LLMs Extract User Preferences for Privacy-Preserving Recommendation Systems

Personal Knowledge Graphs (PKGs) offer a promising method for modeling user preferences while maintaining privacy, especially when dealing with decentralized and unstructured conversational data. A new paper introduces a reproducible pipeline designed to bridge the gap between raw conversational text and structured semantic information. This pipeline leverages lightweight Large Language Models (LLMs) to extract RDF-compliant triples, which are then linked to Wikidata identifiers, from conversational data. This process facilitates the construction of PKGs that can power recommendation systems. The study specifically evaluated Qwen- and Gemma-based models on their ability to accurately extract these triples and assessed the practical utility of the resulting knowledge graphs in a recommendation task. The findings indicate that certain models performed effectively, demonstrating a strong correlation between their triple extraction accuracy and their subsequent performance in generating recommendations.