Synergistic data-resource participant selection for efficient Federated Edge Learning in IoT ecosystems

The Internet of Things (IoT), as a concept, is becoming increasingly integral to our daily lives, enabling smart environments through sensing, communication, and computation. However, real-world edge devices exhibit pronounced heterogeneity and inherent limitations in both computational resources and data distributions, posing significant challenges for deploying robust, efficient, and adaptive edge intelligence. We propose FedCDRP, a novel Federated Edge Learning (FEEL) methodology for addressing the realistic dynamic heterogeneity of IoT environments, through effective participant selection. The key novelty of FedCDRP is the dual-centric selection of the most contributing and resource-efficient clients to take part in the FEEL process. Specifically, FedCDRP jointly tackles both imbalanced heterogeneous data and system performance fluctuations through our class diversity and resource performance-aware client selection strategy, by considering local data richness and resource occupancy forecasting of IoT edge devices. Through real-world circumstances, the conducted experiments demonstrated the superior computational efficiency of our FedCDRP strategy, outperforming the state-of-the-art techniques with an improvement rate as high as 63.14% in the learning capabilities of the global model. This is further evidenced by significant ecosystem-level gains, including the completion of 40 additional communication rounds within identical time intervals and achieving high model performance up to 42 rounds earlier than competing approaches. These results are further supported by statistical significance across multiple runs, robustness via sensitivity analysis, validation on a real-world FEEL platform, and stronger out-of-distribution generalization.