Utility-aware Collaboration Structure Optimization in Federated Learning

Federated learning enables collaborative modeling across data silos while preserving data privacy. However, distribution heterogeneity and data imbalance across clients often degrade the performance of global models, even causing negative transfer. Existing works attempt to form collaborative clusters to mitigate this issue, but most rely on explicit distribution estimation and complex parameter optimization, resulting in high computation and poor stability. This paper proposes a utility-aware collaboration structure optimization framework for federated learning. By identifying optimal collaborator sets for each client based on individual utility, we construct a benefit graph and detect stable coalitions that maximize model performance. We theoretically analyze the error bound and empirically validate our method on real-world medical datasets.