This paper proposes a novel generative model that overcomes the limitations of the assumption of social homogeneity (the assumption that individuals with social connections share similar preferences) to enhance the effectiveness of social recommendations on online platforms. To address the problem that the assumption of social homogeneity does not always hold due to the complexity and noise of real-world social networks, we propose a score-based generative model, the Score-based Generative Model for Social Recommendation (SGSR). SGSR applies a stochastic differential equation (SDE)-based diffusion model to social recommendations, utilizes a co-curricular learning strategy to mitigate the missing supervised signal problem, and utilizes self-supervised learning techniques to align knowledge across social and collaborative domains. Experimental results using real-world datasets demonstrate that SGSR effectively filters out unnecessary social information and improves recommendation performance.
