Reproduce “FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens” on OpenResearch

About this paper

Chao Wang, Yixin Song, Jinhui Ye, Chuan Qin, Dazhong Shen, Lingfeng Liu, Xiang Wang, Yanyong Zhang

The FACE framework directly maps continuous collaborative filtering (CF) embeddings into discrete, semantically meaningful textual tokens that large language models (LLMs) can directly interpret without fine-tuning. This approach consistently enhances recommendation performance across various CF models and enables LLMs to provide interpretable explanations for user-item interactions.

Key insights

• Directly mapping CF embeddings into pre-trained LLM textual tokens enables LLMs to inherently understand recommendations using their native literal knowledge, addressing the core semantic incompatibility.
• Disentangling CF representations into concept-specific vectors via multi-projectors and quantizing them with an LLM vocabulary-based codebook allows for a granular, interpretable conversion of continuous to discrete information.
• Semantic alignment through contrastive learning with LLM-generated textual summaries ensures that the discrete descriptors are not just tokens, but semantically meaningful and consistent with human language.

Summary by alphaXiv.