FedLGN: Federated Lightweight Graph Neural Network for Recommendation

Recommender systems have become a key part of our daily digital experiences, powering personalized content discovery across various online platforms. However, the increasing privacy concerns and the distributed nature of user data pose significant challenges for traditional centralized recommendation models. Existing federated recommendation models often suffer from high computational complexity, which makes them unsuitable for client devices with limited computational and storage resources. To address these challenges, we propose a FedLGN framework in this work. FedLGN leverages the power of graph neural networks to capture complex user-item interactions, while the federated learning paradigm ensures that the model can be trained collaboratively across multiple client devices without the need to share sensitive user data. Meanwhile, our framework also enables efficient federated optimization, reducing the communication overhead and improving the overall system scalability. The experiment results demonstrate the convergence efficiency and communication cost of federated learning while maintaining recommendation accuracy, enhancing the practicality of the system. Additionally, we explore the recommendation performance of the proposed framework with different numbers of clients.