B1606
Title: Identifiability of binary Bayesian networks with one latent variable
Authors: Hisayuki Hara - Doshisha University (Japan) [presenting]
Abstract: Parameter identifiability of graphical models with latent variables is a challenging problem. For Gaussian directed graphical models with one latent variable, some useful sufficient conditions for a model to be generically identifiable up to sign change are known. We discuss binary graphical models defined by directed acyclic graphs with one latent variable which is parental to all observable variables. We show that a part of previous results is applicable also to binary graphical models, and provide a sufficient condition for a model to be generically identifiable up to label swapping of a latent variable.
