Reproduce “Cluster-driven Graph Federated Learning over Multiple Domains” on OpenResearch

About this paper

Debora Caldarola, Massimiliano Mancini, Fabio Galasso, Marco Ciccone, Emanuele Rodolà, Barbara Caputo

Cluster-driven Graph Federated Learning (FedCG) addresses statistical heterogeneity in federated learning by identifying data distributions across clients and sharing knowledge between these domains. The framework utilizes federated clustering, domain-specific model components, and a graph convolutional network, leading to improved performance on non-i.i.d. data and better adaptation to new domains at inference time compared to prior methods.

Key insights

• Identifying underlying data distributions (domains) across clients is crucial for addressing statistical heterogeneity in FL.
• A graph convolutional network can effectively model relationships and share knowledge between these identified domains, improving generalization.
• Combining shared model components with domain-specific residual modules, guided by soft domain assignments, allows for adaptive inference on both seen and unseen data distributions.

Summary by alphaXiv.