LLMs generate unbiased labels for recommendation system pre-ranking

Researchers have developed Generative Pseudo-Labeling (GPL), a new framework that uses large language models (LLMs) to create unbiased pseudo-labels for pre-ranking in recommendation systems. This method addresses the challenge of train-serving discrepancy, where models are trained on limited exposed interactions but must score all potential items, including unexposed long-tail content. By generating user-specific interest anchors and matching them in a semantic space, GPL provides high-quality training data without increasing online latency. When implemented in a large-scale production system, GPL improved click-through rates by 3.07% and enhanced recommendation diversity. AI

IMPACT Enhances recommendation system performance by addressing sample selection bias and improving long-tail content discovery.

RANK_REASON The cluster contains an academic paper detailing a new method for improving recommendation systems using LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Daixuan Cheng
• Generative Pseudo-Labeling
• LLMs

• paper
• product
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →