Multi-center hierarchical federated learning for personalized user experiences in the metaverse

The Metaverse, an emerging digital frontier integrating virtual worlds and augmented reality, offers a diverse and dynamic range of experiences that blend the digital and physical realms. These interactions necessitate personalized services to ensure users have an immersive and satisfying experience. Multi-Center Federated Learning (MFL) provides personalized model training for individual users by categorizing user types and learning multiple global models. However, the vast user base in the Metaverse exerts significant communication pressure on MFL, leading to unacceptable delays. To address these challenges, we propose the Multi-Center Hierarchical Federated Learning (MHFL) framework. This framework allows edge servers to perform partial aggregation, thereby alleviating the burden on global centers. The most critical problem in implementing MHFL within edge computing environments lies in the limited resources of edge servers, which hinder their ability to deploy all central models and perform partial aggregation services effectively. We conduct a detailed analysis of this problem and identify the key ideas for its resolution. Based on this, a method is designed to optimize the deployment categories for edge servers and assign users to appropriate edge servers for aggregation, enhancing learning efficiency by reducing model transmission time. Additionally, we propose an asynchronous aggregation strategy to address the varying aggregation times caused by resource heterogeneity across different user categories, thereby enhancing training efficiency. Extensive experiments demonstrate the superior performance of our method in the Metaverse computing environment.