Selective Participation in Federated Learning under Temporally Correlated Client Data

Abstract

Federated learning typically assumes that repeatedly selecting available clients improves training efficiency by increasing the amount of data incorporated into optimization. This assumption becomes problematic when client observations are temporally correlated. Devices participating frequently within a short period may contribute highly redundant samples, causing the global model to overrepresent recent local states while underutilizing less frequent but more diverse observations. This paper investigates selective client participation from the perspective of temporal information redundancy. We introduce a participation criterion that estimates the marginal contribution of a client update relative to its recent training history and the current global model state. Clients whose updates provide little additional information can temporarily defer participation without permanently excluding their local data. The resulting procedure separates client availability from client usefulness and can be integrated with standard federated optimization without requiring access to raw local observations. We analyze the behavior of the method under different degrees of temporal correlation and non-IID data distributions, and examine its effects on convergence, communication demand, and representation across clients. The results suggest that maximizing participation frequency is not always an efficient strategy when local data evolve slowly over time.