improved fairness and recommendation quality across eight datasets.
Problem
Fairness research often focuses on consumer or producer fairness separately, although recommender systems connect both stakeholder groups in a marketplace.
Source: §1 · pp. 1–2Method
The paper proposes a re-ranking approach that jointly optimizes consumer and producer fairness objectives.
Source: §1 · p. 2Results
Across eight datasets, the method reduces unfair outcomes for consumers and providers while improving overall recommendation quality.
Source: §7 · p. 10Takeaways & Limitations
The results demonstrate that algorithmic fairness may play a critical role in addressing unfair outcomes affecting consumers and providers.
Source: §7 · p. 10The method segments consumer and provider groups using internal attributes based on interaction counts.
Source: §2 · p. 3
Abstract
from arXiv · showhide
Recently, there has been a rising awareness that when machine learning (ML) algorithms are used to automate choices, they may treat/affect individuals unfairly, with legal, ethical, or economic consequences. Recommender systems are prominent examples of such ML systems that assist users in making high-stakes judgments. A common trend in the previous literature research on fairness in recommender systems is that the majority of works treat user and item fairness concerns separately, ignoring the fact that recommender systems operate in a two-sided marketplace. In this work, we present an optimization-based re-ranking approach that seamlessly integrates fairness constraints from both the consumer and producer-side in a joint objective framework. We demonstrate through large-scale experiments on 8 datasets that our proposed method is capable of improving both consumer and producer fairness without reducing overall recommendation quality, demonstrating the role algorithms may play in minimizing data biases.
1 INTRODUCTION
Recommender systems affect users’ decisions and businesses, yet fairness research often addresses consumer and producer concerns separately. CP-fairness jointly addresses both sides, and the reported comparison shows improved fairness alongside higher accuracy and novelty.
1 INTRODUCTION
- Recommender systems influence consumed content, beliefs, decisions, and business success, including in high-stakes contexts such as job search and loan applications.The paper frames these impacts as motivation for considering fairness in recommendation.Source: §1 · p. 1
- Accuracy rises from 0.2832 to 0.2868 under CP-fairness, which also improves consumer and producer fairness and increases novelty through more long-tail recommendations.Producer-only fairness reduces accuracy from 0.2832 to 0.2756 and worsens DCF from 0.3280 to 0.3234.Source: Figure 1 · p. 1, §1 · p. 2
- Fairness research often targets consumer or producer concerns individually, while fewer than 10% of surveyed studies address both despite their shared marketplace setting.The survey reports 49.1% focusing on consumer fairness and 41.8% on producer fairness.Source: §1 · pp. 1–2
- The paper proposes model-agnostic re-ranking to jointly provide fair producer exposure and reduce disparate treatment of consumer groups without considerable sacrifice in total accuracy.The introduction describes consumer fairness through relevance and producer fairness through exposure.Source: §1 · pp. 1–2
- The proposed framework formalizes both stakeholder objectives as integer programming and uses an efficient greedy algorithm to find an optimal trade-off in polynomial time.The paper reports experiments on eight real-world datasets with explicit and implicit feedback.Source: §1 · p. 2
2 BACKGROUND AND RELATED WORK
Prior work examines consumer, producer, and two-sided fairness, but much research focuses on a single stakeholder. This paper uses interaction counts to segment both sides and proposes a joint fairness algorithm.
2 BACKGROUND AND RELATED WORK
- Only 9.1% of screened recent fairness publications address multi-sided fairness, and 3.6% propose a two-sided fairness-aware algorithm.The cited survey screening concerns papers published in fairness during the past two years.Source: §2 · p. 3
- Consumer-fairness work includes individual and group approaches, while producer-fairness work addresses equitable exposure and popularity constraints; prior work also studies changing popularity over time.The passage cites individual fairness, group fairness, and producer-side approaches, [15], and.Source: §2 · p. 3
- Fairness groups may be segmented using internal interaction-derived attributes or external protected attributes; this work uses interaction counts for consumers and providers.Source: §2 · p. 3
- The paper’s approach jointly addresses user-group and item-group fairness objectives through optimization intended to increase overall system accuracy.The related-work discussion distinguishes metric development, fair-model frameworks, and data modification as approaches in the literature.Source: §2 · p. 3
3 MOTIVATING CP-FAIRNESS CONCERNS
The motivating analysis finds uneven interaction distributions and recommendation quality across user and item groups in Epinion and Gowalla. Active users receive higher quality recommendations, while popular items account for a large share of recommendations.
3 MOTIVATING CP-FAIRNESS CONCERNS
- The motivating dataset analysis focuses on Epinion and Gowalla, with the authors reporting similar patterns in other datasets’ baseline models.Source: §3 · p. 3
- Most users have few interactions, and most items are long-tail, motivating broader long-tail coverage and engagement among users with fewer interactions.Source: §3 · p. 3
- Active users receive much higher recommendation quality than inactive users, while recommendations disproportionately include short-head items despite the much larger long-tail item group.For NeuMF and VAECF, active-user performance is almost 1.0 and inactive-user performance is almost 0.6; the overall figure is only slightly above 0.6.Source: §3 · p. 4, Figure 3 · p. 4
4 PROPOSED CP-FAIRNESS ALGORITHM
The section defines consumer and producer fairness as parity deviations, then formulates a post-hoc re-ranking approach that balances recommendation relevance with both fairness objectives. Its chosen fairness targets and exposure scope constrain what kinds of fairness it addresses.
4.1 Marketplace Fairness Objectives
- The framework adopts parity between groups as its fairness target, leaving exposure or relevance proportionality targets for future investigation.The target distribution is an application-dependent design choice.Source: §4.1 · p. 4
4.1.1 Fairness constraint on producer side.
- Producer fairness is measured as the exposure-parity deviation between short-head and long-tail item groups, and this work considers binary exposure only.Other producer fairness constraints, including relevance-based fairness, are left for future work.Source: §4.1.1 · p. 4
4.1.2 Fairness constraint on the consumer side.
- Consumer fairness measures the relevance-parity deviation between frequent and less-frequent users; DCF and DPF approaching zero indicate parity but may coincide with lower overall accuracy.The user groups are based on activity level, and relevance can be measured using metrics such as nDCG, F1, or Recall.Source: §4.1.2 · p. 5
4.2 Proposed Algorithm
- The post-hoc objective maximizes total relevance while penalizing consumer and producer parity deviations, with λ1 and λ2 controlling their relative weights.Parity is not enforced as a hard constraint, to avoid significantly diminishing total accuracy; setting both weights to zero recovers the baseline list.Source: §4.2 · p. 5
- The Fair Re-ranking Greedy algorithm solves the MILP optimally in polynomial time by reducing it to a fixed-size, equal-weight knapsack selection; its worst-case complexity is O(n × N).The formulation selects exactly K items per user from the baseline top-N list; the section also notes that the general MILP is NP-hard.Source: §4.2 · p. 5
5 EXPERIMENTAL METHODOLOGY
The experiments evaluate the re-ranking approach on eight public datasets and compare it with unilateral fairness variants and multiple recommendation baselines. The methodology defines user and item groups, a two-sided fairness metric, and parameter selection using fairness and accuracy.
Experimental Setup
- The project reports publicly releasing its code and dataset and using the Cornac open-source recommendation toolkit for evaluation.The passage also describes the study as aiming to foster reproducibility.Source: §5 · p. 6
Datasets
- The study uses eight public datasets spanning movie, point-of-interest, music, book, e-commerce, and opinion domains, with k-core preprocessing to ensure sufficient training feedback.Examples include MovieLens, Epinion, Gowalla, and LastFM.Source: §5 · p. 6
Evaluation Method
- A 70/10/20 train-validation-test split is used, with active users defined as the top 5% by training activity and popular items as the top 20% by training interactions.mCPF is a designer-weighted average of DPF and DCF, and lower values indicate fairer models; λ1 and λ2 are selected to maximize average mCPF and nDCG.Source: §5 · p. 6
Baselines and Evaluation Settings
- The comparisons include PF, WMF, NeuMF, and VAECF, plus Random and MostPop baselines, with CP re-ranking evaluated against unilateral C- and P-fairness variants.Baseline settings follow their original papers; the experiments use 50-dimensional embeddings and select models based on validation performance.Source: §5 · pp. 6–7
6 RESULTS
Across eight datasets, CP-fairness achieves lower mCPF than baselines and one-sided fairness methods while maintaining competitive nDCG@10. Results also show that unilateral fairness can trade off stakeholder outcomes, while CP-fairness performance depends on dataset and parameter choices.
6 RESULTS
- Across four baseline CF models, aggregated averages indicate inherent consumer and supplier biases, with Epinion substantially worse than Gowalla: mean DCF values are 0.3797 and 0.2205, and mean DPF values are 0.8404 and 0.4576.Source: §6 · p. 7
- NeuMF attains the best Epinion accuracy (0.0447) while mCPF rises from 0.621 to 0.7127; VACEF provides the strongest fairness-accuracy balance across the two datasets.VACEF ranks second on Epinion and first on Gowalla for this balance; NeuMF's performance varies substantially with dataset characteristics.Source: §6 · p. 7
- P-fairness more often harms user fairness than C-fairness; active users are only 5% of users, limiting user-relevance optimization's impact on item-group exposure.In Epinion, P-fairness changes DCF from 0.3797 to 0.3877, while C-fairness changes DPF from 0.8404 to 0.8382; on Gowalla, P-fairness changes DPF from 0.45760 to 0.45525.Source: §6 · p. 7
- mCPF improvement over N averaged 3.83%, 9.25%, and 12.84% for C, P, and CP on Epinion, versus 10.36%, 31.28%, and 46.90% on Gowalla.Source: §6 · p. 7
- Under CP-fairness, average mCPF/All fell from 18.73 to 16.53 on Epinion and from 7.31 to 4.51 on Gowalla versus N, without sacrificing baseline accuracy.Source: §6 · pp. 7–8
- Across all eight datasets, CP-fairness achieves lower mCPF than baselines and unilateral fairness methods, with competitive overall nDCG@10.Figure 4 reports mCPF distributions and overall nDCG@10 for the methods.Source: §6 · p. 9, Figure 4 · p. 9
- λ1 is more accuracy-centric and has little effect on item exposure, whereas λ2 is exposure-centric and can affect both beyond-accuracy outcomes and accuracy.Excessive λ2 may not improve overall accuracy because many long-tail items lack sufficient preference interactions; choosing parameters requires balancing fairness and marketplace utility.Source: §6 · pp. 9–10
7 CONCLUSIONS
The paper concludes that personalized CP-fairness re-ranking can mitigate unfair outcomes for consumers and providers while maintaining recommendation quality. Experiments across eight datasets show improved overall recommendation quality.
7 CONCLUSIONS
- Across eight datasets, personalized CP-fairness constrained re-ranking reduces unfair outcomes for consumers and providers while improving overall recommendation quality.Source: §7 · p. 10