A Practical Framework for Federated Fuzzy Clusterwise Regression, Preprocessing and Hyperparameter Tuning

Federated learning has the potential to enable novel use cases in environments with data sharing constraints. While practitioners acknowledge the potential and researchers show applicability to many different use case domains, this new paradigm has not been widely adopted in practice yet. Various studies identify impact factors on adoption, including the communication overhead of the FL approach, non-identically or independently distributed (non-IID) training data, data quality problems, complexity of implementing best practices from (centralized) machine learning operations (MLOps), just to name a few. Moreover, federated regression problems remain an underexplored area as recent research activity has shown. To address the barriers to federated learning adoption, we introduce an end-to-end federated regression framework. Its core component is a novel federated fuzzy clusterwise regression method (FedFCR), complemented by federated protocols for preprocessing (to ensure data quality) and hyperparameter optimization. Through extensive numerical experiments, we find that FedFCR achieves competitive predictive performance, comparable to other recently introduced federated fuzzy regression methods, but with lower communication overhead. Moreover, our experiments show that FedFCR is robust against non-IIDness in training data. In summary, we hope that our practical end-to-end regression framework with federated protocols for preprocessing and hyperparameter optimization, low communication overhead, and robustness against non-IID data helps magnifying the adoption of federated learning in practice.