[Paper Review] Collaborative Large Language Model for Recommender Systems

Yaochen Zhu, Liang Wu+3 more|arXiv (Cornell University)|Nov 2, 2023

Topic Modeling4 citations

TL;DR

This paper proposes CLLM4Rec, the first generative recommender system that tightly integrates large language models (LLMs) with the traditional ID paradigm by extending LLM vocabularies with user/item ID tokens and using a soft+hard prompting strategy to jointly model collaborative and content semantics. The method achieves state-of-the-art performance through mutual regularization and recommendation-oriented fine-tuning, enabling efficient, hallucination-free multi-item generation without fine-tuning the LLM backbone.

ABSTRACT

Recently, there has been growing interest in developing the next-generation recommender systems (RSs) based on pretrained large language models (LLMs). However, the semantic gap between natural language and recommendation tasks is still not well addressed, leading to multiple issues such as spuriously correlated user/item descriptors, ineffective language modeling on user/item data, inefficient recommendations via auto-regression, etc. In this paper, we propose CLLM4Rec, the first generative RS that tightly integrates the LLM paradigm and ID paradigm of RSs, aiming to address the above challenges simultaneously. We first extend the vocabulary of pretrained LLMs with user/item ID tokens to faithfully model user/item collaborative and content semantics. Accordingly, a novel soft+hard prompting strategy is proposed to effectively learn user/item collaborative/content token embeddings via language modeling on RS-specific corpora, where each document is split into a prompt consisting of heterogeneous soft (user/item) tokens and hard (vocab) tokens and a main text consisting of homogeneous item tokens or vocab tokens to facilitate stable and effective language modeling. In addition, a novel mutual regularization strategy is introduced to encourage CLLM4Rec to capture recommendation-related information from noisy user/item content. Finally, we propose a novel recommendation-oriented finetuning strategy for CLLM4Rec, where an item prediction head with multinomial likelihood is added to the pretrained CLLM4Rec backbone to predict hold-out items based on soft+hard prompts established from masked user-item interaction history, where recommendations of multiple items can be generated efficiently without hallucination. Codes are released at https://github.com/yaochenzhu/llm4rec.

Motivation & Objective

• To address the semantic gap between natural language and recommendation tasks in LLM-based recommender systems.
• To overcome issues like spurious correlations from ID tokenization and poor language modeling on user/item data.
• To enable efficient, multi-item recommendation generation without hallucination using a unified LLM framework.
• To preserve pretrained LLM knowledge while adapting to recommendation-specific semantics through targeted fine-tuning.
• To demonstrate the effectiveness of combining collaborative and content signals via mutual regularization in a unified LLM-based architecture.

Proposed method

• Extends the vocabulary of pretrained LLMs with dedicated user and item ID tokens to represent users and items directly in the token space.
• Introduces a soft+hard prompting strategy where prompts consist of heterogeneous soft (user/item) tokens and hard (vocabulary) tokens, enabling stable and effective language modeling on RS-specific corpora.
• Employs a mutual regularization strategy that encourages the collaborative LLM and content LLM to share knowledge, improving robustness to noisy user/item content.
• Proposes a recommendation-oriented fine-tuning strategy with a multinomial likelihood-based item prediction head to generate multiple items efficiently from masked interaction history.
• Uses a two-stage training process: first, mutual regularization with soft+hard prompting for joint collaborative and content modeling; second, fine-tuning with a dedicated prediction head for inference efficiency.
• Projects learned user and item token embeddings into 128 dimensions to create a lightweight embedding baseline (CLLM4Rec-Emb) for industrial deployment.

Experimental results

Research questions

• RQ1 Can integrating user/item ID tokens into LLMs effectively bridge the semantic gap between natural language and recommendation tasks?
• RQ2 How can soft+hard prompting improve language modeling stability and effectiveness on hybrid user-item interaction and content data?
• RQ3 Does mutual regularization between collaborative and content LLMs enhance robustness to noisy user/item features?
• RQ4 Can recommendation-oriented fine-tuning enable efficient, multi-item generation without hallucination, while preserving pretrained LLM knowledge?
• RQ5 How does CLLM4Rec compare to strong baselines in both offline evaluation and real-world industrial deployment scenarios?

Key findings

• CLLM4Rec outperforms strong baselines, including the two-tower model used at LinkedIn, by a large margin on the LinkedIn dataset, demonstrating its effectiveness in real-world settings.
• CLLM4Rec achieves state-of-the-art performance on the AM-Beauty, AM-Toys, AM-Sports, and Yelp datasets, with consistent improvements across all metrics.
• The ablation study shows that full fine-tuning of the LLM backbone degrades performance, indicating that preserving pretrained knowledge is critical.
• The optimal mutual regularization strength is around λc = 1, where performance peaks across all datasets, confirming the importance of balanced regularization.
• CLLM4Rec-Emb, a lightweight version using only learned user and item token embeddings, outperforms both the original two-tower model and M6-Retrieval, showing strong potential for low-latency industrial deployment.
• The recommendation-oriented fine-tuning strategy enables efficient, multi-item generation without hallucination, significantly improving inference efficiency over autoregressive baselines.