Diffusion social augmentation for social recommendation

Social recommendation utilizes social information to learn users’ social preferences, which improves the accuracy of user preference modeling. However, social network often contains a significant amount of noisy social relationships, which can degrade the quality of user preference modeling. Existing approaches cannot effectively solve the noise problem existing in social network, resulting in model performance being affected by noisy social relationships. To address this problem, we propose a diffusion social augmentation for social recommendation (DiffuSAR). Our approach introduces a diffusion model to generate denoised semantics from social collaboration information, effectively reducing the noise effect in the social latent space. By realizing the knowledge transfer from denoised social semantics to the latent space of users in interactions through a meta-network, our model can accomplish diffusion social augmentation based on social graph data. Additionally, we incorporate a social-aware information augmentation strategy into the self-supervised learning framework, enhancing the task’s ability to discern noise amidst varying social dependencies. Experiments on three real datasets demonstrate that the model is improved over other baseline models.