Adversarial regularized diffusion model for fair recommendations

Abstract

With the widespread deployment of recommendation systems, concerns have grown over algorithmic fairness and representation bias in recommendation outcomes. Existing debiasing methods primarily suffer from two critical limitations: (1) Explicit feature removal strategies risk eliminating semantic signals entangled with sensitive attributes, inevitably degrading recommendation performance. (2) Conventional adversarial learning frameworks impose rigid gradient reversal to enforce independence from sensitive attributes, yet cause semantic distortion in latent representations through uncontrolled adversarial conflicts between fairness objectives and recommendation goals.

To address these challenges, we propose a fairness-aware recommendation framework leveraging the dynamic equilibrium of diffusion model. During the forward diffusion process, we introduce adaptive gradient-aware noise injection, where fairness discriminators from the reverse denoising process guide Gaussian perturbations through their aggregated gradient statistics, achieving feature-aware bias dissociation while preserving user interest semantics. The reverse denoising process employs adversarial regularization with sensitivity-aware gradient constraints, iteratively purifying recommendation-oriented embeddings through alternating optimization of denoising prediction and fairness discrimination tasks. To further enhance fairness-utility tradeoffs, we design an interest fusion mechanism at denoising initialization and develop a bias-controlled rounding function for candidate generation. Extensive experiments on three real-world datasets with sensitive attributes demonstrate that our model outperforms state-of-the-art methods in recommendation accuracy and fairness. We publish the source code at https://github.com/YangRan993/DiffuFair.