Federated Data Analytics: Theory and Application

Abstract

This report develops three data analytics frameworks that solve the challenges in the engineering system, with application to quality and reliability engineering. (i) Developing FDA algorithms that target fairness: the proposed framework – GIFAIR - imposes group and individual fairness to the FDA setting. By adding a regularization term, GIFAIR penalizes the spread in the loss of client groups to drive the optimizer to a fair solution. (ii) Developing FDA algorithms that go beyond deep learning: the proposed approach extends linear models and sparse linear models to federated scenarios, presenting solutions for hypothesis testing, uncertainty quantification, variable selection, and deriving engineering insight. (iii) Tackling data heterogeneity through personalization: the proposed approach builds personalized Bayesian models that tackle data heterogeneity among different devices. Furthermore, it also provides the first theoretical results on FDA convergence in correlated settings, which is very common in engineering situations. In turn, this may help researchers further investigate FDA within alternative stochastic processes built upon correlations, such as Lévy processes.