Mitigating long-tail bias in recommendations via graph diffusion

Zhuoyang Xia, Meng Jian, Zihan Liu, Yulong Bai, Lifang Wu, Shaona Wang

Abstract: Personalized recommendation systems are pivotal in helping users navigate the expansive digital information landscape. However, these systems often suffer from a long-tail bias that skews recommendations toward popular items, diminishing user experience diversity. Traditional debiasing approaches, such as self-supervised learning models, typically rely on data augmentation and negative sampling strategies that risk diverting attention away from genuine user preferences. To address these limitations, we propose a novel Graph Propagation-based Diffusion (GPD) model that redefines debiased recommendations. Unlike methods relying on negative sampling, GPD refines item embeddings through iterative diffusion processes, naturally balancing the representation of popular and long-tail items without task-irrelevant data manipulations. This approach preserves authentic preference signals while systematically promoting recommendation fairness and accuracy. This work underscores the importance of adapting recommendation systems to reflect evolving user preferences while maintaining equitable treatment across diverse items. Comprehensive evaluations of public datasets reveal that the GPD model outperforms existing graph propagation and contrast-based models in delivering superior personalized recommendation quality.

External IDs:dblp:journals/mms/XiaJLBWW25