Ensuring User-side Fairness in Dynamic Recommender Systems

User-side group fairness is crucial for modern recommender systems, aiming toalleviate performance disparities among user groups defined by sensitiveattributes like gender, race, or age. In the ever-evolving landscape ofuser-item interactions, continual adaptation to newly collected data is crucialfor recommender systems to stay aligned with the latest user preferences.However, we observe that such continual adaptation often exacerbatesperformance disparities. This necessitates a thorough investigation intouser-side fairness in dynamic recommender systems, an area that has beenunexplored in the literature. This problem is challenging due to distributionshifts, frequent model updates, and non-differentiability of ranking metrics.To our knowledge, this paper presents the first principled study on ensuringuser-side fairness in dynamic recommender systems. We start with theoreticalanalyses on fine-tuning v.s. retraining, showing that the best practice isincremental fine-tuning with restart. Guided by our theoretical analyses, wepropose FAir Dynamic rEcommender (FADE), an end-to-end fine-tuning framework todynamically ensure user-side fairness over time. To overcome thenon-differentiability of recommendation metrics in the fairness loss, wefurther introduce Differentiable Hit (DH) as an improvement over the recentNeuralNDCG method, not only alleviating its gradient vanishing issue but alsoachieving higher efficiency. Besides that, we also address the instabilityissue of the fairness loss by leveraging the competing nature between therecommendation loss and the fairness loss. Through extensive experiments onreal-world datasets, we demonstrate that FADE effectively and efficientlyreduces performance disparities with little sacrifice in the overallrecommendation performance.