Social graph diffusion and disentangled representations learning for multimodal recommendation

Multimodal recommendation has become a promising paradigm for building advanced recommendation systems, enabling it to capture the user’s preferences accurately through an integration between the user’s collaborative signals and item modalities (e.g., text, vision, etc.). Current methods typically treat modalities as static supplements on item side and infer user preference jointly with user ID-based collaborative signals. Nevertheless, there are still two limitations: (1) The shared user ID embedding scheme forces the uncontrollable entanglement between the collaborative signal and the multimodal content on the latent space, failing to decouple historical preferences from content semantics when representing user preferences; (2) Data sparsity limits the model’s ability to capture the user’s global interest. Despite the introduction of original social relations as a complementary strategy, the inherent noise problem limits the robustness of the model. To overcome these limitations, we propose a novel Social Graph diffusion and Decoupled representation learning for Multimodal Recommendation (SGDMR). Specifically, it uses a decoupled graph propagation mechanism to separate the cooperative signal from the multimodal features, thereby eliminating the entanglement. Furthermore, to mitigate the adverse effects of noise on the model’s robustness, we introduce a social graph diffusion model to learn robust social structures. Additionally, we enhance local interactions and global semantic consistency through multi-task self-supervised learning. We conduct comprehensive comparison experiments against other state-of-the-art methods on three real-world datasets. The results demonstrate the superiority of SGDMR.