Learning minimal latent directed information trees

Jalal Etesami, Negar Kiyavash, Todd P. Coleman

Abstract: THIS PAPER IS ELIGIBLE FOR THE STUDENT PAPER AWARD - We propose a framework for learning the structure of a minimal latent tree with an associated discrepancy measure. Specifically, we apply this algorithm to recover the minimal latent directed information tree on a mixture of set of observed and unobserved random processes. Directed information trees are a new type of probabilistic graphical model based on directed information that represent the casual dynamics among random processes in a stochastic systems. To the best of our knowledge, this is the first approach that recovers these type of latent graphical models where samples are available only from a subset of processes.