Selecting collaborative learning paradigms for edge IoT: an E–M–O-based comparative simulation with design guidelines

Collaborative learning enables machine learning over edge and Internet of Things (IoT) devices, yet selecting an appropriate learning paradigm remains challenging due to heterogeneous resources, communication constraints, and diverse deployment conditions. Although centralized training, federated learning (FL), and decentralized federated learning (DFL) have been widely studied, they are often evaluated in isolation, providing limited guidance for practical system design. This paper adopts a comparative, regime-based perspective and proposes an Environment–Mechanism–Outcome (E–M– O) framework to analyze when and why different paradigms are preferable under varying edge IoT conditions. Using discrete-time simulations, we evaluate convergence speed, energy consumption, and communication behavior across representative regimes. Results show that no single paradigm dominates: centralized training is optimal in latencydominated settings, FL offers a robust trade-off under uplink constraints, and DFL achieves the lowest energy consumption when proximity communication is inexpensive and connectivity is strong, with performance highly dependent on network topology. These findings translate into practical design guidelines for paradigm selection in edge IoT systems.