DaRec: A Novel Plug-and-Play Alignment Framework for LLMs and Collaborative Models

DaRec is an innovative framework designed to align semantic knowledge between large language models (LLMs) and collaborative models in recommender systems. This framework reduces noise by separating representations into shared and specific components. It also maintains informativeness through uniformity and orthogonal constraints and uses a dual-level structure alignment strategy for effective semantic knowledge transfer. Extensive experiments demonstrate DaRec's superior performance over existing methods across multiple datasets.