Discovering Communities With Clustered Federated Learning
Découvrir des communautés avec l'apprentissage fédéré en clusters
Résumé
We address a community detection problem in a realistic federated learning setup where clients own non-iid data. We propose a Clustered Federated Learning-based method (CFL) that can dynamically discover client communities according to their model distances along the federated rounds. This method is based on Louvain clustering, a relevant model similarity measure and a client aggregated model attribution strategy. The proposed framework enables the unsupervised detection of communities with no prior knowledge while maximizing client task performances. We propose an extensive study based on the Cifar10 dataset to assess the sensitivity of the approach to critical factors including data non-iidness level, model initialization, client participation rates and client cluster attribution strategy. Importantly, both model task accuracy and clustering relevance are evaluated thus extending state-of-the-art standard evaluation. Compared to stateof-the-art on an image classification problem, we show the interest of our continuous clustering and attribution strategy along federated rounds that maintain client migration capability while preserving the aggregated model relevance. This facilitates learning convergence while reducing result variability along trials. This work can be flawlessly integrated in standard FL approaches and opens new directions for both task performance and community detection relevance in a federated learning context. Results show the relevance of the clustering on an image classification task to discover communities of related classes.
Domaines
Intelligence artificielle [cs.AI]Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
licence |