Efficient Clustering with Quality Guardrails for LLM-based Recommender Systems at Industry Scale

The researchers proposed a scalable two-stage clustering algorithm aimed at providing proven sample-level quality assurance for industrial-level recommendation systems driven by large language models (LLMs). The algorithm first uses Mini-batch K-Means to generate initial clusters, and then employs a greedy strategy to select representative points, ensuring that each sample shares the specified minimum embedding similarity and exact attribute matching with the representative points. Experimental verification showed that this method is significantly superior to conventional clustering approaches in terms of running speed and scalability. Currently, the algorithm has been successfully applied in real-world deployments involving 38 million customer data, reducing the cost and running time of downstream large language models by 50 times while maintaining personalized recommendation effectiveness. The persona recommendation system built using this technology has been officially launched, and A/B testing has achieved significant revenue growth and user engagement improvement.