Time-Aware Diffusion based on Preference Disentanglement for Generative Recommendation

Researchers introduce TDPM, a novel generative recommendation framework that applies time-aware diffusion models to improve personalized item suggestions by distinguishing between long-term period preferences and short-term event-triggered preferences. The approach achieves significant performance improvements of up to 29.21% in Hit Rate and 25.45% in NDCG metrics compared to existing methods.

Generative recommendation systems represent a shift from traditional item-matching approaches toward semantic understanding of user preferences through diffusion models. TDPM addresses a critical oversight in existing diffusion-based recommenders: the uniform treatment of all historical interactions regardless of temporal context. This matters because user preferences are inherently dynamic, influenced by both stable long-term patterns and recent catalytic events that dramatically shift immediate behavior.

The distinction between period preference (stable, long-horizon patterns) and point preference (event-driven, short-term shifts) reflects real user behavior dynamics. A customer's general interest in electronics remains constant, yet a specific product launch or personal need creates temporary preference spikes. By explicitly modeling this temporal heterogeneity in the diffusion process rather than treating all interactions equally, TDPM captures nuanced preference evolution that static models miss.

For recommendation platforms and e-commerce services, improved recommendation accuracy translates directly to increased conversion rates and user satisfaction. The 25-29% performance gains suggest meaningful real-world impact, as even marginal improvements in recommendation quality compound across millions of user interactions. This advancement particularly benefits large-scale platforms managing complex temporal preference dynamics.

The research demonstrates that architectural innovations in generative models continue yielding substantial performance gains. As diffusion models mature beyond image generation into sequential decision-making tasks, time-awareness becomes increasingly essential. Future developments likely involve extending temporal modeling to multimodal preference signals and cross-domain recommendation scenarios where temporal patterns vary significantly by content type.

• →TDPM framework achieves 29.21% improvement in Hit Rate by modeling time-aware diffusion in recommendation systems
• →User preferences are disentangled into stable period preferences and event-triggered point preferences for better temporal modeling
• →Diffusion-based generative recommenders now account for preference non-stationarity, addressing fundamental limitations of uniform diffusion approaches
• →Performance gains suggest significant commercial value for e-commerce and content platforms implementing temporal-aware recommendations
• →Time-aware token diffusion emerges as critical architectural component for next-generation recommendation systems