Reproduce “DRE: Generating Recommendation Explanations by Aligning Large Language Models at Data-level” on OpenResearch

About this paper

Shen Gao, Yifan Wang, Jiabao Fang, Lisi Chen, Peng Han, Shuo Shang

The DRE framework generates recommendation explanations by aligning Large Language Models (LLMs) with black-box recommender systems at the data level. It leverages target-aware user preference distillation from item reviews to provide semantically rich and accurate explanations, outperforming existing LLM-based explainers and general-purpose LLMs like ChatGPT and Mistral across multiple benchmark datasets in Aspect Score and Rating Score.

Key insights

• Aligning LLMs with recommender systems at the data-level allows for non-intrusive explanation generation, preserving the performance of the core recommender while providing high-quality explanations.
• Leveraging user review data, specifically through target-aware distillation, is crucial for overcoming semantic information scarcity and generating more nuanced and persuasive explanations.
• LLMs can effectively act as powerful reasoning and distillation engines, processing diverse textual inputs to infer relationships and synthesize information for explainable AI.

Summary by alphaXiv.