Graphical Models for Complex Multivariate Data

ongoing

Modern science increasingly relies on insights gained from sophisticated analyses of large data sets. An ambitious goal of such data-driven discovery is to understand complex systems via statistical analysis of multivariate data on the activity of their interacting units. Probabilistic graphical models, the topic of this project, are tailored to the task. The models facilitate refined yet tractable data exploration by using graphs to represent complex stochastic dependencies between considered variables. Models based on directed graphs, in particular, provide the state-of-the-art approach for detailed exploration of cause-effect relationships. However, modern applications of graphical models face numerous challenges such as key variables being latent (i.e., unobservable/unobserved), lacking temporal resolution in studies of feedback loops, and limited experimental interventions. Often arising in combination, these issues generally result in observed stochastic structure that cannot be characterized using the established notion of conditional independence. As a result, we are left with only a partial understanding of which aspects of a system can be inferred from the available data, and we lack effective methods for fundamental problems such as inference in the presence of feedback loops. The aim of the new project is to move beyond conditional independence structure to obtain a deeper understanding of the inherent limitations on what can be inferred from imperfect measurements, and to design novel statistical methodology to infer estimable quantities. The unique feature of the proposed work is a focus on algebraic relations among moments of probability distributions and the subtle statistical issues arising when such relations are to be exploited in practical methodology.