UniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale
Researchers have developed UniPinRec, a novel system that unifies generative retrieval and ranking models for recommendation systems at Pinterest. This approach uses a single model and training stage, streamlining the process by sharing parameters and compute across both functions. The system has been deployed on Pinterest's core surfaces, resulting in a 1% increase in engagement, an 11.1% reduction in serving latency, and a 63.6% boost in query per second. AI
IMPACT Streamlines recommendation systems, potentially reducing costs and improving user engagement across platforms.