Enhancing LLM-based Recommendation with Preference Hint Discovery from Knowledge Graph

TL;DR

This work tackles the semantic gap in LLM-based recommendations by introducing PIDLR, which discovers instance- and user-level preference hints from an interaction-integrated knowledge graph to guide LLM reasoning. It combines a collaborative hint extraction module with an instance-wise dual-attention mechanism to select concise, informative attributes, then translates these hints into flattened, head-centric prompts for an LLM, optionally fine-tuned via LoRA. The approach yields consistent improvements on pairwise and listwise tasks over traditional and other LLM-based baselines on MovieLens and LastFM, and ablation studies confirm the essential roles of collaborative hints and instance-specific attribute selection. The method demonstrates strong few-shot performance and practical efficiency, highlighting a generalizable strategy for leveraging traditional recommendation insights to enhance large-language-model-based systems.

Abstract

LLMs have garnered substantial attention in recommendation systems. Yet they fall short of traditional recommenders when capturing complex preference patterns. Recent works have tried integrating traditional recommendation embeddings into LLMs to resolve this issue, yet a core gap persists between their continuous embedding and discrete semantic spaces. Intuitively, textual attributes derived from interactions can serve as critical preference rationales for LLMs' recommendation logic. However, directly inputting such attribute knowledge presents two core challenges: (1) Deficiency of sparse interactions in reflecting preference hints for unseen items; (2) Substantial noise introduction from treating all attributes as hints. To this end, we propose a preference hint discovery model based on the interaction-integrated knowledge graph, enhancing LLM-based recommendation. It utilizes traditional recommendation principles to selectively extract crucial attributes as hints. Specifically, we design a collaborative preference hint extraction schema, which utilizes semantic knowledge from similar users' explicit interactions as hints for unseen items. Furthermore, we develop an instance-wise dual-attention mechanism to quantify the preference credibility of candidate attributes, identifying hints specific to each unseen item. Using these item- and user-based hints, we adopt a flattened hint organization method to shorten input length and feed the textual hint information to the LLM for commonsense reasoning. Extensive experiments on both pair-wise and list-wise recommendation tasks verify the effectiveness of our proposed framework, indicating an average relative improvement of over 3.02% against baselines.

Figures (7)

• Figure 1: An illustration of preference hints discovered from interaction-integrated Knowledge Graph.
• Figure 2: Overall Framework of PIDLR. Initially, the Collaborative Preference Hint Extraction module supplements users' potential preference hints preference hints for unseen items based on collaborative filtering principle. Then, the Instance-wise Hint Discovery module identifies the personalized preference hints for each instance. Finally, Head-Centric Translation converts the identified preference hints into flattened texts for LLM-based recommendation.
• Figure 3: Prompt example on LastFM.
• Figure 4: Ablation study on collaborative preference hint extraction and instance-wise preference discovery modules.
• Figure 5: Case Studies. (a) PIDLR infers preference hints for adventure/survival films from the user’s historical interactions in the left-side knowledge graph, achieving successful recommendation of Swiss Family Robinson. (b) Collaborative preference extraction reveals potential interest in Miramax/insect films, enabling accurate recommendation of Microcosmos.