A Personalized Framework for Consumer and Producer Group Fairness Optimization in Recommender Systems

TL;DR

The paper tackles fairness in two-sided recommender systems by addressing both consumer and producer disparities in a unified framework. It introduces CP-FairRank, a model-agnostic re-ranking method that optimizes a joint objective combining total relevance with penalties for consumer ($DCF$) and producer ($DPF$) fairness, controlled by hyperparameters $\lambda_1$ and $\lambda_2$, under the constraint that each user receives exactly $K$ items. The authors formalize marketplace fairness objectives, define parity-based consumer and producer groups, and provide a scalable algorithm that includes a greedy re-ranking variant and a relaxed LP-based solver, with an overall $O(n\times N)$ time footprint. Extensive experiments across eight real-world datasets and four CF models show that CP-FairRank improves the joint fairness metric $mCPF$ while maintaining or enhancing accuracy and novelty, and that unilateral fairness approaches can harm the non-targeted stakeholder. The study demonstrates that jointly optimizing for consumer and producer fairness can mitigate underlying data biases without substantially sacrificing utility, offering practical implications for fair two-sided marketplaces and guiding future extensions to broader fairness notions and larger-scale deployments.

Abstract

In recent years, there has been an increasing recognition that when machine learning (ML) algorithms are used to automate decisions, they may mistreat individuals or groups, with legal, ethical, or economic implications. Recommender systems are prominent examples of these machine learning (ML) systems that aid users in making decisions. The majority of past literature research on RS fairness treats user and item fairness concerns independently, ignoring the fact that recommender systems function in a two-sided marketplace. In this paper, we propose CP-FairRank, an optimization-based re-ranking algorithm that seamlessly integrates fairness constraints from both the consumer and producer side in a joint objective framework. The framework is generalizable and may take into account varied fairness settings based on group segmentation, recommendation model selection, and domain, which is one of its key characteristics. For instance, we demonstrate that the system may jointly increase consumer and producer fairness when (un)protected consumer groups are defined on the basis of their activity level and main-streamness, while producer groups are defined according to their popularity level. For empirical validation, through large-scale on eight datasets and four mainstream collaborative filtering (CF) recommendation models, we demonstrate that our proposed strategy is able to improve both consumer and producer fairness without compromising or very little overall recommendation quality, demonstrating the role algorithms may play in avoiding data biases.

Figures (9)

• Figure 1: (a) The percentage of the research studied different aspects of fairness in recommender systems. (b) N (Fairness-unaware), C (user-oriented), P (item-oriented), and CP (two-sided fairness-performance) of recommendation algorithms on the accuracy, novelty, and fairness performance. Consumer fairness evaluation (DCF) and producer fairness evaluation (DPF) are the two metrics that make up the mCPF. Note that CP represents the core of our contribution. The results on each bar show the average of 32 experiments across datasets and baseline CF models. $\uparrow$ means the higher the better while $\downarrow$ means the lower the better. Acc., Nov., and mCPF refer to accuracy, novelty, and consumer-producer fairness evaluation, respectively.
• Figure 2: The difference between short-head and long-tail item groups on item exposure and novelty.
• Figure 3: Figures (a) and (d) show the difference between active and inactive user groups and Figures (e) and (d) indicate the difference between mainstream and non-mainstream user groups on normalized nDCG@10.
• Figure 4: Correlation plots were produced for the best performing models (in terms of accuracy), NeuMF, and VACEF, to illustrate the potential trade-off between accuracy and fairness metrics. Noting that the points in this example represent datasets, the correlation measures the generalizability of the findings across domains.
• Figure 5: Distributions of mCPF for fairness-unaware (N) and fairness-aware methods (i.e., C, P, and CP) on all 8 datasets. The numbers on top of each plot show the overall performance (i.e., nDCG@10 for all users) according to each fairness methodology.

Theorems & Definitions (4)

• lemma 1
• definition 1
• definition 2
• definition 3