[Paper Review] TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation

Keqin Bao, Jizhi Zhang+4 more|arXiv (Cornell University)|Apr 30, 2023

Topic Modeling66 references13 citations

TL;DR

TALLRec tunes LLaMA-7B with lightweight Alpaca and rec-tuning stages (using LoRA) to align LLMs with recommendation tasks, achieving strong few-shot performance and cross-domain generalization, outperforming traditional and In-context Learning baselines.

ABSTRACT

Large Language Models (LLMs) have demonstrated remarkable performance across diverse domains, thereby prompting researchers to explore their potential for use in recommendation systems. Initial attempts have leveraged the exceptional capabilities of LLMs, such as rich knowledge and strong generalization through In-context Learning, which involves phrasing the recommendation task as prompts. Nevertheless, the performance of LLMs in recommendation tasks remains suboptimal due to a substantial disparity between the training tasks for LLMs and recommendation tasks, as well as inadequate recommendation data during pre-training. To bridge the gap, we consider building a Large Recommendation Language Model by tunning LLMs with recommendation data. To this end, we propose an efficient and effective Tuning framework for Aligning LLMs with Recommendation, namely TALLRec. We have demonstrated that the proposed TALLRec framework can significantly enhance the recommendation capabilities of LLMs in the movie and book domains, even with a limited dataset of fewer than 100 samples. Additionally, the proposed framework is highly efficient and can be executed on a single RTX 3090 with LLaMA-7B. Furthermore, the fine-tuned LLM exhibits robust cross-domain generalization. Our code and data are available at https://github.com/SAI990323/TALLRec.

Motivation & Objective

• Bridge the gap between LLM training tasks and recommendation tasks by building a Large Recommendation Language Model (LRLM).
• Develop an efficient two-stage tuning framework (alpaca tuning and rec-tuning) to adapt LLMs for recommendation with limited data and GPU resources.
• Demonstrate that tuned LRLMs outperform traditional recommenders and In-context Learning baselines in few-shot settings and show cross-domain generalization.

Proposed method

• Format recommendation data as instruction-tuning inputs and outputs for rec-tuning.
• Use Alpaca tuning with self-instruct data to improve generalization of the LLM.
• Apply rec-tuning to tailor the model for predicting user preference on a target item using historical interactions.
• Employ LoRA to enable lightweight parameter-efficient fine-tuning on LLaMA-7B and freeze base parameters.
• Train with two-stage framework (alpaca tuning followed by rec-tuning) implemented on a single RTX 3090.
• Backbone model selection focuses on open-source LLaMA rather than API-based models to reflect practical deployment and data-security considerations.

Experimental results

Research questions

• RQ1 RQ1: How does TALLRec perform compared with current LLM-based and traditional recommendation models?
• RQ2 RQ2: How do the different components in TALLRec affect its effectiveness?
• RQ3 RQ3: How does TALLRec perform under cross-domain recommendation?

Key findings

• TALLRec significantly outperforms both traditional and LLM-based baselines in few-shot settings.
• In-context Learning with GPT-3.5-based prompts performs poorly compared to tuned LLMs under rec-tuning.
• Alpaca tuning contributes to generalization, but rec-tuning is essential for achieving strong recommendation performance in few-shot scenarios.
• TALLRec exhibits strong cross-domain generalization (e.g., movie to book) and can improve further when combining data from multiple domains.
• GRU-BERT and DROS-BERT variants do not significantly improve over their base models in the few-shot setting, indicating textual descriptions alone may not boost traditional models.
• The results include clear improvements in AUC across datasets, with the final TALLRec variant achieving the highest scores in the reported tables.