Conditional Diffusion Denoising and Multi-Source Semantic Fusion for Social Recommendation

Abstract

Social recommendation alleviates data sparsity and cold-start issues by leveraging user social networks. However, real-world social graphs are often contaminated with noisy relations, such as weak or spurious links, which are amplified during graph propagation and degrade user representation learning. To address this issue, a core-interest-guided conditional diffusion denoising approach is proposed, where user core interests extracted from historical interactions are used as semantic conditions to progressively refine noisy social representations in the latent space. In addition, real-world recommendation is influenced not only by individual behaviors and social relations but also by global contextual factors, such as popularity trends and environmental biases. Effectively integrating these heterogeneous signals remains challenging. To this end, a multi-source semantic fusion module is developed, which employs soft-threshold multi-head attention and dynamic residual gating to adaptively align and fuse heterogeneous information. Based on these designs, a unified framework, termed CDD-MSF, is proposed. Extensive experiments on three real-world datasets demonstrate that CDD-MSF consistently outperforms state-of-the-art baselines in terms of Recall@K and NDCG@K, and exhibits strong robustness and generalization under highly sparse and noisy conditions.