Source-linked AI summary
User-oriented Fairness in Recommendation
Yunqi Li, Hanxiong Chen, Zuohui Fu, Yingqiang Ge, Yongfeng Zhang
TL;DR
The paper investigates user-side unfairness arising from activity-related data imbalance in recommender systems. It groups users by activity and applies fairness-constrained re-ranking, finding reduced group disparity and better overall recommendation performance across datasets and algorithms.
Problem
User activity imbalance can bias collaborative-filtering recommenders toward active users, producing lower-quality recommendations for the less active majority.
Method
The paper groups users by observable activity measures and re-ranks any base recommender’s lists using user-oriented group fairness constraints.
Results
The experiments reduce fairness disparity and improve overall recommendation quality across three Amazon datasets and shallow or deep recommendation algorithms.
Takeaways & Limitations
Fairness-aware re-ranking can benefit disadvantaged users while improving overall recommendation results, although advantaged-user performance may be sacrificed.
Takeaways & Limitations
The maximum-price grouping method may capture activity less informatively than interaction count or total consumption, partly because its group distributions are less distinct.
Abstract
from arXiv · showhide
As a highly data-driven application, recommender systems could be affected by data bias, resulting in unfair results for different data groups, which could be a reason that affects the system performance. Therefore, it is important to identify and solve the unfairness issues in recommendation scenarios. In this paper, we address the unfairness problem in recommender systems from the user perspective. We group users into advantaged and disadvantaged groups according to their level of activity, and conduct experiments to show that current recommender systems will behave unfairly between two groups of users. Specifically, the advantaged users (active) who only account for a small proportion in data enjoy much higher recommendation quality than those disadvantaged users (inactive). Such bias can also affect the overall performance since the disadvantaged users are the majority. To solve this problem, we provide a re-ranking approach to mitigate this unfairness problem by adding constraints over evaluation metrics. The experiments we conducted on several real-world datasets with various recommendation algorithms show that our approach can not only improve group fairness of users in recommender systems, but also achieve better overall recommendation performance.
1 INTRODUCTION
The paper frames user-side fairness as a neglected recommendation problem caused by activity-related data imbalance. It groups users by activity and proposes constrained re-ranking to reduce disparities while improving overall recommendation quality.
- User-side algorithmic bias is under-studied relative to item- or provider-side fairness in recommender systems.
- Collaborative filtering can favor active users because they contribute more training data, leaving inactive users with less satisfactory recommendations.
- The paper identifies activity groups using interaction counts, total consumption, or maximum purchase price, labeling more active users advantaged.
- The proposed model-agnostic re-ranking method imposes user-oriented fairness constraints on recommendation lists from any base algorithm.
- Experiments on three Amazon datasets with four shallow or deep algorithms reduce group disparities and improve overall recommendation quality.
2 RELATED WORK
Prior fairness research emphasizes individual, group, item, or provider perspectives, while recommendation fairness also involves users. The paper situates its contribution in this comparatively less-developed user-side area.
- Individual fairness requires similar individuals to receive similar treatment, whereas group fairness compares treatment across protected and advantaged groups.
- Fairness research in rankings commonly focuses on item-side exposure, but recommender systems extend the problem to multiple stakeholders, including users.
- Prior work on ranking proposes mechanisms for individual attention fairness by addressing position bias across repeated rankings.
- Recommendation fairness studies examine varied stakeholder settings, including explainable recommendation over knowledge graphs and popularity bias.
3 MOTIVATING FAIRNESS CONCERNS
The motivating analysis tests whether user activity imbalance corresponds to unequal recommendation quality. Across Amazon datasets and fairness-unaware baselines, a small active group performs substantially better than the majority.
- The analysis evaluates F1@10 and NDCG@10 for four fairness-unaware algorithms across Beauty, Grocery, and Health datasets.
- Most users have limited interactions or consumption, while highly active or consuming users form a small minority.
- The study labels the top 5% by interactions, total consumption, or maximum purchase price as advantaged and the remainder as disadvantaged.
- Across baselines, the 5% advantaged group receives much higher recommendation quality than disadvantaged users on Grocery, with similar trends elsewhere.
- Because disadvantaged users are the majority, their lower recommendation quality is associated with unfair recommendations and degraded overall performance.
4 THE FRAMEWORK
The framework defines user-oriented group fairness as equal recommendation quality across activity-based user groups and uses fairness-constrained re-ranking to produce fair Top-K lists. The optimization selects items from baseline Top-N recommendations while maximizing preference scores under a fairness constraint.
- 4 THE FRAMEWORK: The framework represents each fair recommendation list with binary variables indicating whether candidate item j is selected for user i.Exactly K of the N candidate items must be selected for each user.
- 4.1 User-oriented Group Fairness: User-oriented group fairness requires equal recommendation quality across disjoint user groups defined by activity level.Recommendation quality can be measured with metrics such as NDCG@K or F1 score.
- 4.1 User-oriented Group Fairness: The ε-fairness constraint bounds the difference between average recommendation performance for the two user groups.Smaller ε imposes stricter fairness but can reduce recommendation performance for the advantaged group.
- 4.2 Fairness-aware Algorithm: Fair re-ranking selects exactly K items from each baseline Top-N list to maximize summed preference scores while satisfying the fairness constraint.The procedure uses preference scores produced by a traditional fairness-unaware recommender.
- 4.2 Fairness-aware Algorithm: The fairness-aware optimization is formulated as a 0–1 integer programming problem and solved using fast heuristics such as the Gurobi solver.The authors note that heuristic solutions may not reach the global optimum but report satisfactory experimental results.
5 EXPERIMENTS
Experiments on three Amazon datasets compare fair re-ranking with four recommendation baselines across user-grouping methods and fairness settings. The method generally reduces group disparities while improving overall and disadvantaged-user performance, though benefits vary by grouping and fairness strictness.
- Experimental setup: Experiments use Beauty, Grocery, and Health Amazon datasets with four baselines: Biased-MF, PMF, NeuMF, and STAMP.The evaluation uses publicly available Amazon 5-core data and compares shallow, deep, and sequential recommendation models.
- Experimental setup: Users are divided into advantaged and disadvantaged groups using interaction count, total consumption, or maximum purchase price.The advantaged group consists of the top 5% of users under each grouping method, with the remainder labeled disadvantaged.
- Main results: Baseline recommendation quality differs substantially between advantaged and disadvantaged users across all three grouping methods.For NeuMF on Grocery, the NDCG@10 gaps are 32.21%, 27.43%, and 13.41%, while the F1@10 gaps are 23.51%, 20.34%, and 8.86%.
- Main results: Fair re-ranking reduces group disparities while improving overall recommendation performance across the evaluated baselines.The experiments report this pattern across the three datasets and grouping methods.
- Main results: For NeuMF on Grocery, fair re-ranking raises overall NDCG@10 from 38.86% to 40.09% and overall F1@10 from 14.51% to 15.48%.The same example reduces the NDCG@10 group gap from 32.21% to 10.85% and the F1@10 gap from 23.51% to 5.82%.
- Ablation study: Stricter fairness requirements improve disadvantaged-group and overall performance but reduce advantaged-group performance, creating a fairness–quality trade-off.Under an F1@10 constraint, the three groups become nearly equal when 𝜖=0; NDCG@10 follows a similar trend.
6 CONCLUSIONS
The paper studies user-oriented fairness in recommender systems, showing that activity-related data imbalance produces unequal recommendation quality. It introduces fairness-constrained re-ranking and reports reduced group unfairness alongside improved overall performance.
- Conclusion: The paper studies unfairness between user groups with different activity levels in recommendation algorithms.It focuses on fairness from the user perspective.
- Conclusion: Users are grouped as advantaged or disadvantaged, unfairness is quantified at group level, and recommendation lists are re-ranked under fairness constraints.The approach is designed to mitigate disparity while maintaining recommendation quality.
- Conclusion: Experiments show significantly lower unfairness between user groups together with improved overall recommendation performance.The conclusion attributes the overall-performance effect partly to less active users being the majority.