Learning loosely connected Markov random fields

Abstract

We present a general algorithm for learning the structure of discrete Markov random fields from i.i.d. samples. Several algorithms have been proposed for structure learning algorithms earlier and each of these address the learning problem under different assumptions. Our algorithm provides a unified view in the following sense: when our algorithm is applied to each of the special cases, it results in a the same computational complexity as earlier algorithms. More importantly, our approach also provides a new low-computational complexity algorithm for the case of Ising models where the underlying graph is the Erdos-Renyi random graph G(p,c/p).