WhenVariety Seeking Meets Multi-Sided Recommendation Fairness: A Consistent and Personalized Multi-Objective Optimization Framework

Jiayi Guo, Jiangning He, Chenyan Wang, Xinran Wu

Abstract:

Recommendation research has evolved from solely improving accuracy to addressing ethical and fairness concerns. While prior works focus on optimizing fairness from either the user or product perspective, recent research emphasizes the importance of multi-sided
fairness. This issue is inherently challenging due to the competing goals of different stakeholders. To tackle this challenge, we propose
a Consistent and Personalized Fairness Recommendation framework with Multi-Objective Integer Programming (CPFR-MOIP). Our framework introduces two key innovations. First, we develop a novel similarity-based individual fairness metric for the user side and formulate a consistent product-side fairness metric, ensuring that the generated recommendation list aligns with the user prefer
ence distribution and the expected product exposure distribution. Second, we incorporate users’ variety-seeking levels as a moderating factor to adjust fairness trade-offs and introduce personalized weights to balance user-side and product-side fairness. To effectively solve this optimization problem, we devise an alternating algorithm with theoretical guarantee and demonstrate the Pareto optimality of the obtained solutions. Extensive experiments on two real-world datasets demonstrate that our CPFR-MOIP achieves
superior multi-sided fairness while maintaining competitive recommendation accuracy. Furthermore, ablation analysis highlights the advantages of incorporating user variety-seeking levels for personalizing fairness trade-offs. Our work paves the way for more ethical
and personalized recommendation systems. The implementation code is available at: https://github.com/P0ise-Wang/CPFR-MOIP.