CLoVE: Personalized Federated Learning through Clustering of Loss Vector Embeddings

CLoVE (Clustering of Loss Vector Embeddings) is a new algorithm for Clustered Federated Learning (CFL) that identifies client clusters based on their loss vector embeddings, allowing for the optimization of cluster-specific models. It simplifies the clustering process without requiring optimal model initialization, making it robust for real-world applications. Theoretical convergence bounds indicate that CLoVE can accurately recover clusters and achieve optimal model convergence quickly, outperforming existing CFL and Personalized Federated Learning (PFL) methods in various non-IID settings.