Robust Federated Spatio-Temporal Data Modeling

Abstract

In this thesis, the modeling of spatio-temporal data within a federated environment stands as the primary challenge. With the increase of time series data from diverse sources like sensors, network devices, and market assets, there emerges a significant need for techniques that capture the spatio- temporal dynamics of decentralized data, especially in the presence of noise or missing values. First, we explore Matrix estimation-based (ME) time series analysis for denoising and forecasting. Given the challenges of data privacy in centralized training, we propose the Federated Multivariate Singular Spectrum Analysis (Fed-mSSA). This approach aims at high-dimensional modeling of decentralized data, using a consensus optimization for low-rank matrix representation, capturing spatio-temporal dynamics. Furthermore, a federated method is used to learn non-linear temporal dependencies. Second, considering real-world time series as complex networks, to capture the latent interconnections among different clients, we propose the Federated Multi-task Learning Predictor with Graph Laplacian Regularization (FML-GL). This dual-framework simultaneously trains the predictor while learning the graph Laplacian to reveal hidden networks. Third, given the inherent noise in time series that impacts neural networks' effectiveness, we propose the Robust Dual Recurrent Neural Networks (RDRNN). RDRNN involves two RNNs to classify samples into noisy or clear categories, aiming to extract the main data patterns and alleviate noise effects. Empirical tests in sectors like electricity and finance validate our techniques.