Résumé:
Fully distributed learning schemes based on opportunistic exchanges among nodes, such as Gossip Learning (GL), have recently attracted attention due to their superior scalability, robustness, and enhanced privacy protection. However, their performance has only been characterized in static or application-specific trace-driven mobility scenarios, overlooking the issue of understanding how the structure of the interactions among nodes over time affects the learning process. To address this gap, we propose a new assessment approach for GL in dynamic settings, based on two novel classes of time-varying random graphs, which extend Erdős-Rényi (ER) and Barabási-Albert (BA) random graphs to characterize generic real-world dynamic networks (e.g., social or wireless networks), while accounting for node churn and the rate at which the graph evolves. Evaluating GL on such time-varying graphs allows us to abstract the relationship between the key parameters of GL algorithms, the communication and topology patterns, and the learning performance from factors specific to the experimental contexts, generalizing our findings to a large class of real-world networks. Simulation results show that the sparser the graph, the higher the positive impact edge dynamicity has on GL mean accuracy and convergence time. Surprisingly, we observe that in networks with the same average connectivity degree, regardless of their nodes’ attachment style, a higher edge persistence reduces GL mean accuracy and convergence time, highlighting the value of a varied interaction among nodes. Finally, results show that real-world networks that exhibit preferential attachment can preserve a better GL performance in terms of mean accuracy and convergence time than random networks (i.e., ER-like), even under high node churn and edge dynamicity.
SWITCH edu-ID
Administration
