Abstract
Online learning has demonstrated notable potential to dynamically allocatelimited resources to monitor a large population of processes, effectivelybalancing the exploitation of processes yielding high rewards, and theexploration of uncertain processes. However, most online learning algorithmswere designed under 1) a centralized setting that requires data sharing acrossprocesses to obtain an accurate prediction or 2) a homogeneity assumption thatestimates a single global model from the decentralized data. To facilitate theonline learning of heterogeneous processes from the decentralized data, wepropose a federated collaborative online monitoring method, which captures thelatent representative models inherent in the population through representationlearning and designs a novel federated collaborative UCB algorithm to estimatethe representative models from sequentially observed decentralized data. Theefficiency of our method is illustrated through theoretical analysis,simulation studies, and decentralized cognitive degradation monitoring inAlzheimer's disease.