Title: Time-Aware Diffusion via Preference Disentanglement in Generative Recommendation

Abstract: Generative Recommenders (GRs) have recently revolutionized the recommendation landscape by substituting conventional item identifiers with semantic indices (SIDs). Leveraging the robust generative power of diffusion models, several initial studies have investigated diffusion-based architectures for GRs. Nevertheless, a critical shortcoming of current diffusion-driven GRs lies in their uniform application of the diffusion process across all items in a user’s historical interactions. This approach fails to account for the fact that user preferences are influenced by complex, time-varying factors, resulting in a non-stationary temporal distribution. To address this limitation, we introduce TDPM, a novel GR framework that implements time-aware diffusion on SID tokens. TDPM explicitly incorporates the dynamics of evolving user preferences into the diffusion mechanism. Specifically, it decomposes user preference into two distinct components: (i) period preference, which demonstrates stability over extended periods, and (ii) point preference, which arises from recent salient events. Comprehensive evaluations on three real-world public datasets confirm TDPM’s substantial advantage over state-of-the-art baseline methods. TDPM yields average performance gains of up to 29.21% in HR@20 and 25.45% in NDCG@20. Furthermore, ablation studies highlight the essential role of time-aware token diffusion in enhancing diffusion-based GRs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC