Ensuring User-side Fairness in Dynamic Recommender Systems

TL;DR

The paper investigates user-side fairness in dynamic recommender systems and identifies distribution shifts, frequent updates, and non-differentiable ranking metrics as primary obstacles. It proposes FADE, an end-to-end framework that combines incremental fine-tuning with a differentiable fairness loss driven by Differentiable Hit (DH), enabling adaptive mitigation of disparities over time. Theoretical results favor incremental fine-tuning with restart, and extensive experiments on Movielens and ModCloth show substantial reductions in performance disparity (≈48% on average) with minimal loss to overall recommendation quality, alongside strong time efficiency. This work advances fair dynamic recommendation by delivering a practical, scalable method that maintains user-level equity during continual model updates, with broad implications for two-sided fairness in evolving systems.

Abstract

User-side group fairness is crucial for modern recommender systems, aiming to alleviate performance disparities among user groups defined by sensitive attributes like gender, race, or age. In the ever-evolving landscape of user-item interactions, continual adaptation to newly collected data is crucial for recommender systems to stay aligned with the latest user preferences. However, we observe that such continual adaptation often exacerbates performance disparities. This necessitates a thorough investigation into user-side fairness in dynamic recommender systems, an area that has been unexplored in the literature. This problem is challenging due to distribution shifts, frequent model updates, and non-differentiability of ranking metrics. To our knowledge, this paper presents the first principled study on ensuring user-side fairness in dynamic recommender systems. We start with theoretical analyses on fine-tuning v.s. retraining, showing that the best practice is incremental fine-tuning with restart. Guided by our theoretical analyses, we propose FAir Dynamic rEcommender (FADE), an end-to-end fine-tuning framework to dynamically ensure user-side fairness over time. To overcome the non-differentiability of recommendation metrics in the fairness loss, we further introduce Differentiable Hit (DH) as an improvement over the recent NeuralNDCG method, not only alleviating its gradient vanishing issue but also achieving higher efficiency. Besides that, we also address the instability issue of the fairness loss by leveraging the competing nature between the recommendation loss and the fairness loss. Through extensive experiments on real-world datasets, we demonstrate that FADE effectively and efficiently reduces performance disparities with little sacrifice in the overall recommendation performance.

Figures (20)

• Figure 1: Even though incremental fine-tuning with new data (red curve) upholds recommendation performance compared to pretrain (black curve), the disparity gradually expands over time without fairness regularization. (See §\ref{['sec:experiment']} for detail.)
• Figure 2: The trade-off between recommendation performance (NDCG@20 & F1@20) and absolute performance disparity $|{\operatorname{PD}}|$ of eight compared methods and FADE in Task-R. Employing our fairness loss leads to a substantial reduction in $|{\operatorname{PD}}|$ across all cases, with a modest impact on overall performance. Note that the optimal point should be situated in the bottom-right corner.
• Figure 3: The trend of the absolute performance disparity ($|{\operatorname{PD}}|$) in Task-R. Without the fairness loss, the $|{\operatorname{PD}}|$ is relatively high and often increase, while with the fairness loss, particularly in FADE, the $|{\operatorname{PD}}|$ tends to remain relatively low.
• Figure 4: The effect of the scaling parameter $\lambda$ on the performance of the advantaged and disadvantaged groups.
• Figure 5: Effect of hyperparamters.

Theorems & Definitions (12)

• Definition 1: User-side performance disparity li2021user
• theorem 1: Fine-tuning
• theorem 2: Retraining
• proposition 1
• Definition 1: Subgaussian property
• lemma 1
• proof : Proof of Lemma \ref{['lem:alpha']}
• corollary 1
• proof : Proof of Corollary \ref{['cor:alpha']}
• proof : Proof of Theorem \ref{['thm:ft']}