Inferning trees

Abstract

We consider the problem of learning an Ising model for the purpose of subsequently performing inference from partial observations. This is in contrast to most other work on graphical model learning, which tries to learn the true underlying graph. This objective requires a lower bound on the strength of edges for identifiability of the model. We show that in the relatively simple case of tree models, the Chow-Liu algorithm learns a distribution with accurate low-order marginals despite the model possibly being non-identifiable. In other words, a model that appears rather different from the truth nevertheless allows to carry out inference accurately.