A0423
Title: Federated event predictions with divergence-guided global aggregation
Authors: Xuhui Fan - Macquarie University (Australia) [presenting]
Abstract: Clients-specific events, such as hospital visits, stock market events, and app-based riding-hailing, are typically modelled and predicted in a centralized manner, which may aggregate all the data for training purposes. Such events involve client privacy and are often sparse and uncertain; their existing prediction methods may raise concerns about privacy leakage and require more effort in modelling event sparsity and uncertainty. To enable client-specific privacy-preserving event prediction, the first attempt is made by proposing federated event prediction models (FedEvent), where deep sigmoidal Gaussian Cox processes are developed to generate flexible intensity functions to characterize client-specific event dynamics and capture client event uncertainties simultaneously. Further, a novel framework is proposed using divergence to guide global aggregation over all clients' modeling information, which shares and converges client information and event uncertainty. Divergence measures, including the KL divergence and the Wasserstein distance, are presented to elaborate the approach for uncertain client event prediction. Extensive experimental results verify the advantages of our approach.
