FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens

TL;DR

FACE tackles the challenge of making collaborative-filtering embeddings understandable to pre-trained LLMs by mapping them into discrete, semantically meaningful tokens. It introduces a vector-quantized disentangled mapping with a frozen LLM vocabulary as a codebook and residual quantization to produce descriptors, coupled with a contrastive alignment objective that ties these descriptors to LLM-derived text summaries. The approach is model-agnostic and does not require fine-tuning of LLMs, yet it achieves improved recommendation performance and enhances interpretability through descriptor-level explanations. Experiments on three real-world datasets show broad gains across CF backbones, and interpretability studies validate that the descriptors align with textual signals and can support item recovery and interaction explanations. This framework enables leveraging LLM capabilities for complex downstream tasks, including cold-start scenarios, without sacrificing scalability or requiring LLM fine-tuning.

Abstract

Recently, large language models (LLMs) have been explored for integration with collaborative filtering (CF)-based recommendation systems, which are crucial for personalizing user experiences. However, a key challenge is that LLMs struggle to interpret the latent, non-semantic embeddings produced by CF approaches, limiting recommendation effectiveness and further applications. To address this, we propose FACE, a general interpretable framework that maps CF embeddings into pre-trained LLM tokens. Specifically, we introduce a disentangled projection module to decompose CF embeddings into concept-specific vectors, followed by a quantized autoencoder to convert continuous embeddings into LLM tokens (descriptors). Then, we design a contrastive alignment objective to ensure that the tokens align with corresponding textual signals. Hence, the model-agnostic FACE framework achieves semantic alignment without fine-tuning LLMs and enhances recommendation performance by leveraging their pre-trained capabilities. Empirical results on three real-world recommendation datasets demonstrate performance improvements in benchmark models, with interpretability studies confirming the interpretability of the descriptors. Code is available in https://github.com/YixinRoll/FACE.

Figures (6)

• Figure 1: The overall architecture of our FACE framework. (a) Mapping stage. We employ a framework similar to the RQ-VAE architecture for the final embedding of a CF model with a frozen LLM codebook, encoding the CF embedding into pre-trained LLM tokens. (b) Alignment stage. We leverage contrastive learning to achieve the semantic alignment of descriptors and summaries.
• Figure 2: Item-retrieval task.
• Figure 3: Item-generation task.
• Figure 4: Analysis results on hyperparameter sensitivities.
• Figure 5: An example for item generation task.