Bridging Behavioral Silos with LLMs for Enhanced Multi-Vertical Recommendations

In the realm of multi-vertical e-commerce, platforms like DoorDash face unique challenges, particularly with newer categories such as grocery and retail. A prominent issue is the 'cold start' dilemma, where user data is insufficient for effective personalization. This paper proposes an innovative framework aimed at elevating recommendation systems by leveraging insights from well-established verticals, such as restaurants, to enhance those that are less developed. By utilizing Large Language Models (LLMs), we can generate high-dimensional features that reflect latent user preferences. Our method employs a hierarchical Retrieval-Augmented Generation (RAG) approach to extract multi-level taxonomic features from users' restaurant orders and search behaviors. These generated features capture both enduring cross-vertical interests and immediate user intentions. We integrate these insights into a Multi-Task Learning (MTL) ranking model, demonstrating through rigorous evaluations that our strategy significantly boosts personalization and user engagement in emerging verticals, effectively addressing the gap in behavioral data.