HDRec: Hierarchical Distillation for Enhanced LLM-based Recommendation Systems

Abstract

Large Language Models (LLMs) have shown significant potential in recommendation systems by enhancing the semantic reasoning capabilities derived from user-item interactions. However, existing methods often rely on original reviews as ground truth explanations, with limited attention to uncovering the underlying rationales behind each interaction, which hampers the reasoning performance of LLMs. In this paper, we propose a novel Hierarchical Distillation for Recommendation (HDRec) model that effectively specifies user and item profiles by hierarchically distilling interaction rationales from reviews using LLMs. Additionally, we introduce a review summary task that condenses distilled information, such as user preferences, personality traits, item attributes, and target audience, improving both model training and interpretability. Extensive experiments demonstrate that HDRec achieves state-of-the-art performance on three real-world datasets in both sequential and Top-N recommendation tasks. The source code for HDRec is publicly available at https://github.com/linglingl635/HDRec.