Understanding Personalization and Generalization in Decentralized Federated Learning for Human Activity Recognition

Federated Learning (FL) enables collaborative model training without sharing raw data among participants. However, FL approaches rely on a central server to aggregate local model updates, introducing a single point of failure and requiring a trusted coordination entity. Decentralized Federated Learning (DFL) addresses these limitations by replacing centralized aggregation with peer-to-peer coordination mechanisms. In this work, we investigate selected decentralized federated learning strategies for Human Activity Recognition under a common experimental setting, using centralized FedAvg as a reference baseline. Our primary focus is on how decentralization affects the trade-off between personalization and generalization and, in particular, on how this trade-off evolves throughout the iterative learning process. A comprehensive comparison of existing DFL algorithms, as well as an exhaustive exploration of the parameters governing the DFL process, is outside the scope of this work. Instead, we deliberately select representative decentralized strategies and configure them under comparable experimental conditions. This controlled setting allows us to follow the learning process in detail and gain deeper insight into how different decentralized model dissemination mechanisms shape the evolution of personalization and generalization.