Overview
Traditional recommender systems rely heavily on graph-based ID tracking, which often ignores deep textual context. Recent frameworks integrate LLMs to understand the rich semantics of user behaviors and item descriptions.
Frameworks (by HKUDS)
LLMRec
LLMRec focuses on Large Language Models with Graph Augmentation for Recommendation. It uses LLM-enhanced data augmentors to improve recommendation accuracy over sparse benchmark graphs.
RLMRec
RLMRec is a model-agnostic framework that enhances existing recommenders with LLM-empowered representation learning, capturing intricate semantic aspects of user-item relationships.
RecLM
Focuses on enhancing the capability of capturing user preference diversity using Reinforcement Learning reward functions to facilitate self-augmentation of the language models.
Recursive Language Models (RLMs)
A distinct concept involving an inference strategy where an LLM can decompose and recursively interact with an input context of unbounded length through REPL environments (sandboxes).
TODO: Add architectural overviews on how LLM representations are merged with Graph Neural Networks.