DiffSBR: A diffusion model for session-based recommendation

Session-based recommendation (SBR) focuses on recommending items to anonymous users within short interaction sequences. Existing solutions focus on modeling item representations as fixed embedding vectors within the discriminative learning paradigm, which fail to accurately capture the diverse preferences that user exhibit during dynamic decision-making. We argue that users in the anonymous environment can fundamentally be regarded as a normative implicit group, exhibiting both homogeneous preference and heterogeneous preference when selecting items. To tackle this, we propose a Diffusion Model for Session-based Recommendation (DiffSBR). Specifically, we first model the aforementioned user diverse preferences from both local and global views. Next, we introduce a cluster-aware diffusion model, which directly represents heterogeneous preference clusters as distribution through forward and reverse processes, while indirectly influencing homogeneous preference via the attention mechanism in the final prediction stage, thereby improving the learning of item and session representations and enhancing the next-item recommendation. Experimental results show that DiffSBR outperforms the strong baseline, demonstrating that this sampling-allocation approach accurately reflects the uncertainty and variability in user preferences.