Discovering and Learning Causal Bayesian Models with Latent Variables

Abstract

The causal discovery of Bayesian networks is an active and important topic in artifi?cial intelligence, as sources and volumes of data continue to grow along with the popularity of Bayesian modeling methods. Causal Bayesian networks allow people to investigate causal relationships and modeling under uncertainty in an intuitive fashion. However, in many real world cases, some variables cannot be directly measured or people are simply unaware of their existence; these are called latent variables. In this thesis, we develop a unique explicit process for positing latent variables and incorporating them in a metric-based causal discovery program.