Reproduce “LARR: Large Language Model Aided Real-time Scene Recommendation with Semantic Understanding” on OpenResearch

About this paper

Zhizhong Wan, Bin Yin, Junjie Xie, Fei Jiang, Xiang Li, Wei Lin

Meituan's LARR framework integrates Large Language Models (LLMs) to provide deep semantic understanding of real-time scenes in food delivery recommendation systems, addressing traditional models' limitations. It achieves this efficiently by transforming an LLM into an offline text embedding model, resulting in a 2.5% increase in CTR and 1.2% growth in GMV in online A/B tests.

Key insights

• Continual pretraining with domain-specific corpora and special tokens significantly enhances an LLM's understanding of recommendation context and POI characteristics.
• Transforming the LLM into an offline text embedding model and aggregating its outputs with a lightweight encoder enables efficient, real-time semantic understanding for industrial-scale deployment.
• Multimodal alignment through contrastive learning effectively bridges semantic features from the LLM with collaborative features from traditional RS, leading to more robust and accurate predictions.

Summary by alphaXiv.