Uncertainty-guided Multi-modal Sequential Recommendation

With the rapid development of multimedia systems, multimodal recommendation has emerged as a crucial paradigm by leveraging rich multimodal information. However, multimodal data inherently contains aleatoric uncertainty in their respective feature spaces. This uncertainty becomes particularly critical in sequential recommendation (SR): multimodal interaction data across different time steps may become unreliable due to inherent noise, which propagates and accumulates through temporal dependencies, severely impacting recommendation performance. To address this challenge, we propose Uncertainty-guided Multi-modal Sequential Recommendation (UMSR), which consists of two core modules: (1) Multimodal Optimization via Hierarchical Bridging (MOHB), quantifying intra-modal uncertainty through multivariate Gaussian distributions and learning compact joint representations; (2) Lipschitz-Continuous Attention (LCA) effectively models the dynamic uncertainty inherent in multimodal data across different time steps, while also enhancing collaborative transitivity. Extensive experiments on four real-world datasets demonstrate our model’s superiority over existing methods.