Correlated Feature-Aware Federated Learning for Preserving Interdependencies in Decentralized Edge Systems

Federated Learning (FL) enables decentralized model training, providing privacy and scalability in distributed environments. However, conventional FL frameworks typically assume feature independence and utilize uniform aggregation, treating client updates equally without accounting for feature significance. These assumptions limit effectiveness when input features are strongly correlated and device data distributions vary, reducing model generalization and predictive performance. This paper introduces Correlated Feature-Aware Weighted Federated Learning (CFAW-FL), addressing feature interdependencies by identifying and grouping correlated features across clients. CFAW-FL employs a feature-weighted aggregation strategy, prioritizing updates based on predictive relevance, thereby preserving cross-feature dynamics. Evaluated on various datasets, CFAW-FL achieves up to 20.6% reduction in log loss and 29% faster convergence compared to FedAvg, FedProx,q-FedAvg, and HeteroFL, demonstrating improved performance on heterogeneous, feature-structured data in decentralized settings.