FastPFRec: A Fast Personalized Federated Recommendation with Secure Sharing
New framework cuts training rounds by 32% while improving accuracy by 8.1% on real-world datasets.
A research team led by Zhenxing Yan has introduced FastPFRec (Fast Personalized Federated Recommendation with Secure Sharing), a new framework designed to solve two major problems in graph neural network (GNN)-based federated recommendation systems: slow convergence on graph data and privacy leakage risks during model collaboration. The system accelerates training through an efficient local update strategy and enhances security with a novel privacy-aware parameter sharing mechanism.
Experiments conducted on four real-world datasets—Yelp, Kindle Store, Gowalla-100k, and Gowalla-1m—demonstrate significant performance gains. FastPFRec required 32.0% fewer training rounds to converge and reduced total training time by 34.1% compared to existing baseline methods. Crucially, it also achieved an 8.1% improvement in recommendation accuracy, proving that efficiency gains do not come at the cost of model quality.
The framework's core innovation lies in its balanced approach to the federated learning dilemma. By optimizing how local client models perform updates and share only essential, sanitized parameters, it mitigates the risk of data reconstruction attacks that can plague collaborative systems. This makes FastPFRec a practical advancement for companies needing to build scalable, personalized recommendation engines—like those for e-commerce or social platforms—without centralizing sensitive user data.
- Achieves 34.1% shorter training time and 32.0% fewer training rounds than existing federated recommendation baselines.
- Improves recommendation accuracy by 8.1% on real-world datasets including Yelp, Kindle, and Gowalla.
- Introduces a privacy-aware parameter sharing mechanism to mitigate data leakage risks during collaborative model training.
Why It Matters
Enables faster, more accurate, and privacy-compliant personalized recommendations for e-commerce and social platforms without centralizing user data.