ABSTRACT
This chapter offers a concise introduction to the theory of latent tree models. It highlights the role of tree metrics in the structural description of this model class, in designing learning algorithms, and in understanding fundamental limits of what and when can be learned. The chapter discusses latent tree models and provides motivation to work with this model class. It presents Gaussian and general Markov models as subclasses of latent tree models that admits tractable and rigorous analysis. The chapter also discusses parameter identifiability, sample complexity, and model selection methods. Latent tree models are graphical models defined on trees, in which only a subset of variables is observed. They were first discussed by Judea Pearl as tree-decomposable distributions to generalize star-decomposable distributions such as the latent class model. Latent tree models represent a larger family of probability distributions than fully-observed tree models but retain some of their computational advantages, which is particularly important in high-dimensional settings.
