.png){kind=icon}
.png){kind=icon}
Return
FairDiff: Masked condition diffusion for fairness-aware recommendation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.131720.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The investigation of fairness in recommendation systems has emerged as a vital research area because of the substantial influence of item exposure on user preferences. Nevertheless, current implicit collaborative filtering models often overlook fairness considerations while fairness-aware recommendation methods struggle to strike a balance between fairness and utility. To address this issue, we propose FairDiff, a novel approach that utilizes masked condition diffusion to tackle fairness concerns in recommendation systems. FairDiff harnesses the robust generative capabilities of diffusion models for collaborative filtering by incorporating an enhanced fairness loss. This loss resolves the non-differentiability problem in fairness metrics by innovatively employing an activation function for approximating the ranking function. Furthermore, we significantly optimize the computational efficiency of fairness loss through the utilization of a random sampling method. Our approach aims to maximize recommendation utility by optimizing both utility loss and fairness loss. In addition, we enhance fine-grained control over fairness by the manipulation of the fairness threshold and incorporating fairness loss as a constraint. Extensive experiments conducted on three real-world datasets demonstrate that FairDiff outperforms the state-of-the-art methods. Codes are available at https://anonymous.4open.science/status/FairDiff-E84C 1.
Keywords:
Fairness-aware recommendation
Diffusion models
Collaborative filtering
Fairness loss
Recommendation utility
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W