A new self-supervised hard negative sampling technique has been developed for large-scale two-tower retrieval models, commonly used in recommendation systems. This method utilizes a large language model (LLM) to cluster and generate challenging negative samples in real-time during training. The approach aims to improve model performance by providing more informative negatives than traditional in-batch or out-of-batch methods. Experiments and deployment in a large-scale online system indicate that this technique surpasses current industry standards, helps mitigate feedback loops, and reduces popularity bias. AI
IMPACT Enhances recommendation system performance by improving training data quality and reducing bias.
RANK_REASON Academic paper detailing a new technique for AI model training. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
