Federated Personal Knowledge Graph Completion with Lightweight Large Language Models for Personalized Recommendations
New AI framework combines federated learning with personal knowledge graphs, achieving 4x F1-score improvement over benchmarks.
Researchers Fernando Spadea and Oshani Seneviratne have introduced FedTREK-LM, a novel framework designed to tackle the privacy-personalization paradox in AI recommendations. The system cleverly combines several advanced techniques: lightweight large language models (specifically the Qwen3 family with 0.6B, 1.7B, and 4B parameters), evolving personal knowledge graphs (PKGs) that store individual user preferences, and federated learning (FL) to train models without centralizing sensitive data. A key innovation is the use of Kahneman-Tversky Optimization, which incorporates principles from behavioral economics to better model human decision-making biases. By prompting the smaller LLMs with structured data from a user's personal knowledge graph, FedTREK-LM performs context-aware reasoning to generate tailored suggestions.
In rigorous testing, FedTREK-LM demonstrated a massive leap in performance, consistently and substantially outperforming established state-of-the-art knowledge graph completion and federated recommendation systems like HAKE, KBGAT, and FedKGRec. It achieved more than a 4x improvement in the critical F1-score metric on benchmarks for movie and food recommendations. The research also delivered a crucial finding: synthetic data is a poor substitute for real user information, degrading model performance by up to 46%, underscoring the need for frameworks that can leverage genuine private data ethically.
This work establishes a practical new paradigm for adaptive, LLM-powered services that respect user privacy. Instead of sending personal data to a central server, the intelligence—in the form of a lightweight LLM and a personal knowledge graph—resides and evolves on the user's own device. The framework is designed to generalize across decentralized and constantly changing user profiles, offering a scalable path forward for personalized AI in a privacy-conscious world.
- Achieved >4x F1-score improvement over benchmarks (HAKE, KBGAT, FedKGRec) using lightweight Qwen3 LLMs (0.6B-4B parameters).
- Leverages federated learning & personal knowledge graphs (PKGs) to enable private, on-device learning without data centralization.
- Found synthetic training data degrades performance by up to 46%, proving the necessity of frameworks that use real user data privately.
Why It Matters
Enables highly accurate, personalized AI services like streaming and shopping recommendations without compromising user data privacy.