Generative Recommendation for Large-Scale Advertising (arxiv.org)

A new generative recommendation system tailored for large-scale advertising, named GR4AD (Generative Recommendation for ADdvertising), has been announced, addressing the challenges associated with deploying real-time generative models. GR4AD features innovations like the Unified Advertisement Semantic ID (UA-SID) for more complex business insights and LazyAR, a lazy autoregressive decoder that minimizes layer-wise dependencies, thereby enhancing the efficiency of short, multi-candidate generation. Notably, it also incorporates Value-Aware Supervised Learning (VSL) and RSPO, a ranking-aware reinforcement learning algorithm, which optimizes for business metrics in dynamic environments. The significance of GR4AD lies in its ability to improve ad revenue by up to 4.2% when compared to existing Deep Learning Recommendation Models (DLRMs), all while maintaining high throughput and real-time service for Kuaishou's massive user base of over 400 million. This model illustrates a shift towards more adaptable and effective generative recommendation systems that not only enhance user engagement but also align closely with business objectives, making it a valuable advancement for the AI/ML community, especially in the advertising sector.