Debiased recommendation via diffusion-guided implicit graph augmentation with long-tail awareness

Implicit feedback is widely adopted in recommender systems because it is easy to collect and provides broad behavioral coverage. However, observed implicit interactions may contain noise, suffer from popularity bias, and exhibit high sparsity, whereas unobserved user-item pairs cannot be directly treated as negative feedback. These issues constrain the model's ability to learn reliable user preferences and provide effective recommendations for long-tail items. Therefore, this study proposes a latent Preference Diffusion-based interaction Completion Graph Recommendation model, termed PDCGR. Specifically, we first design a graph-guided latent preference recovery process that exploits graph structural information to guide latent-space preference diffusion, thereby improving the robustness of user preference representations under sparse implicit feedback. Second, we develop a latent preference diffusion objective integrated with spectral regularization to prevent the recovered representations from becoming excessively concentrated along a few dominant singular directions, thereby reducing the influence of popularity-dominated patterns. Furthermore, we construct a diffusion-guided interaction completion mechanism that selects high-confidence candidate interactions from strictly unobserved user-item pairs and provides additional relations for subsequent graph propagation and recommendation learning. Experimental results on three public datasets demonstrate that PDCGR outperforms representative graph-based and diffusion-based recommendation methods. On average across all datasets, compared with the strongest baseline, PDCGR improves HR and NDCG by 6.69% and 6.50%, respectively. In addition, we also conduct extensive experiments to verify that PDCGR can reduce dependence on popular items, balance item exposure, and improve the ranking quality of long-tail items.