Enhancing LLM-based Recommendation through Semantic-Aligned Collaborative Knowledge

TL;DR

The paper tackles the problem of semantic misalignment between LLM knowledge and collaborative signals in recommender systems, proposing SeLLa-Rec to fuse these sources through semantic-aligned tokens.It introduces a three-layer architecture and a hierarchical training procedure that includes LoRA-based fine-tuning of the LLM, bidirectional contrastive alignment with collaborative embeddings, and a Hybrid Projection Layer that injects three tokens into task prompts.Key contributions include a pre-aligned collaborative foundation, a projection-based token integration mechanism, and extensive experiments showing state-of-the-art performance on MovieLens-1M and Amazon Book, with notable gains in both warm and cold data regimes.The work demonstrates that aligning semantic spaces between LLMs and Collab models facilitates more effective knowledge integration, offering practical benefits for large-scale, knowledge-rich recommendation systems.

Abstract

Large Language Models (LLMs) demonstrate remarkable capabilities in leveraging comprehensive world knowledge and sophisticated reasoning mechanisms for recommendation tasks. However, a notable limitation lies in their inability to effectively model sparse identifiers (e.g., user and item IDs), unlike conventional collaborative filtering models (Collabs.), thus hindering LLM to learn distinctive user-item representations and creating a performance bottleneck. Prior studies indicate that integrating collaborative knowledge from Collabs. into LLMs can mitigate the above limitations and enhance their recommendation performance. Nevertheless, the significant discrepancy in knowledge distribution and semantic space between LLMs and Collab. presents substantial challenges for effective knowledge transfer. To tackle these challenges, we propose a novel framework, SeLLa-Rec, which focuses on achieving alignment between the semantic spaces of Collabs. and LLMs. This alignment fosters effective knowledge fusion, mitigating the influence of discriminative noise and facilitating the deep integration of knowledge from diverse models. Specifically, three special tokens with collaborative knowledge are embedded into the LLM's semantic space through a hybrid projection layer and integrated into task-specific prompts to guide the recommendation process. Experiments conducted on two public benchmark datasets (MovieLens-1M and Amazon Book) demonstrate that SeLLa-Rec achieves state-of-the-art performance.

Figures (8)

• Figure 1: The overall framework of SeLLa-Rec. The right part depicts SeLLa-Rec's structure, while the left illustrates the origins of the pre-trained components. The numbered markers ①, ②, and ③ denote the sequential training stages for SeLLa-Rec.
• Figure 2: The prompt structure for Lora finetuning.
• Figure 3: The prompt structure for LLM distillation.
• Figure 4: The prompt structure for Projection Training.
• Figure 5: Models' performance on warm/cold data.