Researchers use LLMs to break recommendation silos

Researchers Nimesh Sinha, Raghav Saboo, Martin Wang, and Sudeep Das published a paper titled 'Mind the Gap: Bridging Behavioral Silos with LLMs in Multi-Vertical Recommendations' on arXiv, proposing a framework to tackle the cold-start problem in emerging e-commerce verticals.

The team leveraged Large Language Models (LLMs) with a hierarchical Retrieval-Augmented Generation (RAG) pipeline to synthesize sparse, high-dimensional user features from data-rich verticals (e.g., restaurants) for data-sparse verticals (e.g., groceries). These generated features, encoding both long-term and short-term user preferences, were integrated into a Multi-Task Learning (MTL) ranking model. Extensive offline and online evaluations showed significant improvements in personalization and engagement, effectively bridging behavioral data gaps in platforms like DoorDash.