Graph Diffusion Social Recommendation

Download PDF
Open Webpage

Jinglong Gao, Yi Zhang, Yiwen Zhang

Published: 2025, Last Modified: 16 Jan 2026IEEE Trans. Consumer Electron. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Social recommendation integrates social networks into the recommendation task, leveraging social relationships to improve the recommendation performance and alleviate the data sparsity issue. Recently, graph-based social recommendation models have shown good results by capturing high-order social influence. Most graph-based social recommendation models directly incorporate social networks into the paradigm and influence user preferences through the homogeneity of social networks, providing more possibilities for users to choose items. Although these methods are effective, the time overhead significantly increases due to the complex structures of social networks and the intricate computations required by Graph Neural Networks. More importantly, existing studies often ignore the fact that there is a large amount of recommendation-irrelevant noise in user interaction records and social networks in the real world. As the graph structure propagates, these noise will seriously interfere with the recommendation process. To address these challenges, we propose an denoising social recommendation framework called Graph Diffusion Social Recommendation (GDSR). From a technical perspective, GDSR mainly consists of two parts: a denoising recommendation framework based on social networks and a recommendation denoising framework based on preference fusion. Specifically, GDSR introduces Gaussian noise to the interaction graph and social networks structures during training to disrupting the original structures. In the prediction phase, a parameterized neural network iteratively reconstructs the graph structure, which is then utilized to rank and predict non-interacted items. Eight baseline methods were assessed on three real-world datasets. The GDSR model outperformed these methods, improving Recall@20 by 11.44%, 9.55%, and 6.33% on the Douban-Book, Douban-Movie, and Ciao datasets, respectively. Comprehensive experiments confirm GDSR’s superior recommendation performance over existing approaches.