【报告摘要】Dynamic recommendation systems, where users interact with items continuously over time, have been widely deployed in real-world online applications. Most existing recommendation algorithms assume static user-item relationships and focus on traditional static recommendation paradigms, generating fixed recommendation lists during the prediction stage. Unlike static scenarios, dynamic systems must contend with rapid environmental changes where continuous user-item interactions lead to constant shifts in user interests, item popularity, and underlying interaction patterns. Traditional methods are inadequate for capturing the rapid evolution of user preferences and item characteristics due to the continuous stream of interactions. A critical challenge is determining whether new interactions genuinely reflect preference changes or temporary deviations. Meanwhile, current approaches frequently model temporal dynamics independently without considering the inherent co-evolution between users and items. In this talk, I will introduce our findings on addressing these challenges through novel GNN-based methods that capture collaborative temporal dynamics, enable efficient real-time updates, and optimize long-term user satisfaction in dynamic recommendation environments.
