Learning Bayesian networks with and without hidden variables

Bayesian networks (BNs) represent patterns of conditional independence between random variables and sometimes also represent causal relations. There are two main approaches to learning BNs from data: "search-and-score" and "constraint-based". I will describe both of these approaches and consider their competing merits. One advantage of the "constraint-based" approach is that it lends itself naturally to detecting when there must be hidden (aka "latent") variables in the data-generating process. I will briefly describe ongoing work where I attempt to learn BNs with hidden variables by learning a set of BNs without them

 

The video can be viewed by anyone @york.ac.uk: Watch the video.